You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@beam.apache.org by Maximilian Michels <mx...@apache.org> on 2018/11/07 19:50:18 UTC
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list
argument
+1
If the preferred approach is to eventually have the JobServer serve the
options, then the best intermediate solution is to replicate common
options in the SDKs.
If we went down the "--runner_option" path, we would end up with
multiple ways of specifying the same options. We would eventually have
to deprecate "runner options" once we have the JobServer approach. I'd
like to avoid that.
For the upcoming release we can revert the changes again and add the
most common missing options to the SDKs. Then hopefully we should have
fetching implemented for the release after.
Do you think that is feasible?
Thanks,
Max
On 30.10.18 23:00, Lukasz Cwik wrote:
> I still like #3 the most, just can't devote the time to get it done.
>
> Instead of going with a fully implemented #3, we could hardcode the a
> subset of options and types within each SDK until the job server is
> ready to provide this information and then migrate to the "full" list.
> This would be an easy path for SDKs to take on. They could "know" of a
> few well known options, and if they want to support all options, they
> implement the integration with the job server.
>
> On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels <mxm@apache.org
> <ma...@apache.org>> wrote:
>
> > I would prefer we don't introduce a (quirky) way of passing
> unknown options that forces users to type JSON into the command line
> (or similar acrobatics)
> Same here, the JSON approach seems technically nice but too bulky
> for users.
>
> > To someone wanting to run a pipeline, all options are equally
> important, whether they are application specific, SDK specific or
> runner specific.
>
> I'm also reluctant to force users to use `--runner_option=` because the
> division into "Runner" options and other options seems rather arbitrary
> to users. Most built-in options are also Runner-related.
>
> > It should be possible to *optionally* qualify/scope (to cover
> cases where there is ambiguity), but otherwise I prefer the format
> we currently have.
>
> Yes, namespacing is a problem. What happens if the user defines a
> custom
> PipelineOption which clashes with one of the builtin ones? If both are
>
> set, which one is actually applied?
>
>
> Note that PipelineOptions so far has been treating name equality to mean
> option equality and the Java implementation has a bunch of strict checks
> to make sure that default values aren't used for duplicate definitions,
> they have the same type, etc...
> With 1), you fail the job if the runner can't understand your option
> because its not represented the same way. User then needs to fix-up
> their declaration of the option name.
> With 2), there are no name conflicts, the SDK will need to validate that
> the option isn't set in both formats and error out if it is before
> pipeline submission time.
> With 3), you can prefetch all the options and error out to the user
> during argument parsing time.
>
>
>
> Here is a summary of the possible paths going forward:
>
>
> 1) Validate PipelineOptions at Runner side
> ==========================================
>
> The main issue raised here was that we want to move away from parsing
> arguments which look like options without validating them. An easy fix
> would be to actually validate them on the Runner side. This could be
> done by changing the deserialization code of PipelineOptions which so
> far ignores unknown JSON options.
>
> See: PipelineOptionsTranslation.fromProto(Struct protoOptions)
>
> Actually, this wouldn't work for user-defined PipelineOptions because
> they might not be known to the Runner (if they are defined in Python).
>
>
> 2) Introduce a Runner-Option Flag
> =================================
>
> In this approach we would try to add as many pipeline options for a
> Runner to the SDK, but allow additional Runner options to be passed
> using the `--runner-option=key=val` flag. The Runner, like in 1), would
> have to ensure validation. I think this has been the most favored
> way so
> far. Going forward, that means that `--parallelism=4` and
> `--runner-option=parallelism=4` will have the same effect for the Flink
> Runner.
>
>
> 3) Implement Fetching of Options from JobServer
> ===============================================
>
> The options are retrieved from the JobServer before submitting the
> pipeline. I think this would be ideal but, as mentioned before, it
> increases the complexity for implementing new SDKs and might overall
> just not be worth the effort.
>
>
> What do you think? I'd implement 2) for the next release, unless there
> are advocates for a different approach.
>
> Cheers,
> Max
>
> On 25.10.18 21:19, Thomas Weise wrote:
> > Reminder that this is something we ideally address before the next
> > release...
> >
> > Considering the discussion so far, my preference is that we get away
> > from unknown options and discover valid options from the runner (by
> > expanding the job service).
> >
> > Once the SDK is aware of all valid options, it is possible to
> provide
> > meaningful feedback to the user (validate or help), and correctly
> handle
> > scopes and types.
> >
> > I would prefer we don't introduce a (quirky) way of passing unknown
> > options that forces users to type JSON into the command line (or
> similar
> > acrobatics). To someone wanting to run a pipeline, all options are
> > equally important, whether they are application specific, SDK
> specific
> > or runner specific. It should be possible to *optionally*
> qualify/scope
> > (to cover cases where there is ambiguity), but otherwise I prefer
> the
> > format we currently have.
> >
> > Regarding type inference: Correct handling of numeric types
> matters, see
> > following issue with protobuf (not JSON):
> > https://issues.apache.org/jira/browse/BEAM-5509
> >
> > Thomas
> >
> >
> > On Thu, Oct 18, 2018 at 6:55 AM Robert Bradshaw
> <robertwb@google.com <ma...@google.com>
> > <mailto:robertwb@google.com <ma...@google.com>>> wrote:
> >
> > On Wed, Oct 17, 2018 at 11:35 PM Lukasz Cwik
> <lcwik@google.com <ma...@google.com>
> > <mailto:lcwik@google.com <ma...@google.com>>> wrote:
> >
> >
> > On Tue, Oct 16, 2018 at 11:51 AM Robert Bradshaw
> > <robertwb@google.com <ma...@google.com>
> <mailto:robertwb@google.com <ma...@google.com>>> wrote:
> >
> > On Tue, Oct 16, 2018 at 7:03 PM Lukasz Cwik
> > <lcwik@google.com <ma...@google.com>
> <mailto:lcwik@google.com <ma...@google.com>>> wrote:
> > >
> > > For all unknown options, the SDK can require that all
> > flag values be specified explicitly as a valid JSON type.
> > > starts with { -> object
> > > starts with [ -> list
> > > starts with " -> string
> > > is null / true / false -> null / true / false
> > > otherwise is number.
> > >
> > > This isn't great for strings but works well for
> all the
> > other types.
> > >
> > > Thus for known options, the additional typing
> information
> > would disambiguate whether something should be a
> > string/boolean/number/object/list but for unknown
> options we
> > would expect the user to use valid JSON explicitly
> and write:
> > > --foo={"object": "value"}
> > > --foo=["value", "value2"]
> > > --foo="string value"
> >
> > Due to shell escaping, one would have to write
> >
> > --foo=\"string value\"
> >
> > or actually, due to the space
> >
> > --foo='"string value"'
> >
> > or some other variation on that, which is really
> > unfortunate. (The JSON list/objects would need similar
> > quoting, but that's less surprising.) Also, does this
> mean
> > we'd only have one kind of number (not integer vs. float,
> > i.e. --parallelism=5.0 works)? I suppose that is JSON.
> >
> >
> > Yes, I was suspecting that users would need to type the
> second
> > variant as \"...\" I found more burdensome then '"..."'
> >
> >
> > > --foo=3.5 --foo=-4
> > > --foo=true --foo=false
> > > --foo=null
> > > This also works if the flag is repeated, so --foo=3.5
> > --foo=-4 is [3.5, -4]
> >
> > The thing that sparked this discussion was what to do
> when
> > unknown foo is repeated, but only one value given.
> >
> >
> > If the person only specifies one value, then they have to
> > disambiguate and put it in a list, only if they specify more
> > then one value will they have to turn it into a list.
> >
> > I believe we could come up with other schemes on how to
> convert
> > unknown options to JSON where we prefer strings over
> non-string
> > types like null/true/false/numbers/list/object and
> require the
> > user to escape out of the string default but anything that is
> > too different from strict JSON would cause headaches when
> > attempting to explain the format to users. I think a happy
> > middle ground would be that we will only require escaping for
> > strings which are ambiguous, so things like true, null,
> false,
> > ... to be treated as strings would require the user to
> escape them.
> >
> >
> > I'd prefer to avoid inferring the type of an unknown argument
> based
> > on its contents, which can lead to surprises. We could
> declare every
> > unknown type to be repeated string, and let any
> parsing/validation
> > occur on the runner. If desired, we could pass these around as a
> > single "runner options" dict that runners could inspect and
> use to
> > populate the actual dict rather than mixing parsed and unparsed
> > options.
> >
> >
> >
> > > On Tue, Oct 16, 2018 at 7:56 AM Thomas Weise
> > <thw@apache.org <ma...@apache.org>
> <mailto:thw@apache.org <ma...@apache.org>>> wrote:
> > >>
> > >> Discovering options from the job server seems
> preferable
> > over replicating runner options in SDKs.
> > >>
> > >> Runners evolve on their own, and with portability the
> > SDK does not need to know anything about the runner.
> > >>
> > >> Regarding --runner-option. It is true that this looks
> > less user friendly. On the other hand it eliminates the
> > possibility of name collisions.
> > >>
> > >> But if options are discovered, the SDK can
> perform full
> > validation. It would only be necessary to use explicit
> > scoping when there is ambiguity.
> > >>
> > >> Thomas
> > >>
> > >>
> > >> On Tue, Oct 16, 2018 at 3:48 AM Maximilian Michels
> > <mxm@apache.org <ma...@apache.org>
> <mailto:mxm@apache.org <ma...@apache.org>>> wrote:
> > >>>
> > >>> Fetching options directly from the Runner's
> JobServer
> > seems like the
> > >>> ideal solution. I agree with Robert that it creates
> > additional
> > >>> complexity for SDK authors, so the `--runner-option`
> > flag would be an
> > >>> easy and explicit way to specify additional
> Runner options.
> > >>>
> > >>> The format I prefer would be:
> --runner_option=key1=val1
> > >>> --runner_option=key2=val2
> > >>>
> > >>> Now, from the perspective of end users, I think
> it is
> > neither convenient
> > >>> nor reasonable to require the use of the
> > `--runner-option` flag. To the
> > >>> user it seems nebulous why some pipeline options
> live
> > in the top-level
> > >>> option namespace while others need to be nested
> within
> > an option. This
> > >>> is amplified by there being two Runners the user
> needs
> > to be aware of,
> > >>> i.e. PortableRunner and the actual Runner
> > (Dataflow/Flink/Spark..).
> > >>>
> > >>> I feel like we would eventually replicate all
> options
> > in the SDK because
> > >>> otherwise users have to use the
> `--runner-option`, but
> > at least we can
> > >>> specify options which have not been replicated yet.
> > >>>
> > >>> -Max
> > >>>
> > >>> On 16.10.18 10:27, Robert Bradshaw wrote:
> > >>> > Yes, we don't know how to parse and/or
> validate it.
> > >>> >
> > >>> > On Tue, Oct 16, 2018 at 1:14 AM Lukasz Cwik
> > <lcwik@google.com <ma...@google.com>
> <mailto:lcwik@google.com <ma...@google.com>>
> > >>> > <mailto:lcwik@google.com
> <ma...@google.com> <mailto:lcwik@google.com
> <ma...@google.com>>>>
> > wrote:
> > >>> >
> > >>> > I see, is the issue that we currently are
> using a
> > JSON
> > >>> > representation for options when being
> serialized
> > and when we get
> > >>> > some unknown option, we don't know how to
> convert
> > it into its JSON form?
> > >>> >
> > >>> > On Mon, Oct 15, 2018 at 2:41 PM Robert
> Bradshaw
> > <robertwb@google.com <ma...@google.com>
> <mailto:robertwb@google.com <ma...@google.com>>
> > >>> > <mailto:robertwb@google.com
> <ma...@google.com>
> > <mailto:robertwb@google.com
> <ma...@google.com>>>> wrote:
> > >>> >
> > >>> > On Mon, Oct 15, 2018 at 11:30 PM
> Lukasz Cwik
> > <lcwik@google.com <ma...@google.com>
> <mailto:lcwik@google.com <ma...@google.com>>
> > >>> > <mailto:lcwik@google.com
> <ma...@google.com>
> > <mailto:lcwik@google.com <ma...@google.com>>>>
> wrote:
> > >>> > >
> > >>> > > On Mon, Oct 15, 2018 at 1:17 PM Robert
> > Bradshaw
> > >>> > <robertwb@google.com
> <ma...@google.com>
> > <mailto:robertwb@google.com
> <ma...@google.com>> <mailto:robertwb@google.com
> <ma...@google.com>
> > <mailto:robertwb@google.com
> <ma...@google.com>>>> wrote:
> > >>> > >>
> > >>> > >> On Mon, Oct 15, 2018 at 7:50 PM
> Lukasz Cwik
> > >>> > <lcwik@google.com
> <ma...@google.com> <mailto:lcwik@google.com
> <ma...@google.com>>
> > <mailto:lcwik@google.com <ma...@google.com>
> <mailto:lcwik@google.com <ma...@google.com>>>> wrote:
> > >>> > >> >
> > >>> > >> > I agree with the sentiment for
> better
> > error checking.
> > >>> > >> >
> > >>> > >> > We can try to make it such that
> the SDK
> > can "fetch" the
> > >>> > set of options that the runner supports by
> > making a call to the
> > >>> > Job API. The API could return a list of
> > option names
> > >>> > (descriptions for --help purposes and also
> > potentially the
> > >>> > expected format) which would remove
> the worry
> > around "unknown"
> > >>> > options. Yes I understand to be able
> to make
> > the Job API call,
> > >>> > we may need to parse some options from the
> > args parameters first
> > >>> > and then parse the unknown options
> after they
> > are fetched.
> > >>> > >>
> > >>> > >> This is an interesting idea, but
> seems it
> > could get quite
> > >>> > complicated.
> > >>> > >> E.g. for delegating runners, one would
> > first read the options to
> > >>> > >> determine which runner to fetch the
> > options from, which
> > >>> > would then
> > >>> > >> return a set of options that possibly
> > depends on the values
> > >>> > of some of
> > >>> > >> its options...
> > >>> > >>
> > >>> > >> > Alternatively, we can choose an
> > explicit format upfront.
> > >>> > >> > To expand on the exact format for
> > --runner_option=...,
> > >>> > here are some different ideas:
> > >>> > >> > 1) Specified multiple times,
> each one
> > is an explicit flag
> > >>> > >> > --runner_option=--blah=bar
> > --runner_option=--foo=baz1
> > >>> > --runner_option=--foo=baz2
> > >>> > >>
> > >>> > >> I'm -1 on this format. We should move
> > away from the idea
> > >>> > that options
> > >>> > >> == flags (as that doesn't compose well
> > with other libraries
> > >>> > that do
> > >>> > >> their own flags parsing). The
> ability to
> > parse a set of
> > >>> > flags into
> > >>> > >> options is just a convenience that an
> > author may (or may
> > >>> > not) choose
> > >>> > >> to use (e.g. when running pipelines a
> > long-lived process like a
> > >>> > >> service or a notebook, the command
> line
> > flags are almost
> > >>> > certainly not
> > >>> > >> the right interface).
> > >>> > >>
> > >>> > >> > 2) specified multiple times, we drop
> > the explicit flag
> > >>> > >> > --runner_option=blah=bar
> > --runner_option=foo=baz1
> > >>> > --runner_option=foo=baz2
> > >>> > >>
> > >>> > >> This or (4) is my preference.
> > >>> > >>
> > >>> > >> > 3) we use a string which the
> runner can
> > choose to
> > >>> > interpret however they want (JSON/XML
> shown
> > below)
> > >>> > >> > --runner_option='{"blah": "bar",
> "foo":
> > ["baz1", "baz2"]}'
> > >>> > >> >
> > >>> >
> >
> --runner_option='<options><blah>bar</blah><foo>baz1</foo><foo>baz2</foo></options>'
> > >>> > >>
> > >>> > >> This would make validation hard.
> Also, I
> > think it makes
> > >>> > sense for some
> > >>> > >> runner options to be "shared"
> > (parallelism") by convention,
> > >>> > so letting
> > >>> > >> it be a free-form string wouldn't
> allow
> > different runners to
> > >>> > inspect
> > >>> > >> different bits.
> > >>> > >>
> > >>> > >> We should consider if we should
> use urns
> > for namespacing, and
> > >>> > >> assigning semantic meaning to
> strings, here.
> > >>> > >>
> > >>> > >> > 4) we use a string which must be a
> > specific format such as
> > >>> > JSON (allows the SDK to do simple
> validation):
> > >>> > >> > --runner_option='{"blah": "bar",
> "foo":
> > ["baz1", "baz2"]}'
> > >>> > >>
> > >>> > >> I like this in that at least some
> > validation can be
> > >>> > performed, and
> > >>> > >> expectations of how to format richer
> > types. On the other
> > >>> > hand it gets
> > >>> > >> a bit verbose, given that most (I'd
> > imagine) options will be
> > >>> > simple.
> > >>> > >> As with normal options,
> > >>> > >>
> > >>> > >> --option1=value1 --option2=value2
> > >>> > >>
> > >>> > >> is shorthand for {"option1": value1,
> > "option2": value2}.
> > >>> > >>
> > >>> > > I lean to 4 the most. With 2, you
> run into
> > issues of what
> > >>> > does --runner_option=foo=["a", "b"]
> > --runner_option=foo=["c",
> > >>> > "d"] mean?
> > >>> > > Is it an error or list of lists or
> > concatenated. Similar
> > >>> > issues for map types represented via JSON
> > object {...}
> > >>> >
> > >>> > We can err to be on the safe side
> > unless/until an argument can
> > >>> > be made
> > >>> > that merging is more natural. I just think
> > this will be excessively
> > >>> > verbose to use.
> > >>> >
> > >>> > >> > I would strongly suggest that we go
> > with the "fetch"
> > >>> > approach, since this makes the set of
> options
> > discoverable and
> > >>> > helps users find errors much earlier
> in their
> > pipeline.
> > >>> > >>
> > >>> > >> This seems like an advanced
> feature that
> > SDKs may want to
> > >>> > support, but
> > >>> > >> I wouldn't want to require this
> > complexity for bootstrapping
> > >>> > an SDK.
> > >>> > >>
> > >>> > > SDKs that are starting off wouldn't
> need
> > to "fetch" options,
> > >>> > they could choose to not support runner
> > options or they could
> > >>> > choose to pass all options through to the
> > runner blindly.
> > >>> > Fetching the options only provides the SDK
> > the ability to
> > >>> > provide error checking upfront and useful
> > error/help messages.
> > >>> >
> > >>> > But how to even pass all options through
> > blindly is exactly the
> > >>> > difficulty we're running into here.
> > >>> >
> > >>> > >> Regarding always keeping runner
> options
> > separate, +1, though
> > >>> > I'm not
> > >>> > >> sure the line is always clear.
> > >>> > >>
> > >>> > >>
> > >>> > >> > On Mon, Oct 15, 2018 at 8:04 AM
> Robert
> > Bradshaw
> > >>> > <robertwb@google.com
> <ma...@google.com>
> > <mailto:robertwb@google.com
> <ma...@google.com>> <mailto:robertwb@google.com
> <ma...@google.com>
> > <mailto:robertwb@google.com
> <ma...@google.com>>>> wrote:
> > >>> > >> >>
> > >>> > >> >> On Mon, Oct 15, 2018 at 3:58 PM
> > Maximilian Michels
> > >>> > <mxm@apache.org
> <ma...@apache.org> <mailto:mxm@apache.org <ma...@apache.org>>
> > <mailto:mxm@apache.org <ma...@apache.org>
> <mailto:mxm@apache.org <ma...@apache.org>>>> wrote:
> > >>> > >> >> >
> > >>> > >> >> > I agree that the current approach
> > breaks the pipeline
> > >>> > options contract
> > >>> > >> >> > because "unknown" options get
> parsed
> > in the same way as
> > >>> > options which
> > >>> > >> >> > have been defined by the user.
> > >>> > >> >>
> > >>> > >> >> FWIW, I think we're already
> breaking
> > this "contract."
> > >>> > Unknown options
> > >>> > >> >> are silently ignored; with this
> change
> > we just change how
> > >>> > we record
> > >>> > >> >> them. It still feels a bit
> hacky though.
> > >>> > >> >>
> > >>> > >> >> > I'm not sure the
> `experiments` flag
> > works for us. AFAIK
> > >>> > it only allows
> > >>> > >> >> > true/false flags. We want to pass
> > all types of pipeline
> > >>> > options to the
> > >>> > >> >> > Runner.
> > >>> > >> >>
> > >>> > >> >> Experiments is an arbitrary set of
> > strings, which can be
> > >>> > of the form
> > >>> > >> >> "param=value" if that's useful.
> > (Dataflow does this.)
> > >>> > There is, again,
> > >>> > >> >> no namespacing on the param
> names, but
> > we could user urns
> > >>> > or impose
> > >>> > >> >> some other structure here.
> > >>> > >> >>
> > >>> > >> >> > How to solve this?
> > >>> > >> >> >
> > >>> > >> >> > 1) Add all options of all
> Runners to
> > each SDK
> > >>> > >> >> > We added some of the FlinkRunner
> > options to the Python
> > >>> > SDK but realized
> > >>> > >> >> > syncing is rather cumbersome
> in the
> > long term. However,
> > >>> > we want the most
> > >>> > >> >> > important options to be
> validated on
> > the client side.
> > >>> > >> >>
> > >>> > >> >> I don't think this is
> sustainable in
> > the long run.
> > >>> > However, thinking
> > >>> > >> >> about this, in the worse case
> > validation happens after
> > >>> > construction
> > >>> > >> >> but before execution (as with
> much of
> > our other
> > >>> > validation) so it
> > >>> > >> >> isn't that bad.
> > >>> > >> >>
> > >>> > >> >> > 2) Pass "unknown" options via a
> > separate list in the
> > >>> > Proto which can
> > >>> > >> >> > only be accessed internally
> by the
> > Runners. This still
> > >>> > allows passing
> > >>> > >> >> > arbitrary options but we wouldn't
> > leak unknown options
> > >>> > and display them
> > >>> > >> >> > as top-level options.
> > >>> > >> >>
> > >>> > >> >> I think there needs to be a way for
> > the user to
> > >>> > communicate values
> > >>> > >> >> directly to the runner
> regardless of
> > the SDK. My
> > >>> > preference would be
> > >>> > >> >> to make this explicit, e.g.
> (repeated)
> > >>> > --runner_option=..., rather
> > >>> > >> >> than scooping up all unknown
> flags at
> > command line
> > >>> > parsing time.
> > >>> > >> >> Perhaps an SDK that is aware of
> some
> > runners could choose
> > >>> > to lift
> > >>> > >> >> these as top-level options, but
> still
> > pass them as runner
> > >>> > options.
> > >>> > >> >>
> > >>> > >> >> > On 13.10.18 02:34, Charles
> Chen wrote:
> > >>> > >> >> > > The current release branch
> > >>> > >> >> > >
> > >>> >
> >
> (https://github.com/apache/beam/commits/release-2.8.0) was cut
> > >>> > after the
> > >>> > >> >> > > revert went in. Sent out
> > >>> > https://github.com/apache/beam/pull/6683 as a
> > >>> > >> >> > > revert of the revert.
> Regarding
> > your comment above,
> > >>> > I can help out with
> > >>> > >> >> > > the design / PR reviews for
> common
> > Python code as you
> > >>> > suggest.
> > >>> > >> >> > >
> > >>> > >> >> > > On Fri, Oct 12, 2018 at 4:48 PM
> > Thomas Weise
> > >>> > <thw@apache.org
> <ma...@apache.org> <mailto:thw@apache.org <ma...@apache.org>>
> > <mailto:thw@apache.org <ma...@apache.org>
> <mailto:thw@apache.org <ma...@apache.org>>>
> > >>> > >> >> > > <mailto:thw@apache.org
> <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>
> <mailto:thw@apache.org <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>>>> wrote:
> > >>> > >> >> > >
> > >>> > >> >> > > Thanks, will tag you and
> > looking forward to
> > >>> > feedback so we can
> > >>> > >> >> > > ensure that changes
> work for
> > everyone.
> > >>> > >> >> > >
> > >>> > >> >> > > Looking at the PR, I see
> > agreement from Max to
> > >>> > revert the change on
> > >>> > >> >> > > the release branch, but
> not in
> > master. Would you
> > >>> > mind to restore it
> > >>> > >> >> > > in master?
> > >>> > >> >> > >
> > >>> > >> >> > > Thanks
> > >>> > >> >> > >
> > >>> > >> >> > > On Fri, Oct 12, 2018 at
> 4:40
> > PM Ahmet Altay
> > >>> > <altay@google.com
> <ma...@google.com> <mailto:altay@google.com
> <ma...@google.com>>
> > <mailto:altay@google.com <ma...@google.com>
> <mailto:altay@google.com <ma...@google.com>>>
> > >>> > >> >> > >
> <mailto:altay@google.com <ma...@google.com>
> > <mailto:altay@google.com <ma...@google.com>>
> > >>> > <mailto:altay@google.com
> <ma...@google.com>
> > <mailto:altay@google.com
> <ma...@google.com>>>>> wrote:
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > > On Fri, Oct 12, 2018 at
> > 11:31 AM, Charles
> > >>> > Chen <ccy@google.com
> <ma...@google.com> <mailto:ccy@google.com <ma...@google.com>>
> > <mailto:ccy@google.com <ma...@google.com>
> <mailto:ccy@google.com <ma...@google.com>>>
> > >>> > >> >> > >
> <mailto:ccy@google.com <ma...@google.com>
> > <mailto:ccy@google.com <ma...@google.com>>
> > >>> > <mailto:ccy@google.com
> <ma...@google.com>
> > <mailto:ccy@google.com <ma...@google.com>>>>> wrote:
> > >>> > >> >> > >
> > >>> > >> >> > > What I mean is
> that a
> > user may find that
> > >>> > it works for them
> > >>> > >> >> > > to pass
> "--myarg blah"
> > and access it as
> > >>> > "options.myarg"
> > >>> > >> >> > > without explicitly
> > defining a "my_arg"
> > >>> > flag due to the added
> > >>> > >> >> > > logic. This is not
> > the intended behavior
> > >>> > and we may want to
> > >>> > >> >> > > change this
> > implementation detail in the
> > >>> > future. However,
> > >>> > >> >> > > having this
> logic in a
> > released version
> > >>> > makes it hard to
> > >>> > >> >> > > change this
> behavior
> > since users may
> > >>> > erroneously depend on
> > >>> > >> >> > > this undocumented
> > behavior. Instead, we
> > >>> > should namespace /
> > >>> > >> >> > > scope this so
> that it
> > is obvious that
> > >>> > this is meant for
> > >>> > >> >> > > runner (and not
> Beam
> > user) consumption.
> > >>> > >> >> > >
> > >>> > >> >> > > On Fri, Oct 12,
> 2018
> > at 10:48 AM Thomas Weise
> > >>> > >> >> > > <thw@apache.org
> <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>
> <mailto:thw@apache.org <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>>
> > >>> > <mailto:thw@apache.org
> <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>
> <mailto:thw@apache.org <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>>>> wrote:
> > >>> > >> >> > >
> > >>> > >> >> > > Can you please
> > elaborate more what
> > >>> > practical problems
> > >>> > >> >> > > this introduces
> > for users?
> > >>> > >> >> > >
> > >>> > >> >> > > I can see that
> > this change allows a
> > >>> > user to specify a
> > >>> > >> >> > > runner specific
> > option, which in the
> > >>> > future may change
> > >>> > >> >> > > because we
> decide
> > to scope
> > >>> > differently. If this only
> > >>> > >> >> > > affects
> users of
> > the portable Flink
> > >>> > runner (like us),
> > >>> > >> >> > > then no need to
> > revert, because at
> > >>> > this early stage we
> > >>> > >> >> > > prefer
> something
> > that works over
> > >>> > being blocked.
> > >>> > >> >> > >
> > >>> > >> >> > > It would
> also be
> > really great if some
> > >>> > of the core Python
> > >>> > >> >> > > SDK developers
> > could help out with
> > >>> > the design aspects
> > >>> > >> >> > > and PR
> reviews of
> > changes that affect
> > >>> > common Python
> > >>> > >> >> > > code.
> Anyone who
> > specifically wants
> > >>> > to be tagged on
> > >>> > >> >> > > relevant
> JIRAs and
> > PRs?
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > > I would be happy to be
> > tagged, and I can also
> > >>> > help with
> > >>> > >> >> > > including other
> relevant
> > folks whenever
> > >>> > possible. In general I
> > >>> > >> >> > > think Robert, Charles,
> > myself are good
> > >>> > candidates.
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > > Thanks
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > > On Fri, Oct 12,
> > 2018 at 10:20 AM
> > >>> > Ahmet Altay
> > >>> > >> >> > >
> <altay@google.com <ma...@google.com>
> > <mailto:altay@google.com <ma...@google.com>>
> > >>> > <mailto:altay@google.com
> <ma...@google.com>
> > <mailto:altay@google.com <ma...@google.com>>>
> <mailto:altay@google.com <ma...@google.com>
> > <mailto:altay@google.com <ma...@google.com>>
> > >>> > <mailto:altay@google.com
> <ma...@google.com>
> > <mailto:altay@google.com
> <ma...@google.com>>>>> wrote:
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > > On Fri, Oct
> > 12, 2018 at 10:11 AM,
> > >>> > Charles Chen
> > >>> > >> >> > >
> > <ccy@google.com <ma...@google.com>
> <mailto:ccy@google.com <ma...@google.com>>
> > >>> > <mailto:ccy@google.com
> <ma...@google.com>
> > <mailto:ccy@google.com <ma...@google.com>>>
> <mailto:ccy@google.com <ma...@google.com>
> > <mailto:ccy@google.com <ma...@google.com>>
> > >>> > <mailto:ccy@google.com
> <ma...@google.com>
> > <mailto:ccy@google.com <ma...@google.com>>>>> wrote:
> > >>> > >> >> > >
> > >>> > >> >> > > For
> > context, I made comments on
> > >>> > >> >> > >
> > https://github.com/apache/beam/pull/6600 noting
> > >>> > >> >> > >
> that the
> > changes being made
> > >>> > were not good for
> > >>> > >> >> > > Beam
> > >>> > backwards-compatibility. The change as is
> > >>> > >> >> > > allows
> > users to use pipeline
> > >>> > options without
> > >>> > >> >> > >
> explicitly
> > defining them,
> > >>> > which is not the type
> > >>> > >> >> > > of
> usage
> > we would like to
> > >>> > encourage since we
> > >>> > >> >> > >
> prefer to
> > be explicit
> > >>> > whenever possible. If
> > >>> > >> >> > > users
> > write pipelines with
> > >>> > this sort of pattern,
> > >>> > >> >> > >
> they will
> > potentially
> > >>> > encounter pain when
> > >>> > >> >> > >
> upgrading
> > to a later version
> > >>> > since this is an
> > >>> > >> >> > >
> > implementation detail and not
> > >>> > an officially
> > >>> > >> >> > >
> supported
> > pattern. I agree
> > >>> > with the comments
> > >>> > >> >> > >
> above that
> > this is ultimately
> > >>> > a scoping issue.
> > >>> > >> >> > > I would
> > not have a problem
> > >>> > with these changes if
> > >>> > >> >> > >
> they were
> > explicitly scoped
> > >>> > under either a
> > >>> > >> >> > >
> runner or
> > unparsed options
> > >>> > namespace.
> > >>> > >> >> > >
> > >>> > >> >> > > As a
> > second note, since the
> > >>> > 2.8.0 release is
> > >>> > >> >> > >
> being cut
> > right now, because
> > >>> > of these
> > >>> > >> >> > >
> > backwards-compatibility
> > >>> > concerns, I would
> > >>> > >> >> > > suggest
> > reverting these
> > >>> > changes, at least until
> > >>> > >> >> > >
> 2.8.0 is
> > cut, so we can have
> > >>> > a discussion here
> > >>> > >> >> > > before
> > committing to and
> > >>> > releasing any API-level
> > >>> > >> >> > >
> changes.
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > > +1 I would
> > like to revert the
> > >>> > changes in order not
> > >>> > >> >> > > rush
> this into
> > the release. Once
> > >>> > this discussion
> > >>> > >> >> > > results
> in an
> > agreement changes
> > >>> > can be brought back.
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > > On Fri,
> > Oct 12, 2018 at 9:26
> > >>> > AM Henning Rohde
> > >>> > >> >> > >
> > <herohde@google.com <ma...@google.com>
> <mailto:herohde@google.com <ma...@google.com>>
> > >>> > <mailto:herohde@google.com
> <ma...@google.com>
> > <mailto:herohde@google.com
> <ma...@google.com>>> <mailto:herohde@google.com
> <ma...@google.com>
> > <mailto:herohde@google.com <ma...@google.com>>
> > >>> > <mailto:herohde@google.com
> <ma...@google.com>
> > <mailto:herohde@google.com
> <ma...@google.com>>>>>
> > >>> > >> >> > > wrote:
> > >>> > >> >> > >
> > >>> > >> >> > >
> Agree
> > that pipeline
> > >>> > options lack some
> > >>> > >> >> > >
> > mechanism for scoping. It
> > >>> > is also not always
> > >>> > >> >> > >
> > possible distinguish
> > >>> > options meant to be
> > >>> > >> >> > >
> > consumed at pipeline
> > >>> > construction time, by
> > >>> > >> >> > > the
> > runner, by the SDK
> > >>> > harness, by the user
> > >>> > >> >> > >
> code
> > or any combination
> > >>> > -- and this causes
> > >>> > >> >> > >
> > confusion every now and then.
> > >>> > >> >> > >
> > >>> > >> >> > > For
> > Dataflow, we have
> > >>> > been using
> > >>> > >> >> > >
> > "experiments" for
> > >>> > arbitrary runner-specific
> > >>> > >> >> > >
> > options. It's simply a
> > >>> > string list pipeline
> > >>> > >> >> > >
> option
> > that all SDKs
> > >>> > support and, for Go at
> > >>> > >> >> > >
> least,
> > is sent to
> > >>> > portable runners. Flink
> > >>> > >> >> > >
> can do
> > the same in the
> > >>> > short term to move
> > >>> > >> >> > >
> forward.
> > >>> > >> >> > >
> > >>> > >> >> > >
> Henning
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > > On
> > Fri, Oct 12, 2018 at
> > >>> > 8:50 AM Thomas Weise
> > >>> > >> >> > >
> > <thw@apache.org <ma...@apache.org>
> <mailto:thw@apache.org <ma...@apache.org>>
> > >>> > <mailto:thw@apache.org
> <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>>
> <mailto:thw@apache.org <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>
> > >>> > <mailto:thw@apache.org
> <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>>>> wrote:
> > >>> > >> >> > >
> > >>> > >> >> > >
> > [moving to the list]
> > >>> > >> >> > >
> > >>> > >> >> > >
> > The requirement
> > >>> > driving this part of the
> > >>> > >> >> > >
> > change was to allow a
> > >>> > user to specify
> > >>> > >> >> > >
> > pipeline options that
> > >>> > a runner supports
> > >>> > >> >> > >
> > without having to
> > >>> > declare those in each
> > >>> > >> >> > >
> > language SDK.
> > >>> > >> >> > >
> > >>> > >> >> > >
> In
> > the specific
> > >>> > scenario, we have
> > >>> > >> >> > >
> > options that the
> > >>> > Flink runner supports
> > >>> > >> >> > >
> > (and can validate),
> > >>> > that are not
> > >>> > >> >> > >
> > enumerated in the
> > >>> > Python SDK.
> > >>> > >> >> > >
> > >>> > >> >> > >
> I
> > think we have a
> > >>> > bigger problem scoping
> > >>> > >> >> > >
> > pipeline options. For
> > >>> > example, the
> > >>> > >> >> > >
> > runner options are
> > >>> > dumped into the SDK
> > >>> > >> >> > >
> > worker. There is also
> > >>> > a possibility of
> > >>> > >> >> > >
> > name collisions. So I
> > >>> > think this would
> > >>> > >> >> > >
> > benefit from broader
> > >>> > feedback.
> > >>> > >> >> > >
> > >>> > >> >> > >
> > Thanks,
> > >>> > >> >> > >
> Thomas
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > >
> > ---------- Forwarded
> > >>> > message ---------
> > >>> > >> >> > >
> > From: *Charles Chen*
> > >>> > >> >> > >
> > >>> > <notifications@github.com
> <ma...@github.com>
> > <mailto:notifications@github.com
> <ma...@github.com>>
> > <mailto:notifications@github.com
> <ma...@github.com>
> > <mailto:notifications@github.com
> <ma...@github.com>>>
> > >>> > >> >> > >
> > >>> > <mailto:notifications@github.com
> <ma...@github.com>
> > <mailto:notifications@github.com
> <ma...@github.com>>
> > >>> > <mailto:notifications@github.com
> <ma...@github.com>
> > <mailto:notifications@github.com
> <ma...@github.com>>>>>
> > >>> > >> >> > >
> > Date: Fri, Oct 12,
> > >>> > 2018 at 8:36 AM
> > >>> > >> >> > >
> > Subject: Re:
> > >>> > [apache/beam] [BEAM-5442]
> > >>> > >> >> > >
> > Store duplicate
> > >>> > unknown options in a
> > >>> > >> >> > >
> > list argument (#6600)
> > >>> > >> >> > >
> > To: apache/beam
> > >>> > <beam@noreply.github.com
> <ma...@noreply.github.com>
> > <mailto:beam@noreply.github.com
> <ma...@noreply.github.com>>
> > <mailto:beam@noreply.github.com
> <ma...@noreply.github.com>
> > <mailto:beam@noreply.github.com
> <ma...@noreply.github.com>>>
> > >>> > >> >> > >
> > >>> > <mailto:beam@noreply.github.com
> <ma...@noreply.github.com>
> > <mailto:beam@noreply.github.com
> <ma...@noreply.github.com>>
> > <mailto:beam@noreply.github.com
> <ma...@noreply.github.com>
> > <mailto:beam@noreply.github.com
> <ma...@noreply.github.com>>>>>
> > >>> > >> >> > >
> > Cc: Thomas Weise
> > >>> > <thomas.weise@gmail.com
> <ma...@gmail.com>
> > <mailto:thomas.weise@gmail.com
> <ma...@gmail.com>>
> > <mailto:thomas.weise@gmail.com
> <ma...@gmail.com> <mailto:thomas.weise@gmail.com
> <ma...@gmail.com>>>
> > >>> > >> >> > >
> > >>> > <mailto:thomas.weise@gmail.com
> <ma...@gmail.com>
> > <mailto:thomas.weise@gmail.com
> <ma...@gmail.com>>
> > <mailto:thomas.weise@gmail.com
> <ma...@gmail.com>
> > <mailto:thomas.weise@gmail.com
> <ma...@gmail.com>>>>>,
> > >>> > >> >> > >
> > Mention
> > >>> > <mention@noreply.github.com
> <ma...@noreply.github.com>
> > <mailto:mention@noreply.github.com
> <ma...@noreply.github.com>>
> > <mailto:mention@noreply.github.com
> <ma...@noreply.github.com>
> > <mailto:mention@noreply.github.com
> <ma...@noreply.github.com>>>
> > >>> > >> >> > >
> > >>> > <mailto:mention@noreply.github.com
> <ma...@noreply.github.com>
> > <mailto:mention@noreply.github.com
> <ma...@noreply.github.com>>
> > >>> > <mailto:mention@noreply.github.com
> <ma...@noreply.github.com>
> > <mailto:mention@noreply.github.com
> <ma...@noreply.github.com>>>>>
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > >
> > CC: @tweise
> > >>> > <https://github.com/tweise>
> > >>> > >> >> > >
> > >>> > >> >> > >
> —
> > >>> > >> >> > >
> > You are receiving
> > >>> > this because you were
> > >>> > >> >> > >
> > mentioned.
> > >>> > >> >> > >
> > Reply to this email
> > >>> > directly, view it on
> > >>> > >> >> > >
> GitHub
> > >>> > >> >> > >
> > >>> >
> >
> <https://github.com/apache/beam/pull/6600#issuecomment-429367754>,
> > >>> > >> >> > >
> or
> > mute the thread
> > >>> > >> >> > >
> > >>> >
> >
> <https://github.com/notifications/unsubscribe-auth/AAQGDwwt15R85eq9pySUisyxq2HYz-Vyks5ukLcLgaJpZM4XMo-T>.
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> > >> >> > >
> > >>> >
> >
>
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list
argument
Posted by Maximilian Michels <mx...@apache.org>.
Thomas was so kind to implement Option 3) in
https://github.com/apache/beam/pull/7597
Heads-up to the Go SDK people to eventually implement the new
DescribePipelineOptionsRequest. Tracking issue:
https://issues.apache.org/jira/browse/BEAM-6549
Also related, we will have to follow-up with proper scoping of pipeline options:
https://issues.apache.org/jira/browse/BEAM-6537
Thanks,
Max
On 13.11.18 19:05, Robert Burke wrote:
> +1 to Option 3
>
> I'd rather have each SDK have a single point of well defined complexity to do
> something general, than have to make tiny but simple changes. Less toil and
> maintenance in the long run per SDK.
>
> Similarly I don't have time to make it happen right now.
>
> On Tue, Nov 13, 2018, 9:22 AM Thomas Weise <thw@apache.org
> <ma...@apache.org>> wrote:
>
> Discovering options from the job server would be the only way to perform
> full validation (and provide upfront help to the user).
>
> The runner cannot perform full validation, since it is not aware of the user
> and SDK options (that it has to forward to the SDK worker).
>
> Special runner options flag to forward unknown options also wouldn't fully
> solve the problem (besides being subject to change in the future). Let's say
> runner understands --fancy-int-option and the user repeats that option
> multiple times. Not knowing the type of option, the SDK will pass it as a
> list and the runner will fail.
>
> Replicating SDK options is a workaround for known runners but it really goes
> against the idea of portability (making assumptions outside the API
> contract). We already have runners implemented outside of Beam and hope for
> the ecosystem to grow. What we do for options should also work for those
> runners.
>
> I'm with Luke here that options discovery provides the best user experience
> and can address the other issues. Even the scenario of multiple intermediate
> runners could be addressed by forwarding the unparsed options with the
> discovery call. I don't see SDK implementation complexity as a significant
> drawback so far.
>
> Thomas
>
>
> On Mon, Nov 12, 2018 at 2:30 PM Lukasz Cwik <lcwik@google.com
> <ma...@google.com>> wrote:
>
>
> On Mon, Nov 12, 2018 at 9:38 AM Maximilian Michels <mxm@apache.org
> <ma...@apache.org>> wrote:
>
> Thank you Robert and Lukasz for your points.
>
> > Note that I believe that we will want to have multiple URLs to
> support cross language pipelines since we will want to be able to
> ask other SDK languages/versions for their "list" of supported
> PipelineOptions.
>
> Why is that? The Runner itself is the source of truth for its options.
>
>
> Because other languages (or even different versions of the same
> language) may have their own options. For example, the Go SDK talks to a
> Java service which applies a SQL transform and returns the expanded form
> (this may require knowledge of some options like credentials for the
> filesystem, ...) and then talks to a Python service that performs
> another transform expansion. Finally the pipeline containing Go, Java
> and Python transforms is submitted to a runner and it performs its own
> internal replacements/expansions related to executing the pipeline.
>
> Everything else is SDK-related and should be validated there.
>
> I imagined the process to go like this:
>
> a) Parse options to find JobServer URL
> a) Retrieve options from JobServer
> c) Parse all options
> ...continue as always...
>
> An option is just represented by a name and a type. There is nothing
> more to it, at least as of now. So it should be possible to parse them
> in the SDK without much further work.
>
> Nevertheless, I agree with your sentiment, Robert. The "runner_option"
> flag would prevent additional complexity. I still don't prefer it
> because it's not nice from an end user perspective. If we were to
> implement it, I would definitely go for the "option promotion" which
> you
> mentioned.
>
> I hadn't thought about delegating runners, although the PortableRunner
> is basically a delegating Runner. If that was an important feature, I
> suppose the "runner_option" would be the preferred way.
>
> All in all, since there doesn't seem to be an excitement to implement
> JobServer option retrieval and we will need the help of all SDK
> developers, "runner_option" seems to be the more likely path.
>
>
> I would say its a lack of time for people to improve this story over
> others but it can be revisited at some point in the future and I agree
> that using --runner_option as an interim provides value.
>
>
> -Max
>
> On 08.11.18 21:50, Lukasz Cwik wrote:
> > The purpose of the spec would be to provide the names, type and
> > descriptions of the options. We don't need anything beyond the JSON
> > types (string, number, bool, object, list) because the only
> ambiguity we
> > get is how do we parse command line string into the JSON type
> (and that
> > ambiguity is actually only between string and non-string since
> all the
> > other JSON types are unambiguous).
> >
> > Also, I believe the flow would be
> > 1) Parse options
> > a) Find the URL from args specified and/or additional methods on
> > PipelineOptions that exposes a programmatic way to set the URL
> during
> > parsing.
> > b) Query URL for option specs
> > c) Parse the remainder of the options
> > 2) Construct pipeline
> > 3) Choose runner
> > 4) Submit job to runner
> >
> > Note that I believe that we will want to have multiple URLs to
> support
> > cross language pipelines since we will want to be able to ask
> other SDK
> > languages/versions for their "list" of supported PipelineOptions.
> >
> > On Thu, Nov 8, 2018 at 11:51 AM Robert Bradshaw
> <robertwb@google.com <ma...@google.com>
> > <mailto:robertwb@google.com <ma...@google.com>>> wrote:
> >
> > There's two questions here:
> >
> > (A) What do we do in the short term?
> >
> > I think adding every runner option to every SDK is not
> sustainable
> > (n*m work, assuming every SDK knows about every runner), and
> having a
> > patchwork of options that were added as one-offs to SDKs is not
> > desirable either. Furthermore, it seems difficult to parse
> unknown
> > options as if they were valid options, so my preference here
> would be
> > to just use a special runner_option flag. (One could also
> pass a set
> > of unparsed/unvalidated runner options to the runner, even if
> they're
> > not distinguished for the user, and runners (or any
> intermediates)
> > could run a "promote" operation that promotes any of these
> unknowns
> > that they recognize to real options before further
> processing. The
> > parsing would be done as repeated-string, and not be
> intermingled with
> > the actually validated options. This is essential a variant of
> > option 1.)
> >
> > (B) What do do in the long term? While the JobServer approach
> sounds
> > nice, I think it introduces a lot of complexity (we have too
> much of
> > that already) and still doesn't completely solve the problem. In
> > particular, it changes the flow from
> >
> > 1. Parse options
> > 2. Construct pipeline
> > 3. Choose runner
> > 4. Submit job to runner
> >
> > to
> >
> > 1. Parse options
> > 2. Construct pipeline
> > 3. Choose runner
> > 4a. Query runner for option specs
> > 4b. Re-parse options
> > 4c. Submit job to runner
> >
> > In particular, doing 4b in the SDK rather than just let the
> runner
> > itself do the validation as part of (4) doesn't save much and
> forces
> > us to come up with a (probably incomplete) spec as to how to
> define
> > options, their types, and their validations. It also means that a
> > delegating runner must choose and interact with its downstream
> > runner(s) synchronously, else we haven't actually solved the
> issue.
> >
> > For these reasons, I don't think we even want to go with the
> JobServer
> > approach in the long term, which has bearing on (A).
> >
> > - Robert
> >
> >
> > On Wed, Nov 7, 2018 at 8:50 PM Maximilian Michels
> <mxm@apache.org <ma...@apache.org>
> > <mailto:mxm@apache.org <ma...@apache.org>>> wrote:
> > >
> > > +1
> > >
> > > If the preferred approach is to eventually have the JobServer
> > serve the
> > > options, then the best intermediate solution is to
> replicate common
> > > options in the SDKs.
> > >
> > > If we went down the "--runner_option" path, we would end
> up with
> > > multiple ways of specifying the same options. We would
> eventually
> > have
> > > to deprecate "runner options" once we have the JobServer
> > approach. I'd
> > > like to avoid that.
> > >
> > > For the upcoming release we can revert the changes again
> and add the
> > > most common missing options to the SDKs. Then hopefully we
> should
> > have
> > > fetching implemented for the release after.
> > >
> > > Do you think that is feasible?
> > >
> > > Thanks,
> > > Max
> > >
> > > On 30.10.18 23:00, Lukasz Cwik wrote:
> > > > I still like #3 the most, just can't devote the time to
> get it
> > done.
> > > >
> > > > Instead of going with a fully implemented #3, we could
> hardcode
> > the a
> > > > subset of options and types within each SDK until the
> job server is
> > > > ready to provide this information and then migrate to the
> > "full" list.
> > > > This would be an easy path for SDKs to take on. They could
> > "know" of a
> > > > few well known options, and if they want to support all
> > options, they
> > > > implement the integration with the job server.
> > > >
> > > > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels
> > <mxm@apache.org <ma...@apache.org>
> <mailto:mxm@apache.org <ma...@apache.org>>
> > > > <mailto:mxm@apache.org <ma...@apache.org>
> <mailto:mxm@apache.org <ma...@apache.org>>>> wrote:
> > > >
> > > > > I would prefer we don't introduce a (quirky) way
> of passing
> > > > unknown options that forces users to type JSON into the
> > command line
> > > > (or similar acrobatics)
> > > > Same here, the JSON approach seems technically nice
> but too
> > bulky
> > > > for users.
> > > >
> > > > > To someone wanting to run a pipeline, all options are
> > equally
> > > > important, whether they are application specific, SDK
> > specific or
> > > > runner specific.
> > > >
> > > > I'm also reluctant to force users to use
> `--runner_option=`
> > because the
> > > > division into "Runner" options and other options seems
> > rather arbitrary
> > > > to users. Most built-in options are also Runner-related.
> > > >
> > > > > It should be possible to *optionally*
> qualify/scope (to
> > cover
> > > > cases where there is ambiguity), but otherwise I
> prefer the
> > format
> > > > we currently have.
> > > >
> > > > Yes, namespacing is a problem. What happens if the user
> > defines a
> > > > custom
> > > > PipelineOption which clashes with one of the builtin
> ones?
> > If both are
> > > >
> > > > set, which one is actually applied?
> > > >
> > > >
> > > > Note that PipelineOptions so far has been treating name
> > equality to mean
> > > > option equality and the Java implementation has a bunch of
> > strict checks
> > > > to make sure that default values aren't used for duplicate
> > definitions,
> > > > they have the same type, etc...
> > > > With 1), you fail the job if the runner can't understand
> your
> > option
> > > > because its not represented the same way. User then
> needs to fix-up
> > > > their declaration of the option name.
> > > > With 2), there are no name conflicts, the SDK will need to
> > validate that
> > > > the option isn't set in both formats and error out if it
> is before
> > > > pipeline submission time.
> > > > With 3), you can prefetch all the options and error out
> to the user
> > > > during argument parsing time.
> > > >
> > > >
> > > >
> > > > Here is a summary of the possible paths going forward:
> > > >
> > > >
> > > > 1) Validate PipelineOptions at Runner side
> > > > ==========================================
> > > >
> > > > The main issue raised here was that we want to move away
> > from parsing
> > > > arguments which look like options without validating
> them.
> > An easy fix
> > > > would be to actually validate them on the Runner
> side. This
> > could be
> > > > done by changing the deserialization code of
> > PipelineOptions which so
> > > > far ignores unknown JSON options.
> > > >
> > > > See: PipelineOptionsTranslation.fromProto(Struct
> protoOptions)
> > > >
> > > > Actually, this wouldn't work for user-defined
> > PipelineOptions because
> > > > they might not be known to the Runner (if they are
> defined
> > in Python).
> > > >
> > > >
> > > > 2) Introduce a Runner-Option Flag
> > > > =================================
> > > >
> > > > In this approach we would try to add as many pipeline
> > options for a
> > > > Runner to the SDK, but allow additional Runner
> options to
> > be passed
> > > > using the `--runner-option=key=val` flag. The
> Runner, like
> > in 1), would
> > > > have to ensure validation. I think this has been the
> most
> > favored
> > > > way so
> > > > far. Going forward, that means that
> `--parallelism=4` and
> > > > `--runner-option=parallelism=4` will have the same
> effect
> > for the Flink
> > > > Runner.
> > > >
> > > >
> > > > 3) Implement Fetching of Options from JobServer
> > > > ===============================================
> > > >
> > > > The options are retrieved from the JobServer before
> > submitting the
> > > > pipeline. I think this would be ideal but, as mentioned
> > before, it
> > > > increases the complexity for implementing new SDKs and
> > might overall
> > > > just not be worth the effort.
> > > >
> > > >
> > > > What do you think? I'd implement 2) for the next
> release,
> > unless there
> > > > are advocates for a different approach.
> > > >
> > > > Cheers,
> > > > Max
> >
>
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list argument
Posted by Robert Burke <ro...@frantil.com>.
+1 to Option 3
I'd rather have each SDK have a single point of well defined complexity to
do something general, than have to make tiny but simple changes. Less toil
and maintenance in the long run per SDK.
Similarly I don't have time to make it happen right now.
On Tue, Nov 13, 2018, 9:22 AM Thomas Weise <th...@apache.org> wrote:
> Discovering options from the job server would be the only way to perform
> full validation (and provide upfront help to the user).
>
> The runner cannot perform full validation, since it is not aware of the
> user and SDK options (that it has to forward to the SDK worker).
>
> Special runner options flag to forward unknown options also wouldn't fully
> solve the problem (besides being subject to change in the future). Let's
> say runner understands --fancy-int-option and the user repeats that option
> multiple times. Not knowing the type of option, the SDK will pass it as a
> list and the runner will fail.
>
> Replicating SDK options is a workaround for known runners but it really
> goes against the idea of portability (making assumptions outside the API
> contract). We already have runners implemented outside of Beam and hope for
> the ecosystem to grow. What we do for options should also work for those
> runners.
>
> I'm with Luke here that options discovery provides the best user
> experience and can address the other issues. Even the scenario of multiple
> intermediate runners could be addressed by forwarding the unparsed options
> with the discovery call. I don't see SDK implementation complexity as a
> significant drawback so far.
>
> Thomas
>
>
> On Mon, Nov 12, 2018 at 2:30 PM Lukasz Cwik <lc...@google.com> wrote:
>
>>
>> On Mon, Nov 12, 2018 at 9:38 AM Maximilian Michels <mx...@apache.org>
>> wrote:
>>
>>> Thank you Robert and Lukasz for your points.
>>>
>>> > Note that I believe that we will want to have multiple URLs to support
>>> cross language pipelines since we will want to be able to ask other SDK
>>> languages/versions for their "list" of supported PipelineOptions.
>>>
>>> Why is that? The Runner itself is the source of truth for its options.
>>>
>>
>> Because other languages (or even different versions of the same language)
>> may have their own options. For example, the Go SDK talks to a Java service
>> which applies a SQL transform and returns the expanded form (this may
>> require knowledge of some options like credentials for the filesystem, ...)
>> and then talks to a Python service that performs another transform
>> expansion. Finally the pipeline containing Go, Java and Python transforms
>> is submitted to a runner and it performs its own internal
>> replacements/expansions related to executing the pipeline.
>>
>>
>>> Everything else is SDK-related and should be validated there.
>>>
>>> I imagined the process to go like this:
>>>
>>> a) Parse options to find JobServer URL
>>> a) Retrieve options from JobServer
>>> c) Parse all options
>>> ...continue as always...
>>>
>>> An option is just represented by a name and a type. There is nothing
>>> more to it, at least as of now. So it should be possible to parse them
>>> in the SDK without much further work.
>>>
>>> Nevertheless, I agree with your sentiment, Robert. The "runner_option"
>>> flag would prevent additional complexity. I still don't prefer it
>>> because it's not nice from an end user perspective. If we were to
>>> implement it, I would definitely go for the "option promotion" which you
>>> mentioned.
>>>
>>> I hadn't thought about delegating runners, although the PortableRunner
>>> is basically a delegating Runner. If that was an important feature, I
>>> suppose the "runner_option" would be the preferred way.
>>>
>>> All in all, since there doesn't seem to be an excitement to implement
>>> JobServer option retrieval and we will need the help of all SDK
>>> developers, "runner_option" seems to be the more likely path.
>>>
>>
>> I would say its a lack of time for people to improve this story over
>> others but it can be revisited at some point in the future and I agree that
>> using --runner_option as an interim provides value.
>>
>>
>>>
>>> -Max
>>>
>>> On 08.11.18 21:50, Lukasz Cwik wrote:
>>> > The purpose of the spec would be to provide the names, type and
>>> > descriptions of the options. We don't need anything beyond the JSON
>>> > types (string, number, bool, object, list) because the only ambiguity
>>> we
>>> > get is how do we parse command line string into the JSON type (and
>>> that
>>> > ambiguity is actually only between string and non-string since all the
>>> > other JSON types are unambiguous).
>>> >
>>> > Also, I believe the flow would be
>>> > 1) Parse options
>>> > a) Find the URL from args specified and/or additional methods on
>>> > PipelineOptions that exposes a programmatic way to set the URL during
>>> > parsing.
>>> > b) Query URL for option specs
>>> > c) Parse the remainder of the options
>>> > 2) Construct pipeline
>>> > 3) Choose runner
>>> > 4) Submit job to runner
>>> >
>>> > Note that I believe that we will want to have multiple URLs to support
>>> > cross language pipelines since we will want to be able to ask other
>>> SDK
>>> > languages/versions for their "list" of supported PipelineOptions.
>>> >
>>> > On Thu, Nov 8, 2018 at 11:51 AM Robert Bradshaw <robertwb@google.com
>>> > <ma...@google.com>> wrote:
>>> >
>>> > There's two questions here:
>>> >
>>> > (A) What do we do in the short term?
>>> >
>>> > I think adding every runner option to every SDK is not sustainable
>>> > (n*m work, assuming every SDK knows about every runner), and
>>> having a
>>> > patchwork of options that were added as one-offs to SDKs is not
>>> > desirable either. Furthermore, it seems difficult to parse unknown
>>> > options as if they were valid options, so my preference here would
>>> be
>>> > to just use a special runner_option flag. (One could also pass a
>>> set
>>> > of unparsed/unvalidated runner options to the runner, even if
>>> they're
>>> > not distinguished for the user, and runners (or any intermediates)
>>> > could run a "promote" operation that promotes any of these unknowns
>>> > that they recognize to real options before further processing. The
>>> > parsing would be done as repeated-string, and not be intermingled
>>> with
>>> > the actually validated options. This is essential a variant of
>>> > option 1.)
>>> >
>>> > (B) What do do in the long term? While the JobServer approach
>>> sounds
>>> > nice, I think it introduces a lot of complexity (we have too much
>>> of
>>> > that already) and still doesn't completely solve the problem. In
>>> > particular, it changes the flow from
>>> >
>>> > 1. Parse options
>>> > 2. Construct pipeline
>>> > 3. Choose runner
>>> > 4. Submit job to runner
>>> >
>>> > to
>>> >
>>> > 1. Parse options
>>> > 2. Construct pipeline
>>> > 3. Choose runner
>>> > 4a. Query runner for option specs
>>> > 4b. Re-parse options
>>> > 4c. Submit job to runner
>>> >
>>> > In particular, doing 4b in the SDK rather than just let the runner
>>> > itself do the validation as part of (4) doesn't save much and
>>> forces
>>> > us to come up with a (probably incomplete) spec as to how to define
>>> > options, their types, and their validations. It also means that a
>>> > delegating runner must choose and interact with its downstream
>>> > runner(s) synchronously, else we haven't actually solved the issue.
>>> >
>>> > For these reasons, I don't think we even want to go with the
>>> JobServer
>>> > approach in the long term, which has bearing on (A).
>>> >
>>> > - Robert
>>> >
>>> >
>>> > On Wed, Nov 7, 2018 at 8:50 PM Maximilian Michels <mxm@apache.org
>>> > <ma...@apache.org>> wrote:
>>> > >
>>> > > +1
>>> > >
>>> > > If the preferred approach is to eventually have the JobServer
>>> > serve the
>>> > > options, then the best intermediate solution is to replicate
>>> common
>>> > > options in the SDKs.
>>> > >
>>> > > If we went down the "--runner_option" path, we would end up with
>>> > > multiple ways of specifying the same options. We would
>>> eventually
>>> > have
>>> > > to deprecate "runner options" once we have the JobServer
>>> > approach. I'd
>>> > > like to avoid that.
>>> > >
>>> > > For the upcoming release we can revert the changes again and
>>> add the
>>> > > most common missing options to the SDKs. Then hopefully we
>>> should
>>> > have
>>> > > fetching implemented for the release after.
>>> > >
>>> > > Do you think that is feasible?
>>> > >
>>> > > Thanks,
>>> > > Max
>>> > >
>>> > > On 30.10.18 23:00, Lukasz Cwik wrote:
>>> > > > I still like #3 the most, just can't devote the time to get it
>>> > done.
>>> > > >
>>> > > > Instead of going with a fully implemented #3, we could
>>> hardcode
>>> > the a
>>> > > > subset of options and types within each SDK until the job
>>> server is
>>> > > > ready to provide this information and then migrate to the
>>> > "full" list.
>>> > > > This would be an easy path for SDKs to take on. They could
>>> > "know" of a
>>> > > > few well known options, and if they want to support all
>>> > options, they
>>> > > > implement the integration with the job server.
>>> > > >
>>> > > > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels
>>> > <mxm@apache.org <ma...@apache.org>
>>> > > > <mailto:mxm@apache.org <ma...@apache.org>>> wrote:
>>> > > >
>>> > > > > I would prefer we don't introduce a (quirky) way of
>>> passing
>>> > > > unknown options that forces users to type JSON into the
>>> > command line
>>> > > > (or similar acrobatics)
>>> > > > Same here, the JSON approach seems technically nice but
>>> too
>>> > bulky
>>> > > > for users.
>>> > > >
>>> > > > > To someone wanting to run a pipeline, all options are
>>> > equally
>>> > > > important, whether they are application specific, SDK
>>> > specific or
>>> > > > runner specific.
>>> > > >
>>> > > > I'm also reluctant to force users to use
>>> `--runner_option=`
>>> > because the
>>> > > > division into "Runner" options and other options seems
>>> > rather arbitrary
>>> > > > to users. Most built-in options are also Runner-related.
>>> > > >
>>> > > > > It should be possible to *optionally* qualify/scope (to
>>> > cover
>>> > > > cases where there is ambiguity), but otherwise I prefer
>>> the
>>> > format
>>> > > > we currently have.
>>> > > >
>>> > > > Yes, namespacing is a problem. What happens if the user
>>> > defines a
>>> > > > custom
>>> > > > PipelineOption which clashes with one of the builtin ones?
>>> > If both are
>>> > > >
>>> > > > set, which one is actually applied?
>>> > > >
>>> > > >
>>> > > > Note that PipelineOptions so far has been treating name
>>> > equality to mean
>>> > > > option equality and the Java implementation has a bunch of
>>> > strict checks
>>> > > > to make sure that default values aren't used for duplicate
>>> > definitions,
>>> > > > they have the same type, etc...
>>> > > > With 1), you fail the job if the runner can't understand your
>>> > option
>>> > > > because its not represented the same way. User then needs to
>>> fix-up
>>> > > > their declaration of the option name.
>>> > > > With 2), there are no name conflicts, the SDK will need to
>>> > validate that
>>> > > > the option isn't set in both formats and error out if it is
>>> before
>>> > > > pipeline submission time.
>>> > > > With 3), you can prefetch all the options and error out to
>>> the user
>>> > > > during argument parsing time.
>>> > > >
>>> > > >
>>> > > >
>>> > > > Here is a summary of the possible paths going forward:
>>> > > >
>>> > > >
>>> > > > 1) Validate PipelineOptions at Runner side
>>> > > > ==========================================
>>> > > >
>>> > > > The main issue raised here was that we want to move away
>>> > from parsing
>>> > > > arguments which look like options without validating them.
>>> > An easy fix
>>> > > > would be to actually validate them on the Runner side.
>>> This
>>> > could be
>>> > > > done by changing the deserialization code of
>>> > PipelineOptions which so
>>> > > > far ignores unknown JSON options.
>>> > > >
>>> > > > See: PipelineOptionsTranslation.fromProto(Struct
>>> protoOptions)
>>> > > >
>>> > > > Actually, this wouldn't work for user-defined
>>> > PipelineOptions because
>>> > > > they might not be known to the Runner (if they are defined
>>> > in Python).
>>> > > >
>>> > > >
>>> > > > 2) Introduce a Runner-Option Flag
>>> > > > =================================
>>> > > >
>>> > > > In this approach we would try to add as many pipeline
>>> > options for a
>>> > > > Runner to the SDK, but allow additional Runner options to
>>> > be passed
>>> > > > using the `--runner-option=key=val` flag. The Runner, like
>>> > in 1), would
>>> > > > have to ensure validation. I think this has been the most
>>> > favored
>>> > > > way so
>>> > > > far. Going forward, that means that `--parallelism=4` and
>>> > > > `--runner-option=parallelism=4` will have the same effect
>>> > for the Flink
>>> > > > Runner.
>>> > > >
>>> > > >
>>> > > > 3) Implement Fetching of Options from JobServer
>>> > > > ===============================================
>>> > > >
>>> > > > The options are retrieved from the JobServer before
>>> > submitting the
>>> > > > pipeline. I think this would be ideal but, as mentioned
>>> > before, it
>>> > > > increases the complexity for implementing new SDKs and
>>> > might overall
>>> > > > just not be worth the effort.
>>> > > >
>>> > > >
>>> > > > What do you think? I'd implement 2) for the next release,
>>> > unless there
>>> > > > are advocates for a different approach.
>>> > > >
>>> > > > Cheers,
>>> > > > Max
>>> >
>>>
>>
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list argument
Posted by Thomas Weise <th...@apache.org>.
Discovering options from the job server would be the only way to perform
full validation (and provide upfront help to the user).
The runner cannot perform full validation, since it is not aware of the
user and SDK options (that it has to forward to the SDK worker).
Special runner options flag to forward unknown options also wouldn't fully
solve the problem (besides being subject to change in the future). Let's
say runner understands --fancy-int-option and the user repeats that option
multiple times. Not knowing the type of option, the SDK will pass it as a
list and the runner will fail.
Replicating SDK options is a workaround for known runners but it really
goes against the idea of portability (making assumptions outside the API
contract). We already have runners implemented outside of Beam and hope for
the ecosystem to grow. What we do for options should also work for those
runners.
I'm with Luke here that options discovery provides the best user experience
and can address the other issues. Even the scenario of multiple
intermediate runners could be addressed by forwarding the unparsed options
with the discovery call. I don't see SDK implementation complexity as a
significant drawback so far.
Thomas
On Mon, Nov 12, 2018 at 2:30 PM Lukasz Cwik <lc...@google.com> wrote:
>
> On Mon, Nov 12, 2018 at 9:38 AM Maximilian Michels <mx...@apache.org> wrote:
>
>> Thank you Robert and Lukasz for your points.
>>
>> > Note that I believe that we will want to have multiple URLs to support
>> cross language pipelines since we will want to be able to ask other SDK
>> languages/versions for their "list" of supported PipelineOptions.
>>
>> Why is that? The Runner itself is the source of truth for its options.
>>
>
> Because other languages (or even different versions of the same language)
> may have their own options. For example, the Go SDK talks to a Java service
> which applies a SQL transform and returns the expanded form (this may
> require knowledge of some options like credentials for the filesystem, ...)
> and then talks to a Python service that performs another transform
> expansion. Finally the pipeline containing Go, Java and Python transforms
> is submitted to a runner and it performs its own internal
> replacements/expansions related to executing the pipeline.
>
>
>> Everything else is SDK-related and should be validated there.
>>
>> I imagined the process to go like this:
>>
>> a) Parse options to find JobServer URL
>> a) Retrieve options from JobServer
>> c) Parse all options
>> ...continue as always...
>>
>> An option is just represented by a name and a type. There is nothing
>> more to it, at least as of now. So it should be possible to parse them
>> in the SDK without much further work.
>>
>> Nevertheless, I agree with your sentiment, Robert. The "runner_option"
>> flag would prevent additional complexity. I still don't prefer it
>> because it's not nice from an end user perspective. If we were to
>> implement it, I would definitely go for the "option promotion" which you
>> mentioned.
>>
>> I hadn't thought about delegating runners, although the PortableRunner
>> is basically a delegating Runner. If that was an important feature, I
>> suppose the "runner_option" would be the preferred way.
>>
>> All in all, since there doesn't seem to be an excitement to implement
>> JobServer option retrieval and we will need the help of all SDK
>> developers, "runner_option" seems to be the more likely path.
>>
>
> I would say its a lack of time for people to improve this story over
> others but it can be revisited at some point in the future and I agree that
> using --runner_option as an interim provides value.
>
>
>>
>> -Max
>>
>> On 08.11.18 21:50, Lukasz Cwik wrote:
>> > The purpose of the spec would be to provide the names, type and
>> > descriptions of the options. We don't need anything beyond the JSON
>> > types (string, number, bool, object, list) because the only ambiguity
>> we
>> > get is how do we parse command line string into the JSON type (and that
>> > ambiguity is actually only between string and non-string since all the
>> > other JSON types are unambiguous).
>> >
>> > Also, I believe the flow would be
>> > 1) Parse options
>> > a) Find the URL from args specified and/or additional methods on
>> > PipelineOptions that exposes a programmatic way to set the URL during
>> > parsing.
>> > b) Query URL for option specs
>> > c) Parse the remainder of the options
>> > 2) Construct pipeline
>> > 3) Choose runner
>> > 4) Submit job to runner
>> >
>> > Note that I believe that we will want to have multiple URLs to support
>> > cross language pipelines since we will want to be able to ask other SDK
>> > languages/versions for their "list" of supported PipelineOptions.
>> >
>> > On Thu, Nov 8, 2018 at 11:51 AM Robert Bradshaw <robertwb@google.com
>> > <ma...@google.com>> wrote:
>> >
>> > There's two questions here:
>> >
>> > (A) What do we do in the short term?
>> >
>> > I think adding every runner option to every SDK is not sustainable
>> > (n*m work, assuming every SDK knows about every runner), and having
>> a
>> > patchwork of options that were added as one-offs to SDKs is not
>> > desirable either. Furthermore, it seems difficult to parse unknown
>> > options as if they were valid options, so my preference here would
>> be
>> > to just use a special runner_option flag. (One could also pass a set
>> > of unparsed/unvalidated runner options to the runner, even if
>> they're
>> > not distinguished for the user, and runners (or any intermediates)
>> > could run a "promote" operation that promotes any of these unknowns
>> > that they recognize to real options before further processing. The
>> > parsing would be done as repeated-string, and not be intermingled
>> with
>> > the actually validated options. This is essential a variant of
>> > option 1.)
>> >
>> > (B) What do do in the long term? While the JobServer approach sounds
>> > nice, I think it introduces a lot of complexity (we have too much of
>> > that already) and still doesn't completely solve the problem. In
>> > particular, it changes the flow from
>> >
>> > 1. Parse options
>> > 2. Construct pipeline
>> > 3. Choose runner
>> > 4. Submit job to runner
>> >
>> > to
>> >
>> > 1. Parse options
>> > 2. Construct pipeline
>> > 3. Choose runner
>> > 4a. Query runner for option specs
>> > 4b. Re-parse options
>> > 4c. Submit job to runner
>> >
>> > In particular, doing 4b in the SDK rather than just let the runner
>> > itself do the validation as part of (4) doesn't save much and forces
>> > us to come up with a (probably incomplete) spec as to how to define
>> > options, their types, and their validations. It also means that a
>> > delegating runner must choose and interact with its downstream
>> > runner(s) synchronously, else we haven't actually solved the issue.
>> >
>> > For these reasons, I don't think we even want to go with the
>> JobServer
>> > approach in the long term, which has bearing on (A).
>> >
>> > - Robert
>> >
>> >
>> > On Wed, Nov 7, 2018 at 8:50 PM Maximilian Michels <mxm@apache.org
>> > <ma...@apache.org>> wrote:
>> > >
>> > > +1
>> > >
>> > > If the preferred approach is to eventually have the JobServer
>> > serve the
>> > > options, then the best intermediate solution is to replicate
>> common
>> > > options in the SDKs.
>> > >
>> > > If we went down the "--runner_option" path, we would end up with
>> > > multiple ways of specifying the same options. We would eventually
>> > have
>> > > to deprecate "runner options" once we have the JobServer
>> > approach. I'd
>> > > like to avoid that.
>> > >
>> > > For the upcoming release we can revert the changes again and add
>> the
>> > > most common missing options to the SDKs. Then hopefully we should
>> > have
>> > > fetching implemented for the release after.
>> > >
>> > > Do you think that is feasible?
>> > >
>> > > Thanks,
>> > > Max
>> > >
>> > > On 30.10.18 23:00, Lukasz Cwik wrote:
>> > > > I still like #3 the most, just can't devote the time to get it
>> > done.
>> > > >
>> > > > Instead of going with a fully implemented #3, we could hardcode
>> > the a
>> > > > subset of options and types within each SDK until the job
>> server is
>> > > > ready to provide this information and then migrate to the
>> > "full" list.
>> > > > This would be an easy path for SDKs to take on. They could
>> > "know" of a
>> > > > few well known options, and if they want to support all
>> > options, they
>> > > > implement the integration with the job server.
>> > > >
>> > > > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels
>> > <mxm@apache.org <ma...@apache.org>
>> > > > <mailto:mxm@apache.org <ma...@apache.org>>> wrote:
>> > > >
>> > > > > I would prefer we don't introduce a (quirky) way of
>> passing
>> > > > unknown options that forces users to type JSON into the
>> > command line
>> > > > (or similar acrobatics)
>> > > > Same here, the JSON approach seems technically nice but too
>> > bulky
>> > > > for users.
>> > > >
>> > > > > To someone wanting to run a pipeline, all options are
>> > equally
>> > > > important, whether they are application specific, SDK
>> > specific or
>> > > > runner specific.
>> > > >
>> > > > I'm also reluctant to force users to use `--runner_option=`
>> > because the
>> > > > division into "Runner" options and other options seems
>> > rather arbitrary
>> > > > to users. Most built-in options are also Runner-related.
>> > > >
>> > > > > It should be possible to *optionally* qualify/scope (to
>> > cover
>> > > > cases where there is ambiguity), but otherwise I prefer the
>> > format
>> > > > we currently have.
>> > > >
>> > > > Yes, namespacing is a problem. What happens if the user
>> > defines a
>> > > > custom
>> > > > PipelineOption which clashes with one of the builtin ones?
>> > If both are
>> > > >
>> > > > set, which one is actually applied?
>> > > >
>> > > >
>> > > > Note that PipelineOptions so far has been treating name
>> > equality to mean
>> > > > option equality and the Java implementation has a bunch of
>> > strict checks
>> > > > to make sure that default values aren't used for duplicate
>> > definitions,
>> > > > they have the same type, etc...
>> > > > With 1), you fail the job if the runner can't understand your
>> > option
>> > > > because its not represented the same way. User then needs to
>> fix-up
>> > > > their declaration of the option name.
>> > > > With 2), there are no name conflicts, the SDK will need to
>> > validate that
>> > > > the option isn't set in both formats and error out if it is
>> before
>> > > > pipeline submission time.
>> > > > With 3), you can prefetch all the options and error out to the
>> user
>> > > > during argument parsing time.
>> > > >
>> > > >
>> > > >
>> > > > Here is a summary of the possible paths going forward:
>> > > >
>> > > >
>> > > > 1) Validate PipelineOptions at Runner side
>> > > > ==========================================
>> > > >
>> > > > The main issue raised here was that we want to move away
>> > from parsing
>> > > > arguments which look like options without validating them.
>> > An easy fix
>> > > > would be to actually validate them on the Runner side. This
>> > could be
>> > > > done by changing the deserialization code of
>> > PipelineOptions which so
>> > > > far ignores unknown JSON options.
>> > > >
>> > > > See: PipelineOptionsTranslation.fromProto(Struct
>> protoOptions)
>> > > >
>> > > > Actually, this wouldn't work for user-defined
>> > PipelineOptions because
>> > > > they might not be known to the Runner (if they are defined
>> > in Python).
>> > > >
>> > > >
>> > > > 2) Introduce a Runner-Option Flag
>> > > > =================================
>> > > >
>> > > > In this approach we would try to add as many pipeline
>> > options for a
>> > > > Runner to the SDK, but allow additional Runner options to
>> > be passed
>> > > > using the `--runner-option=key=val` flag. The Runner, like
>> > in 1), would
>> > > > have to ensure validation. I think this has been the most
>> > favored
>> > > > way so
>> > > > far. Going forward, that means that `--parallelism=4` and
>> > > > `--runner-option=parallelism=4` will have the same effect
>> > for the Flink
>> > > > Runner.
>> > > >
>> > > >
>> > > > 3) Implement Fetching of Options from JobServer
>> > > > ===============================================
>> > > >
>> > > > The options are retrieved from the JobServer before
>> > submitting the
>> > > > pipeline. I think this would be ideal but, as mentioned
>> > before, it
>> > > > increases the complexity for implementing new SDKs and
>> > might overall
>> > > > just not be worth the effort.
>> > > >
>> > > >
>> > > > What do you think? I'd implement 2) for the next release,
>> > unless there
>> > > > are advocates for a different approach.
>> > > >
>> > > > Cheers,
>> > > > Max
>> >
>>
>
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list argument
Posted by Lukasz Cwik <lc...@google.com>.
On Mon, Nov 12, 2018 at 9:38 AM Maximilian Michels <mx...@apache.org> wrote:
> Thank you Robert and Lukasz for your points.
>
> > Note that I believe that we will want to have multiple URLs to support
> cross language pipelines since we will want to be able to ask other SDK
> languages/versions for their "list" of supported PipelineOptions.
>
> Why is that? The Runner itself is the source of truth for its options.
>
Because other languages (or even different versions of the same language)
may have their own options. For example, the Go SDK talks to a Java service
which applies a SQL transform and returns the expanded form (this may
require knowledge of some options like credentials for the filesystem, ...)
and then talks to a Python service that performs another transform
expansion. Finally the pipeline containing Go, Java and Python transforms
is submitted to a runner and it performs its own internal
replacements/expansions related to executing the pipeline.
> Everything else is SDK-related and should be validated there.
>
> I imagined the process to go like this:
>
> a) Parse options to find JobServer URL
> a) Retrieve options from JobServer
> c) Parse all options
> ...continue as always...
>
> An option is just represented by a name and a type. There is nothing
> more to it, at least as of now. So it should be possible to parse them
> in the SDK without much further work.
>
> Nevertheless, I agree with your sentiment, Robert. The "runner_option"
> flag would prevent additional complexity. I still don't prefer it
> because it's not nice from an end user perspective. If we were to
> implement it, I would definitely go for the "option promotion" which you
> mentioned.
>
> I hadn't thought about delegating runners, although the PortableRunner
> is basically a delegating Runner. If that was an important feature, I
> suppose the "runner_option" would be the preferred way.
>
> All in all, since there doesn't seem to be an excitement to implement
> JobServer option retrieval and we will need the help of all SDK
> developers, "runner_option" seems to be the more likely path.
>
I would say its a lack of time for people to improve this story over others
but it can be revisited at some point in the future and I agree that using
--runner_option as an interim provides value.
>
> -Max
>
> On 08.11.18 21:50, Lukasz Cwik wrote:
> > The purpose of the spec would be to provide the names, type and
> > descriptions of the options. We don't need anything beyond the JSON
> > types (string, number, bool, object, list) because the only ambiguity we
> > get is how do we parse command line string into the JSON type (and that
> > ambiguity is actually only between string and non-string since all the
> > other JSON types are unambiguous).
> >
> > Also, I believe the flow would be
> > 1) Parse options
> > a) Find the URL from args specified and/or additional methods on
> > PipelineOptions that exposes a programmatic way to set the URL during
> > parsing.
> > b) Query URL for option specs
> > c) Parse the remainder of the options
> > 2) Construct pipeline
> > 3) Choose runner
> > 4) Submit job to runner
> >
> > Note that I believe that we will want to have multiple URLs to support
> > cross language pipelines since we will want to be able to ask other SDK
> > languages/versions for their "list" of supported PipelineOptions.
> >
> > On Thu, Nov 8, 2018 at 11:51 AM Robert Bradshaw <robertwb@google.com
> > <ma...@google.com>> wrote:
> >
> > There's two questions here:
> >
> > (A) What do we do in the short term?
> >
> > I think adding every runner option to every SDK is not sustainable
> > (n*m work, assuming every SDK knows about every runner), and having a
> > patchwork of options that were added as one-offs to SDKs is not
> > desirable either. Furthermore, it seems difficult to parse unknown
> > options as if they were valid options, so my preference here would be
> > to just use a special runner_option flag. (One could also pass a set
> > of unparsed/unvalidated runner options to the runner, even if they're
> > not distinguished for the user, and runners (or any intermediates)
> > could run a "promote" operation that promotes any of these unknowns
> > that they recognize to real options before further processing. The
> > parsing would be done as repeated-string, and not be intermingled
> with
> > the actually validated options. This is essential a variant of
> > option 1.)
> >
> > (B) What do do in the long term? While the JobServer approach sounds
> > nice, I think it introduces a lot of complexity (we have too much of
> > that already) and still doesn't completely solve the problem. In
> > particular, it changes the flow from
> >
> > 1. Parse options
> > 2. Construct pipeline
> > 3. Choose runner
> > 4. Submit job to runner
> >
> > to
> >
> > 1. Parse options
> > 2. Construct pipeline
> > 3. Choose runner
> > 4a. Query runner for option specs
> > 4b. Re-parse options
> > 4c. Submit job to runner
> >
> > In particular, doing 4b in the SDK rather than just let the runner
> > itself do the validation as part of (4) doesn't save much and forces
> > us to come up with a (probably incomplete) spec as to how to define
> > options, their types, and their validations. It also means that a
> > delegating runner must choose and interact with its downstream
> > runner(s) synchronously, else we haven't actually solved the issue.
> >
> > For these reasons, I don't think we even want to go with the
> JobServer
> > approach in the long term, which has bearing on (A).
> >
> > - Robert
> >
> >
> > On Wed, Nov 7, 2018 at 8:50 PM Maximilian Michels <mxm@apache.org
> > <ma...@apache.org>> wrote:
> > >
> > > +1
> > >
> > > If the preferred approach is to eventually have the JobServer
> > serve the
> > > options, then the best intermediate solution is to replicate
> common
> > > options in the SDKs.
> > >
> > > If we went down the "--runner_option" path, we would end up with
> > > multiple ways of specifying the same options. We would eventually
> > have
> > > to deprecate "runner options" once we have the JobServer
> > approach. I'd
> > > like to avoid that.
> > >
> > > For the upcoming release we can revert the changes again and add
> the
> > > most common missing options to the SDKs. Then hopefully we should
> > have
> > > fetching implemented for the release after.
> > >
> > > Do you think that is feasible?
> > >
> > > Thanks,
> > > Max
> > >
> > > On 30.10.18 23:00, Lukasz Cwik wrote:
> > > > I still like #3 the most, just can't devote the time to get it
> > done.
> > > >
> > > > Instead of going with a fully implemented #3, we could hardcode
> > the a
> > > > subset of options and types within each SDK until the job
> server is
> > > > ready to provide this information and then migrate to the
> > "full" list.
> > > > This would be an easy path for SDKs to take on. They could
> > "know" of a
> > > > few well known options, and if they want to support all
> > options, they
> > > > implement the integration with the job server.
> > > >
> > > > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels
> > <mxm@apache.org <ma...@apache.org>
> > > > <mailto:mxm@apache.org <ma...@apache.org>>> wrote:
> > > >
> > > > > I would prefer we don't introduce a (quirky) way of
> passing
> > > > unknown options that forces users to type JSON into the
> > command line
> > > > (or similar acrobatics)
> > > > Same here, the JSON approach seems technically nice but too
> > bulky
> > > > for users.
> > > >
> > > > > To someone wanting to run a pipeline, all options are
> > equally
> > > > important, whether they are application specific, SDK
> > specific or
> > > > runner specific.
> > > >
> > > > I'm also reluctant to force users to use `--runner_option=`
> > because the
> > > > division into "Runner" options and other options seems
> > rather arbitrary
> > > > to users. Most built-in options are also Runner-related.
> > > >
> > > > > It should be possible to *optionally* qualify/scope (to
> > cover
> > > > cases where there is ambiguity), but otherwise I prefer the
> > format
> > > > we currently have.
> > > >
> > > > Yes, namespacing is a problem. What happens if the user
> > defines a
> > > > custom
> > > > PipelineOption which clashes with one of the builtin ones?
> > If both are
> > > >
> > > > set, which one is actually applied?
> > > >
> > > >
> > > > Note that PipelineOptions so far has been treating name
> > equality to mean
> > > > option equality and the Java implementation has a bunch of
> > strict checks
> > > > to make sure that default values aren't used for duplicate
> > definitions,
> > > > they have the same type, etc...
> > > > With 1), you fail the job if the runner can't understand your
> > option
> > > > because its not represented the same way. User then needs to
> fix-up
> > > > their declaration of the option name.
> > > > With 2), there are no name conflicts, the SDK will need to
> > validate that
> > > > the option isn't set in both formats and error out if it is
> before
> > > > pipeline submission time.
> > > > With 3), you can prefetch all the options and error out to the
> user
> > > > during argument parsing time.
> > > >
> > > >
> > > >
> > > > Here is a summary of the possible paths going forward:
> > > >
> > > >
> > > > 1) Validate PipelineOptions at Runner side
> > > > ==========================================
> > > >
> > > > The main issue raised here was that we want to move away
> > from parsing
> > > > arguments which look like options without validating them.
> > An easy fix
> > > > would be to actually validate them on the Runner side. This
> > could be
> > > > done by changing the deserialization code of
> > PipelineOptions which so
> > > > far ignores unknown JSON options.
> > > >
> > > > See: PipelineOptionsTranslation.fromProto(Struct
> protoOptions)
> > > >
> > > > Actually, this wouldn't work for user-defined
> > PipelineOptions because
> > > > they might not be known to the Runner (if they are defined
> > in Python).
> > > >
> > > >
> > > > 2) Introduce a Runner-Option Flag
> > > > =================================
> > > >
> > > > In this approach we would try to add as many pipeline
> > options for a
> > > > Runner to the SDK, but allow additional Runner options to
> > be passed
> > > > using the `--runner-option=key=val` flag. The Runner, like
> > in 1), would
> > > > have to ensure validation. I think this has been the most
> > favored
> > > > way so
> > > > far. Going forward, that means that `--parallelism=4` and
> > > > `--runner-option=parallelism=4` will have the same effect
> > for the Flink
> > > > Runner.
> > > >
> > > >
> > > > 3) Implement Fetching of Options from JobServer
> > > > ===============================================
> > > >
> > > > The options are retrieved from the JobServer before
> > submitting the
> > > > pipeline. I think this would be ideal but, as mentioned
> > before, it
> > > > increases the complexity for implementing new SDKs and
> > might overall
> > > > just not be worth the effort.
> > > >
> > > >
> > > > What do you think? I'd implement 2) for the next release,
> > unless there
> > > > are advocates for a different approach.
> > > >
> > > > Cheers,
> > > > Max
> >
>
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list
argument
Posted by Maximilian Michels <mx...@apache.org>.
Thank you Robert and Lukasz for your points.
> Note that I believe that we will want to have multiple URLs to support cross language pipelines since we will want to be able to ask other SDK languages/versions for their "list" of supported PipelineOptions.
Why is that? The Runner itself is the source of truth for its options.
Everything else is SDK-related and should be validated there.
I imagined the process to go like this:
a) Parse options to find JobServer URL
a) Retrieve options from JobServer
c) Parse all options
...continue as always...
An option is just represented by a name and a type. There is nothing
more to it, at least as of now. So it should be possible to parse them
in the SDK without much further work.
Nevertheless, I agree with your sentiment, Robert. The "runner_option"
flag would prevent additional complexity. I still don't prefer it
because it's not nice from an end user perspective. If we were to
implement it, I would definitely go for the "option promotion" which you
mentioned.
I hadn't thought about delegating runners, although the PortableRunner
is basically a delegating Runner. If that was an important feature, I
suppose the "runner_option" would be the preferred way.
All in all, since there doesn't seem to be an excitement to implement
JobServer option retrieval and we will need the help of all SDK
developers, "runner_option" seems to be the more likely path.
-Max
On 08.11.18 21:50, Lukasz Cwik wrote:
> The purpose of the spec would be to provide the names, type and
> descriptions of the options. We don't need anything beyond the JSON
> types (string, number, bool, object, list) because the only ambiguity we
> get is how do we parse command line string into the JSON type (and that
> ambiguity is actually only between string and non-string since all the
> other JSON types are unambiguous).
>
> Also, I believe the flow would be
> 1) Parse options
> a) Find the URL from args specified and/or additional methods on
> PipelineOptions that exposes a programmatic way to set the URL during
> parsing.
> b) Query URL for option specs
> c) Parse the remainder of the options
> 2) Construct pipeline
> 3) Choose runner
> 4) Submit job to runner
>
> Note that I believe that we will want to have multiple URLs to support
> cross language pipelines since we will want to be able to ask other SDK
> languages/versions for their "list" of supported PipelineOptions.
>
> On Thu, Nov 8, 2018 at 11:51 AM Robert Bradshaw <robertwb@google.com
> <ma...@google.com>> wrote:
>
> There's two questions here:
>
> (A) What do we do in the short term?
>
> I think adding every runner option to every SDK is not sustainable
> (n*m work, assuming every SDK knows about every runner), and having a
> patchwork of options that were added as one-offs to SDKs is not
> desirable either. Furthermore, it seems difficult to parse unknown
> options as if they were valid options, so my preference here would be
> to just use a special runner_option flag. (One could also pass a set
> of unparsed/unvalidated runner options to the runner, even if they're
> not distinguished for the user, and runners (or any intermediates)
> could run a "promote" operation that promotes any of these unknowns
> that they recognize to real options before further processing. The
> parsing would be done as repeated-string, and not be intermingled with
> the actually validated options. This is essential a variant of
> option 1.)
>
> (B) What do do in the long term? While the JobServer approach sounds
> nice, I think it introduces a lot of complexity (we have too much of
> that already) and still doesn't completely solve the problem. In
> particular, it changes the flow from
>
> 1. Parse options
> 2. Construct pipeline
> 3. Choose runner
> 4. Submit job to runner
>
> to
>
> 1. Parse options
> 2. Construct pipeline
> 3. Choose runner
> 4a. Query runner for option specs
> 4b. Re-parse options
> 4c. Submit job to runner
>
> In particular, doing 4b in the SDK rather than just let the runner
> itself do the validation as part of (4) doesn't save much and forces
> us to come up with a (probably incomplete) spec as to how to define
> options, their types, and their validations. It also means that a
> delegating runner must choose and interact with its downstream
> runner(s) synchronously, else we haven't actually solved the issue.
>
> For these reasons, I don't think we even want to go with the JobServer
> approach in the long term, which has bearing on (A).
>
> - Robert
>
>
> On Wed, Nov 7, 2018 at 8:50 PM Maximilian Michels <mxm@apache.org
> <ma...@apache.org>> wrote:
> >
> > +1
> >
> > If the preferred approach is to eventually have the JobServer
> serve the
> > options, then the best intermediate solution is to replicate common
> > options in the SDKs.
> >
> > If we went down the "--runner_option" path, we would end up with
> > multiple ways of specifying the same options. We would eventually
> have
> > to deprecate "runner options" once we have the JobServer
> approach. I'd
> > like to avoid that.
> >
> > For the upcoming release we can revert the changes again and add the
> > most common missing options to the SDKs. Then hopefully we should
> have
> > fetching implemented for the release after.
> >
> > Do you think that is feasible?
> >
> > Thanks,
> > Max
> >
> > On 30.10.18 23:00, Lukasz Cwik wrote:
> > > I still like #3 the most, just can't devote the time to get it
> done.
> > >
> > > Instead of going with a fully implemented #3, we could hardcode
> the a
> > > subset of options and types within each SDK until the job server is
> > > ready to provide this information and then migrate to the
> "full" list.
> > > This would be an easy path for SDKs to take on. They could
> "know" of a
> > > few well known options, and if they want to support all
> options, they
> > > implement the integration with the job server.
> > >
> > > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels
> <mxm@apache.org <ma...@apache.org>
> > > <mailto:mxm@apache.org <ma...@apache.org>>> wrote:
> > >
> > > > I would prefer we don't introduce a (quirky) way of passing
> > > unknown options that forces users to type JSON into the
> command line
> > > (or similar acrobatics)
> > > Same here, the JSON approach seems technically nice but too
> bulky
> > > for users.
> > >
> > > > To someone wanting to run a pipeline, all options are
> equally
> > > important, whether they are application specific, SDK
> specific or
> > > runner specific.
> > >
> > > I'm also reluctant to force users to use `--runner_option=`
> because the
> > > division into "Runner" options and other options seems
> rather arbitrary
> > > to users. Most built-in options are also Runner-related.
> > >
> > > > It should be possible to *optionally* qualify/scope (to
> cover
> > > cases where there is ambiguity), but otherwise I prefer the
> format
> > > we currently have.
> > >
> > > Yes, namespacing is a problem. What happens if the user
> defines a
> > > custom
> > > PipelineOption which clashes with one of the builtin ones?
> If both are
> > >
> > > set, which one is actually applied?
> > >
> > >
> > > Note that PipelineOptions so far has been treating name
> equality to mean
> > > option equality and the Java implementation has a bunch of
> strict checks
> > > to make sure that default values aren't used for duplicate
> definitions,
> > > they have the same type, etc...
> > > With 1), you fail the job if the runner can't understand your
> option
> > > because its not represented the same way. User then needs to fix-up
> > > their declaration of the option name.
> > > With 2), there are no name conflicts, the SDK will need to
> validate that
> > > the option isn't set in both formats and error out if it is before
> > > pipeline submission time.
> > > With 3), you can prefetch all the options and error out to the user
> > > during argument parsing time.
> > >
> > >
> > >
> > > Here is a summary of the possible paths going forward:
> > >
> > >
> > > 1) Validate PipelineOptions at Runner side
> > > ==========================================
> > >
> > > The main issue raised here was that we want to move away
> from parsing
> > > arguments which look like options without validating them.
> An easy fix
> > > would be to actually validate them on the Runner side. This
> could be
> > > done by changing the deserialization code of
> PipelineOptions which so
> > > far ignores unknown JSON options.
> > >
> > > See: PipelineOptionsTranslation.fromProto(Struct protoOptions)
> > >
> > > Actually, this wouldn't work for user-defined
> PipelineOptions because
> > > they might not be known to the Runner (if they are defined
> in Python).
> > >
> > >
> > > 2) Introduce a Runner-Option Flag
> > > =================================
> > >
> > > In this approach we would try to add as many pipeline
> options for a
> > > Runner to the SDK, but allow additional Runner options to
> be passed
> > > using the `--runner-option=key=val` flag. The Runner, like
> in 1), would
> > > have to ensure validation. I think this has been the most
> favored
> > > way so
> > > far. Going forward, that means that `--parallelism=4` and
> > > `--runner-option=parallelism=4` will have the same effect
> for the Flink
> > > Runner.
> > >
> > >
> > > 3) Implement Fetching of Options from JobServer
> > > ===============================================
> > >
> > > The options are retrieved from the JobServer before
> submitting the
> > > pipeline. I think this would be ideal but, as mentioned
> before, it
> > > increases the complexity for implementing new SDKs and
> might overall
> > > just not be worth the effort.
> > >
> > >
> > > What do you think? I'd implement 2) for the next release,
> unless there
> > > are advocates for a different approach.
> > >
> > > Cheers,
> > > Max
>
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list argument
Posted by Lukasz Cwik <lc...@google.com>.
I'm not sure how feasible fetching is because no one has said they will do
the work.
I can not pick it up as I am heads down in splittable dofn right now and
integrating that into portability.
On Wed, Nov 7, 2018 at 11:50 AM Maximilian Michels <mx...@apache.org> wrote:
> +1
>
> If the preferred approach is to eventually have the JobServer serve the
> options, then the best intermediate solution is to replicate common
> options in the SDKs.
>
> If we went down the "--runner_option" path, we would end up with
> multiple ways of specifying the same options. We would eventually have
> to deprecate "runner options" once we have the JobServer approach. I'd
> like to avoid that.
>
> For the upcoming release we can revert the changes again and add the
> most common missing options to the SDKs. Then hopefully we should have
> fetching implemented for the release after.
>
> Do you think that is feasible?
>
> Thanks,
> Max
>
> On 30.10.18 23:00, Lukasz Cwik wrote:
> > I still like #3 the most, just can't devote the time to get it done.
> >
> > Instead of going with a fully implemented #3, we could hardcode the a
> > subset of options and types within each SDK until the job server is
> > ready to provide this information and then migrate to the "full" list.
> > This would be an easy path for SDKs to take on. They could "know" of a
> > few well known options, and if they want to support all options, they
> > implement the integration with the job server.
> >
> > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels <mxm@apache.org
> > <ma...@apache.org>> wrote:
> >
> > > I would prefer we don't introduce a (quirky) way of passing
> > unknown options that forces users to type JSON into the command line
> > (or similar acrobatics)
> > Same here, the JSON approach seems technically nice but too bulky
> > for users.
> >
> > > To someone wanting to run a pipeline, all options are equally
> > important, whether they are application specific, SDK specific or
> > runner specific.
> >
> > I'm also reluctant to force users to use `--runner_option=` because
> the
> > division into "Runner" options and other options seems rather
> arbitrary
> > to users. Most built-in options are also Runner-related.
> >
> > > It should be possible to *optionally* qualify/scope (to cover
> > cases where there is ambiguity), but otherwise I prefer the format
> > we currently have.
> >
> > Yes, namespacing is a problem. What happens if the user defines a
> > custom
> > PipelineOption which clashes with one of the builtin ones? If both
> are
> >
> > set, which one is actually applied?
> >
> >
> > Note that PipelineOptions so far has been treating name equality to mean
> > option equality and the Java implementation has a bunch of strict checks
> > to make sure that default values aren't used for duplicate definitions,
> > they have the same type, etc...
> > With 1), you fail the job if the runner can't understand your option
> > because its not represented the same way. User then needs to fix-up
> > their declaration of the option name.
> > With 2), there are no name conflicts, the SDK will need to validate that
> > the option isn't set in both formats and error out if it is before
> > pipeline submission time.
> > With 3), you can prefetch all the options and error out to the user
> > during argument parsing time.
> >
> >
> >
> > Here is a summary of the possible paths going forward:
> >
> >
> > 1) Validate PipelineOptions at Runner side
> > ==========================================
> >
> > The main issue raised here was that we want to move away from parsing
> > arguments which look like options without validating them. An easy
> fix
> > would be to actually validate them on the Runner side. This could be
> > done by changing the deserialization code of PipelineOptions which so
> > far ignores unknown JSON options.
> >
> > See: PipelineOptionsTranslation.fromProto(Struct protoOptions)
> >
> > Actually, this wouldn't work for user-defined PipelineOptions because
> > they might not be known to the Runner (if they are defined in
> Python).
> >
> >
> > 2) Introduce a Runner-Option Flag
> > =================================
> >
> > In this approach we would try to add as many pipeline options for a
> > Runner to the SDK, but allow additional Runner options to be passed
> > using the `--runner-option=key=val` flag. The Runner, like in 1),
> would
> > have to ensure validation. I think this has been the most favored
> > way so
> > far. Going forward, that means that `--parallelism=4` and
> > `--runner-option=parallelism=4` will have the same effect for the
> Flink
> > Runner.
> >
> >
> > 3) Implement Fetching of Options from JobServer
> > ===============================================
> >
> > The options are retrieved from the JobServer before submitting the
> > pipeline. I think this would be ideal but, as mentioned before, it
> > increases the complexity for implementing new SDKs and might overall
> > just not be worth the effort.
> >
> >
> > What do you think? I'd implement 2) for the next release, unless
> there
> > are advocates for a different approach.
> >
> > Cheers,
> > Max
> >
> > On 25.10.18 21:19, Thomas Weise wrote:
> > > Reminder that this is something we ideally address before the next
> > > release...
> > >
> > > Considering the discussion so far, my preference is that we get
> away
> > > from unknown options and discover valid options from the runner
> (by
> > > expanding the job service).
> > >
> > > Once the SDK is aware of all valid options, it is possible to
> > provide
> > > meaningful feedback to the user (validate or help), and correctly
> > handle
> > > scopes and types.
> > >
> > > I would prefer we don't introduce a (quirky) way of passing
> unknown
> > > options that forces users to type JSON into the command line (or
> > similar
> > > acrobatics). To someone wanting to run a pipeline, all options are
> > > equally important, whether they are application specific, SDK
> > specific
> > > or runner specific. It should be possible to *optionally*
> > qualify/scope
> > > (to cover cases where there is ambiguity), but otherwise I prefer
> > the
> > > format we currently have.
> > >
> > > Regarding type inference: Correct handling of numeric types
> > matters, see
> > > following issue with protobuf (not JSON):
> > > https://issues.apache.org/jira/browse/BEAM-5509
> > >
> > > Thomas
> > >
> > >
> > > On Thu, Oct 18, 2018 at 6:55 AM Robert Bradshaw
> > <robertwb@google.com <ma...@google.com>
> > > <mailto:robertwb@google.com <ma...@google.com>>> wrote:
> > >
> > > On Wed, Oct 17, 2018 at 11:35 PM Lukasz Cwik
> > <lcwik@google.com <ma...@google.com>
> > > <mailto:lcwik@google.com <ma...@google.com>>> wrote:
> > >
> > >
> > > On Tue, Oct 16, 2018 at 11:51 AM Robert Bradshaw
> > > <robertwb@google.com <ma...@google.com>
> > <mailto:robertwb@google.com <ma...@google.com>>> wrote:
> > >
> > > On Tue, Oct 16, 2018 at 7:03 PM Lukasz Cwik
> > > <lcwik@google.com <ma...@google.com>
> > <mailto:lcwik@google.com <ma...@google.com>>> wrote:
> > > >
> > > > For all unknown options, the SDK can require that
> all
> > > flag values be specified explicitly as a valid JSON
> type.
> > > > starts with { -> object
> > > > starts with [ -> list
> > > > starts with " -> string
> > > > is null / true / false -> null / true / false
> > > > otherwise is number.
> > > >
> > > > This isn't great for strings but works well for
> > all the
> > > other types.
> > > >
> > > > Thus for known options, the additional typing
> > information
> > > would disambiguate whether something should be a
> > > string/boolean/number/object/list but for unknown
> > options we
> > > would expect the user to use valid JSON explicitly
> > and write:
> > > > --foo={"object": "value"}
> > > > --foo=["value", "value2"]
> > > > --foo="string value"
> > >
> > > Due to shell escaping, one would have to write
> > >
> > > --foo=\"string value\"
> > >
> > > or actually, due to the space
> > >
> > > --foo='"string value"'
> > >
> > > or some other variation on that, which is really
> > > unfortunate. (The JSON list/objects would need similar
> > > quoting, but that's less surprising.) Also, does this
> > mean
> > > we'd only have one kind of number (not integer vs.
> float,
> > > i.e. --parallelism=5.0 works)? I suppose that is JSON.
> > >
> > >
> > > Yes, I was suspecting that users would need to type the
> > second
> > > variant as \"...\" I found more burdensome then '"..."'
> > >
> > >
> > > > --foo=3.5 --foo=-4
> > > > --foo=true --foo=false
> > > > --foo=null
> > > > This also works if the flag is repeated, so
> --foo=3.5
> > > --foo=-4 is [3.5, -4]
> > >
> > > The thing that sparked this discussion was what to do
> > when
> > > unknown foo is repeated, but only one value given.
> > >
> > >
> > > If the person only specifies one value, then they have to
> > > disambiguate and put it in a list, only if they specify
> more
> > > then one value will they have to turn it into a list.
> > >
> > > I believe we could come up with other schemes on how to
> > convert
> > > unknown options to JSON where we prefer strings over
> > non-string
> > > types like null/true/false/numbers/list/object and
> > require the
> > > user to escape out of the string default but anything
> that is
> > > too different from strict JSON would cause headaches when
> > > attempting to explain the format to users. I think a happy
> > > middle ground would be that we will only require escaping
> for
> > > strings which are ambiguous, so things like true, null,
> > false,
> > > ... to be treated as strings would require the user to
> > escape them.
> > >
> > >
> > > I'd prefer to avoid inferring the type of an unknown argument
> > based
> > > on its contents, which can lead to surprises. We could
> > declare every
> > > unknown type to be repeated string, and let any
> > parsing/validation
> > > occur on the runner. If desired, we could pass these around
> as a
> > > single "runner options" dict that runners could inspect and
> > use to
> > > populate the actual dict rather than mixing parsed and
> unparsed
> > > options.
> > >
> > >
> > >
> > > > On Tue, Oct 16, 2018 at 7:56 AM Thomas Weise
> > > <thw@apache.org <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>> wrote:
> > > >>
> > > >> Discovering options from the job server seems
> > preferable
> > > over replicating runner options in SDKs.
> > > >>
> > > >> Runners evolve on their own, and with portability
> the
> > > SDK does not need to know anything about the runner.
> > > >>
> > > >> Regarding --runner-option. It is true that this
> looks
> > > less user friendly. On the other hand it eliminates
> the
> > > possibility of name collisions.
> > > >>
> > > >> But if options are discovered, the SDK can
> > perform full
> > > validation. It would only be necessary to use explicit
> > > scoping when there is ambiguity.
> > > >>
> > > >> Thomas
> > > >>
> > > >>
> > > >> On Tue, Oct 16, 2018 at 3:48 AM Maximilian Michels
> > > <mxm@apache.org <ma...@apache.org>
> > <mailto:mxm@apache.org <ma...@apache.org>>> wrote:
> > > >>>
> > > >>> Fetching options directly from the Runner's
> > JobServer
> > > seems like the
> > > >>> ideal solution. I agree with Robert that it
> creates
> > > additional
> > > >>> complexity for SDK authors, so the
> `--runner-option`
> > > flag would be an
> > > >>> easy and explicit way to specify additional
> > Runner options.
> > > >>>
> > > >>> The format I prefer would be:
> > --runner_option=key1=val1
> > > >>> --runner_option=key2=val2
> > > >>>
> > > >>> Now, from the perspective of end users, I think
> > it is
> > > neither convenient
> > > >>> nor reasonable to require the use of the
> > > `--runner-option` flag. To the
> > > >>> user it seems nebulous why some pipeline options
> > live
> > > in the top-level
> > > >>> option namespace while others need to be nested
> > within
> > > an option. This
> > > >>> is amplified by there being two Runners the user
> > needs
> > > to be aware of,
> > > >>> i.e. PortableRunner and the actual Runner
> > > (Dataflow/Flink/Spark..).
> > > >>>
> > > >>> I feel like we would eventually replicate all
> > options
> > > in the SDK because
> > > >>> otherwise users have to use the
> > `--runner-option`, but
> > > at least we can
> > > >>> specify options which have not been replicated
> yet.
> > > >>>
> > > >>> -Max
> > > >>>
> > > >>> On 16.10.18 10:27, Robert Bradshaw wrote:
> > > >>> > Yes, we don't know how to parse and/or
> > validate it.
> > > >>> >
> > > >>> > On Tue, Oct 16, 2018 at 1:14 AM Lukasz Cwik
> > > <lcwik@google.com <ma...@google.com>
> > <mailto:lcwik@google.com <ma...@google.com>>
> > > >>> > <mailto:lcwik@google.com
> > <ma...@google.com> <mailto:lcwik@google.com
> > <ma...@google.com>>>>
> > > wrote:
> > > >>> >
> > > >>> > I see, is the issue that we currently are
> > using a
> > > JSON
> > > >>> > representation for options when being
> > serialized
> > > and when we get
> > > >>> > some unknown option, we don't know how to
> > convert
> > > it into its JSON form?
> > > >>> >
> > > >>> > On Mon, Oct 15, 2018 at 2:41 PM Robert
> > Bradshaw
> > > <robertwb@google.com <ma...@google.com>
> > <mailto:robertwb@google.com <ma...@google.com>>
> > > >>> > <mailto:robertwb@google.com
> > <ma...@google.com>
> > > <mailto:robertwb@google.com
> > <ma...@google.com>>>> wrote:
> > > >>> >
> > > >>> > On Mon, Oct 15, 2018 at 11:30 PM
> > Lukasz Cwik
> > > <lcwik@google.com <ma...@google.com>
> > <mailto:lcwik@google.com <ma...@google.com>>
> > > >>> > <mailto:lcwik@google.com
> > <ma...@google.com>
> > > <mailto:lcwik@google.com <ma...@google.com>>>>
> > wrote:
> > > >>> > >
> > > >>> > > On Mon, Oct 15, 2018 at 1:17 PM
> Robert
> > > Bradshaw
> > > >>> > <robertwb@google.com
> > <ma...@google.com>
> > > <mailto:robertwb@google.com
> > <ma...@google.com>> <mailto:robertwb@google.com
> > <ma...@google.com>
> > > <mailto:robertwb@google.com
> > <ma...@google.com>>>> wrote:
> > > >>> > >>
> > > >>> > >> On Mon, Oct 15, 2018 at 7:50 PM
> > Lukasz Cwik
> > > >>> > <lcwik@google.com
> > <ma...@google.com> <mailto:lcwik@google.com
> > <ma...@google.com>>
> > > <mailto:lcwik@google.com <ma...@google.com>
> > <mailto:lcwik@google.com <ma...@google.com>>>> wrote:
> > > >>> > >> >
> > > >>> > >> > I agree with the sentiment for
> > better
> > > error checking.
> > > >>> > >> >
> > > >>> > >> > We can try to make it such that
> > the SDK
> > > can "fetch" the
> > > >>> > set of options that the runner
> supports by
> > > making a call to the
> > > >>> > Job API. The API could return a list of
> > > option names
> > > >>> > (descriptions for --help purposes and
> also
> > > potentially the
> > > >>> > expected format) which would remove
> > the worry
> > > around "unknown"
> > > >>> > options. Yes I understand to be able
> > to make
> > > the Job API call,
> > > >>> > we may need to parse some options from
> the
> > > args parameters first
> > > >>> > and then parse the unknown options
> > after they
> > > are fetched.
> > > >>> > >>
> > > >>> > >> This is an interesting idea, but
> > seems it
> > > could get quite
> > > >>> > complicated.
> > > >>> > >> E.g. for delegating runners, one
> would
> > > first read the options to
> > > >>> > >> determine which runner to fetch the
> > > options from, which
> > > >>> > would then
> > > >>> > >> return a set of options that
> possibly
> > > depends on the values
> > > >>> > of some of
> > > >>> > >> its options...
> > > >>> > >>
> > > >>> > >> > Alternatively, we can choose an
> > > explicit format upfront.
> > > >>> > >> > To expand on the exact format for
> > > --runner_option=...,
> > > >>> > here are some different ideas:
> > > >>> > >> > 1) Specified multiple times,
> > each one
> > > is an explicit flag
> > > >>> > >> > --runner_option=--blah=bar
> > > --runner_option=--foo=baz1
> > > >>> > --runner_option=--foo=baz2
> > > >>> > >>
> > > >>> > >> I'm -1 on this format. We should
> move
> > > away from the idea
> > > >>> > that options
> > > >>> > >> == flags (as that doesn't compose
> well
> > > with other libraries
> > > >>> > that do
> > > >>> > >> their own flags parsing). The
> > ability to
> > > parse a set of
> > > >>> > flags into
> > > >>> > >> options is just a convenience that
> an
> > > author may (or may
> > > >>> > not) choose
> > > >>> > >> to use (e.g. when running
> pipelines a
> > > long-lived process like a
> > > >>> > >> service or a notebook, the command
> > line
> > > flags are almost
> > > >>> > certainly not
> > > >>> > >> the right interface).
> > > >>> > >>
> > > >>> > >> > 2) specified multiple times, we
> drop
> > > the explicit flag
> > > >>> > >> > --runner_option=blah=bar
> > > --runner_option=foo=baz1
> > > >>> > --runner_option=foo=baz2
> > > >>> > >>
> > > >>> > >> This or (4) is my preference.
> > > >>> > >>
> > > >>> > >> > 3) we use a string which the
> > runner can
> > > choose to
> > > >>> > interpret however they want (JSON/XML
> > shown
> > > below)
> > > >>> > >> > --runner_option='{"blah": "bar",
> > "foo":
> > > ["baz1", "baz2"]}'
> > > >>> > >> >
> > > >>> >
> > >
> >
> --runner_option='<options><blah>bar</blah><foo>baz1</foo><foo>baz2</foo></options>'
> > > >>> > >>
> > > >>> > >> This would make validation hard.
> > Also, I
> > > think it makes
> > > >>> > sense for some
> > > >>> > >> runner options to be "shared"
> > > (parallelism") by convention,
> > > >>> > so letting
> > > >>> > >> it be a free-form string wouldn't
> > allow
> > > different runners to
> > > >>> > inspect
> > > >>> > >> different bits.
> > > >>> > >>
> > > >>> > >> We should consider if we should
> > use urns
> > > for namespacing, and
> > > >>> > >> assigning semantic meaning to
> > strings, here.
> > > >>> > >>
> > > >>> > >> > 4) we use a string which must be
> a
> > > specific format such as
> > > >>> > JSON (allows the SDK to do simple
> > validation):
> > > >>> > >> > --runner_option='{"blah": "bar",
> > "foo":
> > > ["baz1", "baz2"]}'
> > > >>> > >>
> > > >>> > >> I like this in that at least some
> > > validation can be
> > > >>> > performed, and
> > > >>> > >> expectations of how to format
> richer
> > > types. On the other
> > > >>> > hand it gets
> > > >>> > >> a bit verbose, given that most (I'd
> > > imagine) options will be
> > > >>> > simple.
> > > >>> > >> As with normal options,
> > > >>> > >>
> > > >>> > >> --option1=value1
> --option2=value2
> > > >>> > >>
> > > >>> > >> is shorthand for {"option1":
> value1,
> > > "option2": value2}.
> > > >>> > >>
> > > >>> > > I lean to 4 the most. With 2, you
> > run into
> > > issues of what
> > > >>> > does --runner_option=foo=["a", "b"]
> > > --runner_option=foo=["c",
> > > >>> > "d"] mean?
> > > >>> > > Is it an error or list of lists or
> > > concatenated. Similar
> > > >>> > issues for map types represented via
> JSON
> > > object {...}
> > > >>> >
> > > >>> > We can err to be on the safe side
> > > unless/until an argument can
> > > >>> > be made
> > > >>> > that merging is more natural. I just
> think
> > > this will be excessively
> > > >>> > verbose to use.
> > > >>> >
> > > >>> > >> > I would strongly suggest that we
> go
> > > with the "fetch"
> > > >>> > approach, since this makes the set of
> > options
> > > discoverable and
> > > >>> > helps users find errors much earlier
> > in their
> > > pipeline.
> > > >>> > >>
> > > >>> > >> This seems like an advanced
> > feature that
> > > SDKs may want to
> > > >>> > support, but
> > > >>> > >> I wouldn't want to require this
> > > complexity for bootstrapping
> > > >>> > an SDK.
> > > >>> > >>
> > > >>> > > SDKs that are starting off wouldn't
> > need
> > > to "fetch" options,
> > > >>> > they could choose to not support runner
> > > options or they could
> > > >>> > choose to pass all options through to
> the
> > > runner blindly.
> > > >>> > Fetching the options only provides the
> SDK
> > > the ability to
> > > >>> > provide error checking upfront and
> useful
> > > error/help messages.
> > > >>> >
> > > >>> > But how to even pass all options
> through
> > > blindly is exactly the
> > > >>> > difficulty we're running into here.
> > > >>> >
> > > >>> > >> Regarding always keeping runner
> > options
> > > separate, +1, though
> > > >>> > I'm not
> > > >>> > >> sure the line is always clear.
> > > >>> > >>
> > > >>> > >>
> > > >>> > >> > On Mon, Oct 15, 2018 at 8:04 AM
> > Robert
> > > Bradshaw
> > > >>> > <robertwb@google.com
> > <ma...@google.com>
> > > <mailto:robertwb@google.com
> > <ma...@google.com>> <mailto:robertwb@google.com
> > <ma...@google.com>
> > > <mailto:robertwb@google.com
> > <ma...@google.com>>>> wrote:
> > > >>> > >> >>
> > > >>> > >> >> On Mon, Oct 15, 2018 at 3:58 PM
> > > Maximilian Michels
> > > >>> > <mxm@apache.org
> > <ma...@apache.org> <mailto:mxm@apache.org <mailto:
> mxm@apache.org>>
> > > <mailto:mxm@apache.org <ma...@apache.org>
> > <mailto:mxm@apache.org <ma...@apache.org>>>> wrote:
> > > >>> > >> >> >
> > > >>> > >> >> > I agree that the current
> approach
> > > breaks the pipeline
> > > >>> > options contract
> > > >>> > >> >> > because "unknown" options get
> > parsed
> > > in the same way as
> > > >>> > options which
> > > >>> > >> >> > have been defined by the user.
> > > >>> > >> >>
> > > >>> > >> >> FWIW, I think we're already
> > breaking
> > > this "contract."
> > > >>> > Unknown options
> > > >>> > >> >> are silently ignored; with this
> > change
> > > we just change how
> > > >>> > we record
> > > >>> > >> >> them. It still feels a bit
> > hacky though.
> > > >>> > >> >>
> > > >>> > >> >> > I'm not sure the
> > `experiments` flag
> > > works for us. AFAIK
> > > >>> > it only allows
> > > >>> > >> >> > true/false flags. We want to
> pass
> > > all types of pipeline
> > > >>> > options to the
> > > >>> > >> >> > Runner.
> > > >>> > >> >>
> > > >>> > >> >> Experiments is an arbitrary set
> of
> > > strings, which can be
> > > >>> > of the form
> > > >>> > >> >> "param=value" if that's useful.
> > > (Dataflow does this.)
> > > >>> > There is, again,
> > > >>> > >> >> no namespacing on the param
> > names, but
> > > we could user urns
> > > >>> > or impose
> > > >>> > >> >> some other structure here.
> > > >>> > >> >>
> > > >>> > >> >> > How to solve this?
> > > >>> > >> >> >
> > > >>> > >> >> > 1) Add all options of all
> > Runners to
> > > each SDK
> > > >>> > >> >> > We added some of the
> FlinkRunner
> > > options to the Python
> > > >>> > SDK but realized
> > > >>> > >> >> > syncing is rather cumbersome
> > in the
> > > long term. However,
> > > >>> > we want the most
> > > >>> > >> >> > important options to be
> > validated on
> > > the client side.
> > > >>> > >> >>
> > > >>> > >> >> I don't think this is
> > sustainable in
> > > the long run.
> > > >>> > However, thinking
> > > >>> > >> >> about this, in the worse case
> > > validation happens after
> > > >>> > construction
> > > >>> > >> >> but before execution (as with
> > much of
> > > our other
> > > >>> > validation) so it
> > > >>> > >> >> isn't that bad.
> > > >>> > >> >>
> > > >>> > >> >> > 2) Pass "unknown" options via
> a
> > > separate list in the
> > > >>> > Proto which can
> > > >>> > >> >> > only be accessed internally
> > by the
> > > Runners. This still
> > > >>> > allows passing
> > > >>> > >> >> > arbitrary options but we
> wouldn't
> > > leak unknown options
> > > >>> > and display them
> > > >>> > >> >> > as top-level options.
> > > >>> > >> >>
> > > >>> > >> >> I think there needs to be a way
> for
> > > the user to
> > > >>> > communicate values
> > > >>> > >> >> directly to the runner
> > regardless of
> > > the SDK. My
> > > >>> > preference would be
> > > >>> > >> >> to make this explicit, e.g.
> > (repeated)
> > > >>> > --runner_option=..., rather
> > > >>> > >> >> than scooping up all unknown
> > flags at
> > > command line
> > > >>> > parsing time.
> > > >>> > >> >> Perhaps an SDK that is aware of
> > some
> > > runners could choose
> > > >>> > to lift
> > > >>> > >> >> these as top-level options, but
> > still
> > > pass them as runner
> > > >>> > options.
> > > >>> > >> >>
> > > >>> > >> >> > On 13.10.18 02:34, Charles
> > Chen wrote:
> > > >>> > >> >> > > The current release branch
> > > >>> > >> >> > >
> > > >>> >
> > >
> > (https://github.com/apache/beam/commits/release-2.8.0) was cut
> > > >>> > after the
> > > >>> > >> >> > > revert went in. Sent out
> > > >>> > https://github.com/apache/beam/pull/6683 as a
> > > >>> > >> >> > > revert of the revert.
> > Regarding
> > > your comment above,
> > > >>> > I can help out with
> > > >>> > >> >> > > the design / PR reviews for
> > common
> > > Python code as you
> > > >>> > suggest.
> > > >>> > >> >> > >
> > > >>> > >> >> > > On Fri, Oct 12, 2018 at
> 4:48 PM
> > > Thomas Weise
> > > >>> > <thw@apache.org
> > <ma...@apache.org> <mailto:thw@apache.org <mailto:
> thw@apache.org>>
> > > <mailto:thw@apache.org <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>>
> > > >>> > >> >> > > <mailto:thw@apache.org
> > <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>
> > <mailto:thw@apache.org <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>>>>
> wrote:
> > > >>> > >> >> > >
> > > >>> > >> >> > > Thanks, will tag you and
> > > looking forward to
> > > >>> > feedback so we can
> > > >>> > >> >> > > ensure that changes
> > work for
> > > everyone.
> > > >>> > >> >> > >
> > > >>> > >> >> > > Looking at the PR, I see
> > > agreement from Max to
> > > >>> > revert the change on
> > > >>> > >> >> > > the release branch, but
> > not in
> > > master. Would you
> > > >>> > mind to restore it
> > > >>> > >> >> > > in master?
> > > >>> > >> >> > >
> > > >>> > >> >> > > Thanks
> > > >>> > >> >> > >
> > > >>> > >> >> > > On Fri, Oct 12, 2018 at
> > 4:40
> > > PM Ahmet Altay
> > > >>> > <altay@google.com
> > <ma...@google.com> <mailto:altay@google.com
> > <ma...@google.com>>
> > > <mailto:altay@google.com <ma...@google.com>
> > <mailto:altay@google.com <ma...@google.com>>>
> > > >>> > >> >> > >
> > <mailto:altay@google.com <ma...@google.com>
> > > <mailto:altay@google.com <ma...@google.com>>
> > > >>> > <mailto:altay@google.com
> > <ma...@google.com>
> > > <mailto:altay@google.com
> > <ma...@google.com>>>>> wrote:
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > > On Fri, Oct 12,
> 2018 at
> > > 11:31 AM, Charles
> > > >>> > Chen <ccy@google.com
> > <ma...@google.com> <mailto:ccy@google.com <mailto:
> ccy@google.com>>
> > > <mailto:ccy@google.com <ma...@google.com>
> > <mailto:ccy@google.com <ma...@google.com>>>
> > > >>> > >> >> > >
> > <mailto:ccy@google.com <ma...@google.com>
> > > <mailto:ccy@google.com <ma...@google.com>>
> > > >>> > <mailto:ccy@google.com
> > <ma...@google.com>
> > > <mailto:ccy@google.com <ma...@google.com>>>>>
> wrote:
> > > >>> > >> >> > >
> > > >>> > >> >> > > What I mean is
> > that a
> > > user may find that
> > > >>> > it works for them
> > > >>> > >> >> > > to pass
> > "--myarg blah"
> > > and access it as
> > > >>> > "options.myarg"
> > > >>> > >> >> > > without
> explicitly
> > > defining a "my_arg"
> > > >>> > flag due to the added
> > > >>> > >> >> > > logic. This is
> not
> > > the intended behavior
> > > >>> > and we may want to
> > > >>> > >> >> > > change this
> > > implementation detail in the
> > > >>> > future. However,
> > > >>> > >> >> > > having this
> > logic in a
> > > released version
> > > >>> > makes it hard to
> > > >>> > >> >> > > change this
> > behavior
> > > since users may
> > > >>> > erroneously depend on
> > > >>> > >> >> > > this
> undocumented
> > > behavior. Instead, we
> > > >>> > should namespace /
> > > >>> > >> >> > > scope this so
> > that it
> > > is obvious that
> > > >>> > this is meant for
> > > >>> > >> >> > > runner (and not
> > Beam
> > > user) consumption.
> > > >>> > >> >> > >
> > > >>> > >> >> > > On Fri, Oct 12,
> > 2018
> > > at 10:48 AM Thomas Weise
> > > >>> > >> >> > > <thw@apache.org
> > <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>
> > <mailto:thw@apache.org <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>>
> > > >>> > <mailto:thw@apache.org
> > <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>
> > <mailto:thw@apache.org <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>>>>
> wrote:
> > > >>> > >> >> > >
> > > >>> > >> >> > > Can you
> please
> > > elaborate more what
> > > >>> > practical problems
> > > >>> > >> >> > > this
> introduces
> > > for users?
> > > >>> > >> >> > >
> > > >>> > >> >> > > I can see
> that
> > > this change allows a
> > > >>> > user to specify a
> > > >>> > >> >> > > runner
> specific
> > > option, which in the
> > > >>> > future may change
> > > >>> > >> >> > > because we
> > decide
> > > to scope
> > > >>> > differently. If this only
> > > >>> > >> >> > > affects
> > users of
> > > the portable Flink
> > > >>> > runner (like us),
> > > >>> > >> >> > > then no
> need to
> > > revert, because at
> > > >>> > this early stage we
> > > >>> > >> >> > > prefer
> > something
> > > that works over
> > > >>> > being blocked.
> > > >>> > >> >> > >
> > > >>> > >> >> > > It would
> > also be
> > > really great if some
> > > >>> > of the core Python
> > > >>> > >> >> > > SDK
> developers
> > > could help out with
> > > >>> > the design aspects
> > > >>> > >> >> > > and PR
> > reviews of
> > > changes that affect
> > > >>> > common Python
> > > >>> > >> >> > > code.
> > Anyone who
> > > specifically wants
> > > >>> > to be tagged on
> > > >>> > >> >> > > relevant
> > JIRAs and
> > > PRs?
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > > I would be happy to
> be
> > > tagged, and I can also
> > > >>> > help with
> > > >>> > >> >> > > including other
> > relevant
> > > folks whenever
> > > >>> > possible. In general I
> > > >>> > >> >> > > think Robert,
> Charles,
> > > myself are good
> > > >>> > candidates.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > > Thanks
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > > On Fri, Oct
> 12,
> > > 2018 at 10:20 AM
> > > >>> > Ahmet Altay
> > > >>> > >> >> > >
> > <altay@google.com <ma...@google.com>
> > > <mailto:altay@google.com <ma...@google.com>>
> > > >>> > <mailto:altay@google.com
> > <ma...@google.com>
> > > <mailto:altay@google.com <ma...@google.com>>>
> > <mailto:altay@google.com <ma...@google.com>
> > > <mailto:altay@google.com <ma...@google.com>>
> > > >>> > <mailto:altay@google.com
> > <ma...@google.com>
> > > <mailto:altay@google.com
> > <ma...@google.com>>>>> wrote:
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > > On Fri,
> Oct
> > > 12, 2018 at 10:11 AM,
> > > >>> > Charles Chen
> > > >>> > >> >> > >
> > > <ccy@google.com <ma...@google.com>
> > <mailto:ccy@google.com <ma...@google.com>>
> > > >>> > <mailto:ccy@google.com
> > <ma...@google.com>
> > > <mailto:ccy@google.com <ma...@google.com>>>
> > <mailto:ccy@google.com <ma...@google.com>
> > > <mailto:ccy@google.com <ma...@google.com>>
> > > >>> > <mailto:ccy@google.com
> > <ma...@google.com>
> > > <mailto:ccy@google.com <ma...@google.com>>>>>
> wrote:
> > > >>> > >> >> > >
> > > >>> > >> >> > > For
> > > context, I made comments on
> > > >>> > >> >> > >
> > > https://github.com/apache/beam/pull/6600 noting
> > > >>> > >> >> > >
> > that the
> > > changes being made
> > > >>> > were not good for
> > > >>> > >> >> > > Beam
> > > >>> > backwards-compatibility. The change
> as is
> > > >>> > >> >> > >
> allows
> > > users to use pipeline
> > > >>> > options without
> > > >>> > >> >> > >
> > explicitly
> > > defining them,
> > > >>> > which is not the type
> > > >>> > >> >> > > of
> > usage
> > > we would like to
> > > >>> > encourage since we
> > > >>> > >> >> > >
> > prefer to
> > > be explicit
> > > >>> > whenever possible. If
> > > >>> > >> >> > >
> users
> > > write pipelines with
> > > >>> > this sort of pattern,
> > > >>> > >> >> > >
> > they will
> > > potentially
> > > >>> > encounter pain when
> > > >>> > >> >> > >
> > upgrading
> > > to a later version
> > > >>> > since this is an
> > > >>> > >> >> > >
> > > implementation detail and not
> > > >>> > an officially
> > > >>> > >> >> > >
> > supported
> > > pattern. I agree
> > > >>> > with the comments
> > > >>> > >> >> > >
> > above that
> > > this is ultimately
> > > >>> > a scoping issue.
> > > >>> > >> >> > > I
> would
> > > not have a problem
> > > >>> > with these changes if
> > > >>> > >> >> > >
> > they were
> > > explicitly scoped
> > > >>> > under either a
> > > >>> > >> >> > >
> > runner or
> > > unparsed options
> > > >>> > namespace.
> > > >>> > >> >> > >
> > > >>> > >> >> > > As a
> > > second note, since the
> > > >>> > 2.8.0 release is
> > > >>> > >> >> > >
> > being cut
> > > right now, because
> > > >>> > of these
> > > >>> > >> >> > >
> > > backwards-compatibility
> > > >>> > concerns, I would
> > > >>> > >> >> > >
> suggest
> > > reverting these
> > > >>> > changes, at least until
> > > >>> > >> >> > >
> > 2.8.0 is
> > > cut, so we can have
> > > >>> > a discussion here
> > > >>> > >> >> > >
> before
> > > committing to and
> > > >>> > releasing any API-level
> > > >>> > >> >> > >
> > changes.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > > +1 I
> would
> > > like to revert the
> > > >>> > changes in order not
> > > >>> > >> >> > > rush
> > this into
> > > the release. Once
> > > >>> > this discussion
> > > >>> > >> >> > > results
> > in an
> > > agreement changes
> > > >>> > can be brought back.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > > On
> Fri,
> > > Oct 12, 2018 at 9:26
> > > >>> > AM Henning Rohde
> > > >>> > >> >> > >
> > > <herohde@google.com <ma...@google.com>
> > <mailto:herohde@google.com <ma...@google.com>>
> > > >>> > <mailto:herohde@google.com
> > <ma...@google.com>
> > > <mailto:herohde@google.com
> > <ma...@google.com>>> <mailto:herohde@google.com
> > <ma...@google.com>
> > > <mailto:herohde@google.com <mailto:herohde@google.com
> >>
> > > >>> > <mailto:herohde@google.com
> > <ma...@google.com>
> > > <mailto:herohde@google.com
> > <ma...@google.com>>>>>
> > > >>> > >> >> > >
> wrote:
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > Agree
> > > that pipeline
> > > >>> > options lack some
> > > >>> > >> >> > >
> > > mechanism for scoping. It
> > > >>> > is also not always
> > > >>> > >> >> > >
> > > possible distinguish
> > > >>> > options meant to be
> > > >>> > >> >> > >
> > > consumed at pipeline
> > > >>> > construction time, by
> > > >>> > >> >> > >
> the
> > > runner, by the SDK
> > > >>> > harness, by the user
> > > >>> > >> >> > >
> > code
> > > or any combination
> > > >>> > -- and this causes
> > > >>> > >> >> > >
> > > confusion every now and then.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> For
> > > Dataflow, we have
> > > >>> > been using
> > > >>> > >> >> > >
> > > "experiments" for
> > > >>> > arbitrary runner-specific
> > > >>> > >> >> > >
> > > options. It's simply a
> > > >>> > string list pipeline
> > > >>> > >> >> > >
> > option
> > > that all SDKs
> > > >>> > support and, for Go at
> > > >>> > >> >> > >
> > least,
> > > is sent to
> > > >>> > portable runners. Flink
> > > >>> > >> >> > >
> > can do
> > > the same in the
> > > >>> > short term to move
> > > >>> > >> >> > >
> > forward.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > Henning
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> On
> > > Fri, Oct 12, 2018 at
> > > >>> > 8:50 AM Thomas Weise
> > > >>> > >> >> > >
> > > <thw@apache.org <ma...@apache.org>
> > <mailto:thw@apache.org <ma...@apache.org>>
> > > >>> > <mailto:thw@apache.org
> > <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>>
> > <mailto:thw@apache.org <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>
> > > >>> > <mailto:thw@apache.org
> > <ma...@apache.org>
> > > <mailto:thw@apache.org <ma...@apache.org>>>>>
> wrote:
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > [moving to the list]
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > The requirement
> > > >>> > driving this part of the
> > > >>> > >> >> > >
> > > change was to allow a
> > > >>> > user to specify
> > > >>> > >> >> > >
> > > pipeline options that
> > > >>> > a runner supports
> > > >>> > >> >> > >
> > > without having to
> > > >>> > declare those in each
> > > >>> > >> >> > >
> > > language SDK.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > In
> > > the specific
> > > >>> > scenario, we have
> > > >>> > >> >> > >
> > > options that the
> > > >>> > Flink runner supports
> > > >>> > >> >> > >
> > > (and can validate),
> > > >>> > that are not
> > > >>> > >> >> > >
> > > enumerated in the
> > > >>> > Python SDK.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > I
> > > think we have a
> > > >>> > bigger problem scoping
> > > >>> > >> >> > >
> > > pipeline options. For
> > > >>> > example, the
> > > >>> > >> >> > >
> > > runner options are
> > > >>> > dumped into the SDK
> > > >>> > >> >> > >
> > > worker. There is also
> > > >>> > a possibility of
> > > >>> > >> >> > >
> > > name collisions. So I
> > > >>> > think this would
> > > >>> > >> >> > >
> > > benefit from broader
> > > >>> > feedback.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > Thanks,
> > > >>> > >> >> > >
> > Thomas
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > ---------- Forwarded
> > > >>> > message ---------
> > > >>> > >> >> > >
> > > From: *Charles Chen*
> > > >>> > >> >> > >
> > > >>> > <notifications@github.com
> > <ma...@github.com>
> > > <mailto:notifications@github.com
> > <ma...@github.com>>
> > > <mailto:notifications@github.com
> > <ma...@github.com>
> > > <mailto:notifications@github.com
> > <ma...@github.com>>>
> > > >>> > >> >> > >
> > > >>> > <mailto:notifications@github.com
> > <ma...@github.com>
> > > <mailto:notifications@github.com
> > <ma...@github.com>>
> > > >>> > <mailto:notifications@github.com
> > <ma...@github.com>
> > > <mailto:notifications@github.com
> > <ma...@github.com>>>>>
> > > >>> > >> >> > >
> > > Date: Fri, Oct 12,
> > > >>> > 2018 at 8:36 AM
> > > >>> > >> >> > >
> > > Subject: Re:
> > > >>> > [apache/beam] [BEAM-5442]
> > > >>> > >> >> > >
> > > Store duplicate
> > > >>> > unknown options in a
> > > >>> > >> >> > >
> > > list argument (#6600)
> > > >>> > >> >> > >
> > > To: apache/beam
> > > >>> > <beam@noreply.github.com
> > <ma...@noreply.github.com>
> > > <mailto:beam@noreply.github.com
> > <ma...@noreply.github.com>>
> > > <mailto:beam@noreply.github.com
> > <ma...@noreply.github.com>
> > > <mailto:beam@noreply.github.com
> > <ma...@noreply.github.com>>>
> > > >>> > >> >> > >
> > > >>> > <mailto:beam@noreply.github.com
> > <ma...@noreply.github.com>
> > > <mailto:beam@noreply.github.com
> > <ma...@noreply.github.com>>
> > > <mailto:beam@noreply.github.com
> > <ma...@noreply.github.com>
> > > <mailto:beam@noreply.github.com
> > <ma...@noreply.github.com>>>>>
> > > >>> > >> >> > >
> > > Cc: Thomas Weise
> > > >>> > <thomas.weise@gmail.com
> > <ma...@gmail.com>
> > > <mailto:thomas.weise@gmail.com
> > <ma...@gmail.com>>
> > > <mailto:thomas.weise@gmail.com
> > <ma...@gmail.com> <mailto:thomas.weise@gmail.com
> > <ma...@gmail.com>>>
> > > >>> > >> >> > >
> > > >>> > <mailto:thomas.weise@gmail.com
> > <ma...@gmail.com>
> > > <mailto:thomas.weise@gmail.com
> > <ma...@gmail.com>>
> > > <mailto:thomas.weise@gmail.com
> > <ma...@gmail.com>
> > > <mailto:thomas.weise@gmail.com
> > <ma...@gmail.com>>>>>,
> > > >>> > >> >> > >
> > > Mention
> > > >>> > <mention@noreply.github.com
> > <ma...@noreply.github.com>
> > > <mailto:mention@noreply.github.com
> > <ma...@noreply.github.com>>
> > > <mailto:mention@noreply.github.com
> > <ma...@noreply.github.com>
> > > <mailto:mention@noreply.github.com
> > <ma...@noreply.github.com>>>
> > > >>> > >> >> > >
> > > >>> > <mailto:mention@noreply.github.com
> > <ma...@noreply.github.com>
> > > <mailto:mention@noreply.github.com
> > <ma...@noreply.github.com>>
> > > >>> > <mailto:mention@noreply.github.com
> > <ma...@noreply.github.com>
> > > <mailto:mention@noreply.github.com
> > <ma...@noreply.github.com>>>>>
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > CC: @tweise
> > > >>> > <https://github.com/tweise>
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > —
> > > >>> > >> >> > >
> > > You are receiving
> > > >>> > this because you were
> > > >>> > >> >> > >
> > > mentioned.
> > > >>> > >> >> > >
> > > Reply to this email
> > > >>> > directly, view it on
> > > >>> > >> >> > >
> > GitHub
> > > >>> > >> >> > >
> > > >>> >
> > >
> > <https://github.com/apache/beam/pull/6600#issuecomment-429367754>,
> > > >>> > >> >> > >
> > or
> > > mute the thread
> > > >>> > >> >> > >
> > > >>> >
> > >
> > <
> https://github.com/notifications/unsubscribe-auth/AAQGDwwt15R85eq9pySUisyxq2HYz-Vyks5ukLcLgaJpZM4XMo-T
> >.
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> > >> >> > >
> > > >>> >
> > >
> >
>
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list argument
Posted by Lukasz Cwik <lc...@google.com>.
The purpose of the spec would be to provide the names, type and
descriptions of the options. We don't need anything beyond the JSON types
(string, number, bool, object, list) because the only ambiguity we get is
how do we parse command line string into the JSON type (and that ambiguity
is actually only between string and non-string since all the other JSON
types are unambiguous).
Also, I believe the flow would be
1) Parse options
a) Find the URL from args specified and/or additional methods on
PipelineOptions that exposes a programmatic way to set the URL during
parsing.
b) Query URL for option specs
c) Parse the remainder of the options
2) Construct pipeline
3) Choose runner
4) Submit job to runner
Note that I believe that we will want to have multiple URLs to support
cross language pipelines since we will want to be able to ask other SDK
languages/versions for their "list" of supported PipelineOptions.
On Thu, Nov 8, 2018 at 11:51 AM Robert Bradshaw <ro...@google.com> wrote:
> There's two questions here:
>
> (A) What do we do in the short term?
>
> I think adding every runner option to every SDK is not sustainable
> (n*m work, assuming every SDK knows about every runner), and having a
> patchwork of options that were added as one-offs to SDKs is not
> desirable either. Furthermore, it seems difficult to parse unknown
> options as if they were valid options, so my preference here would be
> to just use a special runner_option flag. (One could also pass a set
> of unparsed/unvalidated runner options to the runner, even if they're
> not distinguished for the user, and runners (or any intermediates)
> could run a "promote" operation that promotes any of these unknowns
> that they recognize to real options before further processing. The
> parsing would be done as repeated-string, and not be intermingled with
> the actually validated options. This is essential a variant of option 1.)
>
> (B) What do do in the long term? While the JobServer approach sounds
> nice, I think it introduces a lot of complexity (we have too much of
> that already) and still doesn't completely solve the problem. In
> particular, it changes the flow from
>
> 1. Parse options
> 2. Construct pipeline
> 3. Choose runner
> 4. Submit job to runner
>
> to
>
> 1. Parse options
> 2. Construct pipeline
> 3. Choose runner
> 4a. Query runner for option specs
> 4b. Re-parse options
> 4c. Submit job to runner
>
> In particular, doing 4b in the SDK rather than just let the runner
> itself do the validation as part of (4) doesn't save much and forces
> us to come up with a (probably incomplete) spec as to how to define
> options, their types, and their validations. It also means that a
> delegating runner must choose and interact with its downstream
> runner(s) synchronously, else we haven't actually solved the issue.
>
> For these reasons, I don't think we even want to go with the JobServer
> approach in the long term, which has bearing on (A).
>
> - Robert
>
>
> On Wed, Nov 7, 2018 at 8:50 PM Maximilian Michels <mx...@apache.org> wrote:
> >
> > +1
> >
> > If the preferred approach is to eventually have the JobServer serve the
> > options, then the best intermediate solution is to replicate common
> > options in the SDKs.
> >
> > If we went down the "--runner_option" path, we would end up with
> > multiple ways of specifying the same options. We would eventually have
> > to deprecate "runner options" once we have the JobServer approach. I'd
> > like to avoid that.
> >
> > For the upcoming release we can revert the changes again and add the
> > most common missing options to the SDKs. Then hopefully we should have
> > fetching implemented for the release after.
> >
> > Do you think that is feasible?
> >
> > Thanks,
> > Max
> >
> > On 30.10.18 23:00, Lukasz Cwik wrote:
> > > I still like #3 the most, just can't devote the time to get it done.
> > >
> > > Instead of going with a fully implemented #3, we could hardcode the a
> > > subset of options and types within each SDK until the job server is
> > > ready to provide this information and then migrate to the "full" list.
> > > This would be an easy path for SDKs to take on. They could "know" of a
> > > few well known options, and if they want to support all options, they
> > > implement the integration with the job server.
> > >
> > > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels <mxm@apache.org
> > > <ma...@apache.org>> wrote:
> > >
> > > > I would prefer we don't introduce a (quirky) way of passing
> > > unknown options that forces users to type JSON into the command
> line
> > > (or similar acrobatics)
> > > Same here, the JSON approach seems technically nice but too bulky
> > > for users.
> > >
> > > > To someone wanting to run a pipeline, all options are equally
> > > important, whether they are application specific, SDK specific or
> > > runner specific.
> > >
> > > I'm also reluctant to force users to use `--runner_option=`
> because the
> > > division into "Runner" options and other options seems rather
> arbitrary
> > > to users. Most built-in options are also Runner-related.
> > >
> > > > It should be possible to *optionally* qualify/scope (to cover
> > > cases where there is ambiguity), but otherwise I prefer the format
> > > we currently have.
> > >
> > > Yes, namespacing is a problem. What happens if the user defines a
> > > custom
> > > PipelineOption which clashes with one of the builtin ones? If both
> are
> > >
> > > set, which one is actually applied?
> > >
> > >
> > > Note that PipelineOptions so far has been treating name equality to
> mean
> > > option equality and the Java implementation has a bunch of strict
> checks
> > > to make sure that default values aren't used for duplicate definitions,
> > > they have the same type, etc...
> > > With 1), you fail the job if the runner can't understand your option
> > > because its not represented the same way. User then needs to fix-up
> > > their declaration of the option name.
> > > With 2), there are no name conflicts, the SDK will need to validate
> that
> > > the option isn't set in both formats and error out if it is before
> > > pipeline submission time.
> > > With 3), you can prefetch all the options and error out to the user
> > > during argument parsing time.
> > >
> > >
> > >
> > > Here is a summary of the possible paths going forward:
> > >
> > >
> > > 1) Validate PipelineOptions at Runner side
> > > ==========================================
> > >
> > > The main issue raised here was that we want to move away from
> parsing
> > > arguments which look like options without validating them. An easy
> fix
> > > would be to actually validate them on the Runner side. This could
> be
> > > done by changing the deserialization code of PipelineOptions which
> so
> > > far ignores unknown JSON options.
> > >
> > > See: PipelineOptionsTranslation.fromProto(Struct protoOptions)
> > >
> > > Actually, this wouldn't work for user-defined PipelineOptions
> because
> > > they might not be known to the Runner (if they are defined in
> Python).
> > >
> > >
> > > 2) Introduce a Runner-Option Flag
> > > =================================
> > >
> > > In this approach we would try to add as many pipeline options for a
> > > Runner to the SDK, but allow additional Runner options to be passed
> > > using the `--runner-option=key=val` flag. The Runner, like in 1),
> would
> > > have to ensure validation. I think this has been the most favored
> > > way so
> > > far. Going forward, that means that `--parallelism=4` and
> > > `--runner-option=parallelism=4` will have the same effect for the
> Flink
> > > Runner.
> > >
> > >
> > > 3) Implement Fetching of Options from JobServer
> > > ===============================================
> > >
> > > The options are retrieved from the JobServer before submitting the
> > > pipeline. I think this would be ideal but, as mentioned before, it
> > > increases the complexity for implementing new SDKs and might
> overall
> > > just not be worth the effort.
> > >
> > >
> > > What do you think? I'd implement 2) for the next release, unless
> there
> > > are advocates for a different approach.
> > >
> > > Cheers,
> > > Max
>
Re: [BEAM-5442] Store duplicate unknown (runner) options in a list argument
Posted by Robert Bradshaw <ro...@google.com>.
There's two questions here:
(A) What do we do in the short term?
I think adding every runner option to every SDK is not sustainable
(n*m work, assuming every SDK knows about every runner), and having a
patchwork of options that were added as one-offs to SDKs is not
desirable either. Furthermore, it seems difficult to parse unknown
options as if they were valid options, so my preference here would be
to just use a special runner_option flag. (One could also pass a set
of unparsed/unvalidated runner options to the runner, even if they're
not distinguished for the user, and runners (or any intermediates)
could run a "promote" operation that promotes any of these unknowns
that they recognize to real options before further processing. The
parsing would be done as repeated-string, and not be intermingled with
the actually validated options. This is essential a variant of option 1.)
(B) What do do in the long term? While the JobServer approach sounds
nice, I think it introduces a lot of complexity (we have too much of
that already) and still doesn't completely solve the problem. In
particular, it changes the flow from
1. Parse options
2. Construct pipeline
3. Choose runner
4. Submit job to runner
to
1. Parse options
2. Construct pipeline
3. Choose runner
4a. Query runner for option specs
4b. Re-parse options
4c. Submit job to runner
In particular, doing 4b in the SDK rather than just let the runner
itself do the validation as part of (4) doesn't save much and forces
us to come up with a (probably incomplete) spec as to how to define
options, their types, and their validations. It also means that a
delegating runner must choose and interact with its downstream
runner(s) synchronously, else we haven't actually solved the issue.
For these reasons, I don't think we even want to go with the JobServer
approach in the long term, which has bearing on (A).
- Robert
On Wed, Nov 7, 2018 at 8:50 PM Maximilian Michels <mx...@apache.org> wrote:
>
> +1
>
> If the preferred approach is to eventually have the JobServer serve the
> options, then the best intermediate solution is to replicate common
> options in the SDKs.
>
> If we went down the "--runner_option" path, we would end up with
> multiple ways of specifying the same options. We would eventually have
> to deprecate "runner options" once we have the JobServer approach. I'd
> like to avoid that.
>
> For the upcoming release we can revert the changes again and add the
> most common missing options to the SDKs. Then hopefully we should have
> fetching implemented for the release after.
>
> Do you think that is feasible?
>
> Thanks,
> Max
>
> On 30.10.18 23:00, Lukasz Cwik wrote:
> > I still like #3 the most, just can't devote the time to get it done.
> >
> > Instead of going with a fully implemented #3, we could hardcode the a
> > subset of options and types within each SDK until the job server is
> > ready to provide this information and then migrate to the "full" list.
> > This would be an easy path for SDKs to take on. They could "know" of a
> > few well known options, and if they want to support all options, they
> > implement the integration with the job server.
> >
> > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels <mxm@apache.org
> > <ma...@apache.org>> wrote:
> >
> > > I would prefer we don't introduce a (quirky) way of passing
> > unknown options that forces users to type JSON into the command line
> > (or similar acrobatics)
> > Same here, the JSON approach seems technically nice but too bulky
> > for users.
> >
> > > To someone wanting to run a pipeline, all options are equally
> > important, whether they are application specific, SDK specific or
> > runner specific.
> >
> > I'm also reluctant to force users to use `--runner_option=` because the
> > division into "Runner" options and other options seems rather arbitrary
> > to users. Most built-in options are also Runner-related.
> >
> > > It should be possible to *optionally* qualify/scope (to cover
> > cases where there is ambiguity), but otherwise I prefer the format
> > we currently have.
> >
> > Yes, namespacing is a problem. What happens if the user defines a
> > custom
> > PipelineOption which clashes with one of the builtin ones? If both are
> >
> > set, which one is actually applied?
> >
> >
> > Note that PipelineOptions so far has been treating name equality to mean
> > option equality and the Java implementation has a bunch of strict checks
> > to make sure that default values aren't used for duplicate definitions,
> > they have the same type, etc...
> > With 1), you fail the job if the runner can't understand your option
> > because its not represented the same way. User then needs to fix-up
> > their declaration of the option name.
> > With 2), there are no name conflicts, the SDK will need to validate that
> > the option isn't set in both formats and error out if it is before
> > pipeline submission time.
> > With 3), you can prefetch all the options and error out to the user
> > during argument parsing time.
> >
> >
> >
> > Here is a summary of the possible paths going forward:
> >
> >
> > 1) Validate PipelineOptions at Runner side
> > ==========================================
> >
> > The main issue raised here was that we want to move away from parsing
> > arguments which look like options without validating them. An easy fix
> > would be to actually validate them on the Runner side. This could be
> > done by changing the deserialization code of PipelineOptions which so
> > far ignores unknown JSON options.
> >
> > See: PipelineOptionsTranslation.fromProto(Struct protoOptions)
> >
> > Actually, this wouldn't work for user-defined PipelineOptions because
> > they might not be known to the Runner (if they are defined in Python).
> >
> >
> > 2) Introduce a Runner-Option Flag
> > =================================
> >
> > In this approach we would try to add as many pipeline options for a
> > Runner to the SDK, but allow additional Runner options to be passed
> > using the `--runner-option=key=val` flag. The Runner, like in 1), would
> > have to ensure validation. I think this has been the most favored
> > way so
> > far. Going forward, that means that `--parallelism=4` and
> > `--runner-option=parallelism=4` will have the same effect for the Flink
> > Runner.
> >
> >
> > 3) Implement Fetching of Options from JobServer
> > ===============================================
> >
> > The options are retrieved from the JobServer before submitting the
> > pipeline. I think this would be ideal but, as mentioned before, it
> > increases the complexity for implementing new SDKs and might overall
> > just not be worth the effort.
> >
> >
> > What do you think? I'd implement 2) for the next release, unless there
> > are advocates for a different approach.
> >
> > Cheers,
> > Max