You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@trafficserver.apache.org by Walt Karas <wk...@verizonmedia.com.INVALID> on 2019/02/01 17:53:50 UTC

Debug out put in Au tests?

Pushkar pointed out to me the link to get the _sanbox for Au tests run
by CI checks:  https://ci.trafficserver.apache.org/autest/

The availability of the _sandbox means we can see trace debug output
if the test configures ATS to generate it.  But enabling this output
will cause some increase in the time to run all the tests, which is
becoming an issue.  What does everyone think, to trace or not to trace
in the merged versions of the Au tests?

Re: Debug out put in Au tests?

Posted by Alan Carroll <so...@verizonmedia.com.INVALID>.
>
> Yes, Susan brought that up yesterday and we were discussing it today in
> our internal meeting. Some timing data would obviously be useful.

Re: Debug out put in Au tests?

Posted by Jason Kenny <jk...@verizonmedia.com.INVALID>.
We can modify the default config file to have this value set for all tests.

On Mon, Feb 4, 2019 at 6:43 PM Walt Karas <wk...@verizonmedia.com> wrote:

> I meant the output that is (optionally) generated by TSDebug() calls
> in trafficsever.
>
> On Mon, Feb 4, 2019 at 5:52 PM Jason Kenny <jk...@verizonmedia.com>
> wrote:
> >
> > which debug output? autest or curl?
> >
> > On Mon, Feb 4, 2019 at 4:29 PM Pushkar Pradhan <
> ppradhan@verizonmedia.com.invalid> wrote:
> >>
> >> Walt,
> >> What is your proposal, is it to run all tests with debug output?
> >> This would require modifying all existing tests. That also brings up
> >> another question, what sections should be enabled?
> >> I.e. http, dns, iocore, ....
> >>
> >> One potential downfall to turning on debugging is that writing to the
> debug
> >> logs is a "synchronous" event which IMO changes the way TS runs during
> >> autests.
> >> As a result, Heisenbugs could slip by. A very low possibility and I am
> sure
> >> most people have good load testing to catch these anyway?
> >>
> >> On Fri, Feb 1, 2019 at 11:57 AM Miles Libbey <ml...@apache.org>
> wrote:
> >>
> >> > Since we have no data what is taking time to run the suite of tests,
> >> > isn't it premature to be considering optimizations that may or may not
> >> > affect the time?
> >> >
> >> > miles
> >> >
> >> > On Fri, Feb 1, 2019 at 11:36 AM Walt Karas
> >> > <wk...@verizonmedia.com.invalid> wrote:
> >> > >
> >> > > Pushkar pointed out to me the link to get the _sanbox for Au tests
> run
> >> > > by CI checks:  https://ci.trafficserver.apache.org/autest/
> >> > >
> >> > > The availability of the _sandbox means we can see trace debug output
> >> > > if the test configures ATS to generate it.  But enabling this output
> >> > > will cause some increase in the time to run all the tests, which is
> >> > > becoming an issue.  What does everyone think, to trace or not to
> trace
> >> > > in the merged versions of the Au tests?
> >> >
>

Re: Debug out put in Au tests?

Posted by Walt Karas <wk...@verizonmedia.com.INVALID>.
I meant the output that is (optionally) generated by TSDebug() calls
in trafficsever.

On Mon, Feb 4, 2019 at 5:52 PM Jason Kenny <jk...@verizonmedia.com> wrote:
>
> which debug output? autest or curl?
>
> On Mon, Feb 4, 2019 at 4:29 PM Pushkar Pradhan <pp...@verizonmedia.com.invalid> wrote:
>>
>> Walt,
>> What is your proposal, is it to run all tests with debug output?
>> This would require modifying all existing tests. That also brings up
>> another question, what sections should be enabled?
>> I.e. http, dns, iocore, ....
>>
>> One potential downfall to turning on debugging is that writing to the debug
>> logs is a "synchronous" event which IMO changes the way TS runs during
>> autests.
>> As a result, Heisenbugs could slip by. A very low possibility and I am sure
>> most people have good load testing to catch these anyway?
>>
>> On Fri, Feb 1, 2019 at 11:57 AM Miles Libbey <ml...@apache.org> wrote:
>>
>> > Since we have no data what is taking time to run the suite of tests,
>> > isn't it premature to be considering optimizations that may or may not
>> > affect the time?
>> >
>> > miles
>> >
>> > On Fri, Feb 1, 2019 at 11:36 AM Walt Karas
>> > <wk...@verizonmedia.com.invalid> wrote:
>> > >
>> > > Pushkar pointed out to me the link to get the _sanbox for Au tests run
>> > > by CI checks:  https://ci.trafficserver.apache.org/autest/
>> > >
>> > > The availability of the _sandbox means we can see trace debug output
>> > > if the test configures ATS to generate it.  But enabling this output
>> > > will cause some increase in the time to run all the tests, which is
>> > > becoming an issue.  What does everyone think, to trace or not to trace
>> > > in the merged versions of the Au tests?
>> >

Re: Debug out put in Au tests?

Posted by Jason Kenny <jk...@verizonmedia.com.INVALID>.
which debug output? autest or curl?

On Mon, Feb 4, 2019 at 4:29 PM Pushkar Pradhan
<pp...@verizonmedia.com.invalid> wrote:

> Walt,
> What is your proposal, is it to run all tests with debug output?
> This would require modifying all existing tests. That also brings up
> another question, what sections should be enabled?
> I.e. http, dns, iocore, ....
>
> One potential downfall to turning on debugging is that writing to the debug
> logs is a "synchronous" event which IMO changes the way TS runs during
> autests.
> As a result, Heisenbugs could slip by. A very low possibility and I am sure
> most people have good load testing to catch these anyway?
>
> On Fri, Feb 1, 2019 at 11:57 AM Miles Libbey <ml...@apache.org> wrote:
>
> > Since we have no data what is taking time to run the suite of tests,
> > isn't it premature to be considering optimizations that may or may not
> > affect the time?
> >
> > miles
> >
> > On Fri, Feb 1, 2019 at 11:36 AM Walt Karas
> > <wk...@verizonmedia.com.invalid> wrote:
> > >
> > > Pushkar pointed out to me the link to get the _sanbox for Au tests run
> > > by CI checks:  https://ci.trafficserver.apache.org/autest/
> > >
> > > The availability of the _sandbox means we can see trace debug output
> > > if the test configures ATS to generate it.  But enabling this output
> > > will cause some increase in the time to run all the tests, which is
> > > becoming an issue.  What does everyone think, to trace or not to trace
> > > in the merged versions of the Au tests?
> >
>

Re: Debug out put in Au tests?

Posted by Pushkar Pradhan <pp...@verizonmedia.com.INVALID>.
Walt,
What is your proposal, is it to run all tests with debug output?
This would require modifying all existing tests. That also brings up
another question, what sections should be enabled?
I.e. http, dns, iocore, ....

One potential downfall to turning on debugging is that writing to the debug
logs is a "synchronous" event which IMO changes the way TS runs during
autests.
As a result, Heisenbugs could slip by. A very low possibility and I am sure
most people have good load testing to catch these anyway?

On Fri, Feb 1, 2019 at 11:57 AM Miles Libbey <ml...@apache.org> wrote:

> Since we have no data what is taking time to run the suite of tests,
> isn't it premature to be considering optimizations that may or may not
> affect the time?
>
> miles
>
> On Fri, Feb 1, 2019 at 11:36 AM Walt Karas
> <wk...@verizonmedia.com.invalid> wrote:
> >
> > Pushkar pointed out to me the link to get the _sanbox for Au tests run
> > by CI checks:  https://ci.trafficserver.apache.org/autest/
> >
> > The availability of the _sandbox means we can see trace debug output
> > if the test configures ATS to generate it.  But enabling this output
> > will cause some increase in the time to run all the tests, which is
> > becoming an issue.  What does everyone think, to trace or not to trace
> > in the merged versions of the Au tests?
>

Re: Debug out put in Au tests?

Posted by Jason Kenny <jk...@verizonmedia.com.INVALID>.
Having more output will slow stuff down. However, that is not the main
issue.

I had pointed out before various testability issues with ATS. We have a
number of tests that have to wait for an amount of time for the test to
work. These delays are absolute, in that we have to wait the full time
given. These delays add up. fast. one test with 3 20 sec delays means one
minute. much of that time probably is not needed, and if the test could
react to a state correctly then the test would run faster. This is the low
hanging fruit that exists today. From there other items could be done or
found to speed up the total time for the build. However I would push first
on making sure states we want to test are detectable, and race issues (
cough cough log files) the delay resolution of information get optimized.

There are a number of knobs in autest to control shutdown and startup
times, however, these all (I believe all is the case) are reactive on a
state of the process which we can test for quickly. I should have already
tweaked these in the setup.cli.ext file.


On Fri, Feb 1, 2019 at 1:58 PM Miles Libbey <ml...@apache.org> wrote:

> Since we have no data what is taking time to run the suite of tests,
> isn't it premature to be considering optimizations that may or may not
> affect the time?
>
> miles
>
> On Fri, Feb 1, 2019 at 11:36 AM Walt Karas
> <wk...@verizonmedia.com.invalid> wrote:
> >
> > Pushkar pointed out to me the link to get the _sanbox for Au tests run
> > by CI checks:  https://ci.trafficserver.apache.org/autest/
> >
> > The availability of the _sandbox means we can see trace debug output
> > if the test configures ATS to generate it.  But enabling this output
> > will cause some increase in the time to run all the tests, which is
> > becoming an issue.  What does everyone think, to trace or not to trace
> > in the merged versions of the Au tests?
>

Re: Debug out put in Au tests?

Posted by Miles Libbey <ml...@apache.org>.
Since we have no data what is taking time to run the suite of tests,
isn't it premature to be considering optimizations that may or may not
affect the time?

miles

On Fri, Feb 1, 2019 at 11:36 AM Walt Karas
<wk...@verizonmedia.com.invalid> wrote:
>
> Pushkar pointed out to me the link to get the _sanbox for Au tests run
> by CI checks:  https://ci.trafficserver.apache.org/autest/
>
> The availability of the _sandbox means we can see trace debug output
> if the test configures ATS to generate it.  But enabling this output
> will cause some increase in the time to run all the tests, which is
> becoming an issue.  What does everyone think, to trace or not to trace
> in the merged versions of the Au tests?