WALOG Design

["Slater, David M." <Da...@jhuapl.edu>]

First, thank you all for the responses on my BatchWriter question, as I was able to increase my ingestion rate by a large factor. I am now hitting disk i/o limits, which is forcing me to look at reducing file copying. My primary thoughts concerning this are reducing the hadoop replication factor as well as reducing the number of major compactions.
However, from what I understand about write ahead logs (in 1.4), even if you remove all major compactions, all data will essentially be written to disk twice: once to the WALOG in the local directory (HDFS is 1.5), then from the WALOG to an RFile on HDFS. Is this understanding correct?
I'm trying to understand what the primary reasons are for having the WALOG.
Is there any way to write directly to an RFile from the In-Memory Map (or have the WALOG in memory)?

Thanks,
David

Re: WALOG Design

[John Vines <vi...@apache.org>]

The WALOG exists to preserve any data that goes into the in memory map in
case of a system failure. You can disable the walog, but then you will have
no ability to recover data which had been written to a tserver but not
minor compacted.


On Wed, Sep 25, 2013 at 4:39 PM, Slater, David M.
<Da...@jhuapl.edu>wrote:

> First, thank you all for the responses on my BatchWriter question, as I
> was able to increase my ingestion rate by a large factor. I am now hitting
> disk i/o limits, which is forcing me to look at reducing file copying. My
> primary thoughts concerning this are reducing the hadoop replication factor
> as well as reducing the number of major compactions.****
>
> However, from what I understand about write ahead logs (in 1.4), even if
> you remove all major compactions, all data will essentially be written to
> disk twice: once to the WALOG in the local directory (HDFS is 1.5), then
> from the WALOG to an RFile on HDFS. Is this understanding correct? ****
>
> I’m trying to understand what the primary reasons are for having the
> WALOG. ****
>
> Is there any way to write directly to an RFile from the In-Memory Map (or
> have the WALOG in memory)?****
>
> ** **
>
> Thanks,
> David****
>

Re: WALOG Design

[Billie Rinaldi <bi...@gmail.com>]

On Thu, Sep 26, 2013 at 8:04 AM, Slater, David M.
<Da...@jhuapl.edu>wrote:

> Awesome, thanks!
>
> For clarification, when the in memory map is full, if it contains data for
> multiple tablets, how does it prioritize which tablets to do minor
> compactions for?


The memory manager chooses a tablet to minor compact based on the tablet's
memory usage and idle time.  It chooses the tablet with the largest (size
in bytes * 2^(minutes idle / 15)).  If memory isn't full, but there are
tablets meeting a per-table idleness threshold, it will select one of those
to minor compact.  It tries to initiate minor compactions before the memory
is completely full in an attempt to avoid blocking writes during minor
compactions.  It uses a dynamic threshold, a percentage of the maximum
memory, to decide when to initiate them.  It adjusts the threshold based on
how full the memory gets before a minor compaction is completed
successfully.


> Is there a separate in memory map for each tablet on a tablet server? When
> the in memory map is full, will it do minor compactions until all of the
> data currently in it is empty, or will it trigger a smaller number of minor
> compactions until it is at a more reasonable size?
>
> It is unclear to me if you can change the minimum size of rfiles written
> by minor compactions.
>
> Also, how does the in memory map handle mutations for tablets that are
> currently doing a minor compaction?
>
> Best,
> David
>
> -----Original Message-----
> From: Christopher [mailto:ctubbsii@apache.org]
> Sent: Wednesday, September 25, 2013 4:49 PM
> To: Accumulo User List
> Subject: Re: WALOG Design
>
> Mutations are written to WALOGS when they are inserted into a TServer's
> in-memory map. The TServer's in-memory map gets flushed to disk
> periodically, but there's a risk that the TServer will die after the data
> has been ingested, but before it is flushed to disk. The WALOGS, when
> enabled, protect against this data loss, by first writing out incoming data
> to a WALOG. The WALOG is more efficient than creating RFiles, because it
> does not contain sorted data or indexes.
> It's just a playback file, so that in case of a failure, Mutations that
> the client believed had been ingested, aren't lost.
>
> Putting the WALOG in memory defeats the purpose of the WALOG, but it can
> be disabled (per-table), if you care more about performance than protection
> against data loss. Don't disable it for the !METADATA table, though...
>
> You can generate RFiles directly (perhaps using a M/R job), and bypass the
> WALOG, and bulk import them into Accumulo.
>
> --
> Christopher L Tubbs II
> http://gravatar.com/ctubbsii
>
>
> On Wed, Sep 25, 2013 at 4:39 PM, Slater, David M.
> <Da...@jhuapl.edu> wrote:
> > First, thank you all for the responses on my BatchWriter question, as
> > I was able to increase my ingestion rate by a large factor. I am now
> > hitting disk i/o limits, which is forcing me to look at reducing file
> > copying. My primary thoughts concerning this are reducing the hadoop
> > replication factor as well as reducing the number of major compactions.
> >
> > However, from what I understand about write ahead logs (in 1.4), even
> > if you remove all major compactions, all data will essentially be
> > written to disk
> > twice: once to the WALOG in the local directory (HDFS is 1.5), then
> > from the WALOG to an RFile on HDFS. Is this understanding correct?
> >
> > I’m trying to understand what the primary reasons are for having the
> WALOG.
> >
> > Is there any way to write directly to an RFile from the In-Memory Map
> > (or have the WALOG in memory)?
> >
> >
> >
> > Thanks,
> > David
>

Re: WALOG Design

[John Vines <vi...@apache.org>]

The in memory map gets divided to be used by all existing tablets on the
tserver. It has a manager which I think is pluggable (at one point it was)
which tries to be intelligent in allocating space to tablets which need/are
using space vs. idle tablets. But when a tablet's portion of the in memory
map is full, their entire portion of the map gets written at once. So you
can sort of manage minor compact files size by manipulating the size of the
in memory map, the number of tablets on a tserver, and this memory manager.

If the tserver has space, I think it gives the tablet a new hunk for it to
start to use to keep things streamlined. However, if things are really
backed up and memory is unable to be provided, the mutations will be
rejected and this error will propagate back to clients.


On Thu, Sep 26, 2013 at 11:04 AM, Slater, David M.
<Da...@jhuapl.edu>wrote:

> Awesome, thanks!
>
> For clarification, when the in memory map is full, if it contains data for
> multiple tablets, how does it prioritize which tablets to do minor
> compactions for? Is there a separate in memory map for each tablet on a
> tablet server? When the in memory map is full, will it do minor compactions
> until all of the data currently in it is empty, or will it trigger a
> smaller number of minor compactions until it is at a more reasonable size?
>
> It is unclear to me if you can change the minimum size of rfiles written
> by minor compactions.
>
> Also, how does the in memory map handle mutations for tablets that are
> currently doing a minor compaction?
>
> Best,
> David
>
> -----Original Message-----
> From: Christopher [mailto:ctubbsii@apache.org]
> Sent: Wednesday, September 25, 2013 4:49 PM
> To: Accumulo User List
> Subject: Re: WALOG Design
>
> Mutations are written to WALOGS when they are inserted into a TServer's
> in-memory map. The TServer's in-memory map gets flushed to disk
> periodically, but there's a risk that the TServer will die after the data
> has been ingested, but before it is flushed to disk. The WALOGS, when
> enabled, protect against this data loss, by first writing out incoming data
> to a WALOG. The WALOG is more efficient than creating RFiles, because it
> does not contain sorted data or indexes.
> It's just a playback file, so that in case of a failure, Mutations that
> the client believed had been ingested, aren't lost.
>
> Putting the WALOG in memory defeats the purpose of the WALOG, but it can
> be disabled (per-table), if you care more about performance than protection
> against data loss. Don't disable it for the !METADATA table, though...
>
> You can generate RFiles directly (perhaps using a M/R job), and bypass the
> WALOG, and bulk import them into Accumulo.
>
> --
> Christopher L Tubbs II
> http://gravatar.com/ctubbsii
>
>
> On Wed, Sep 25, 2013 at 4:39 PM, Slater, David M.
> <Da...@jhuapl.edu> wrote:
> > First, thank you all for the responses on my BatchWriter question, as
> > I was able to increase my ingestion rate by a large factor. I am now
> > hitting disk i/o limits, which is forcing me to look at reducing file
> > copying. My primary thoughts concerning this are reducing the hadoop
> > replication factor as well as reducing the number of major compactions.
> >
> > However, from what I understand about write ahead logs (in 1.4), even
> > if you remove all major compactions, all data will essentially be
> > written to disk
> > twice: once to the WALOG in the local directory (HDFS is 1.5), then
> > from the WALOG to an RFile on HDFS. Is this understanding correct?
> >
> > I’m trying to understand what the primary reasons are for having the
> WALOG.
> >
> > Is there any way to write directly to an RFile from the In-Memory Map
> > (or have the WALOG in memory)?
> >
> >
> >
> > Thanks,
> > David
>

RE: WALOG Design

["Slater, David M." <Da...@jhuapl.edu>]

Awesome, thanks!

For clarification, when the in memory map is full, if it contains data for multiple tablets, how does it prioritize which tablets to do minor compactions for? Is there a separate in memory map for each tablet on a tablet server? When the in memory map is full, will it do minor compactions until all of the data currently in it is empty, or will it trigger a smaller number of minor compactions until it is at a more reasonable size? 

It is unclear to me if you can change the minimum size of rfiles written by minor compactions.

Also, how does the in memory map handle mutations for tablets that are currently doing a minor compaction?

Best,
David

-----Original Message-----
From: Christopher [mailto:ctubbsii@apache.org] 
Sent: Wednesday, September 25, 2013 4:49 PM
To: Accumulo User List
Subject: Re: WALOG Design

Mutations are written to WALOGS when they are inserted into a TServer's in-memory map. The TServer's in-memory map gets flushed to disk periodically, but there's a risk that the TServer will die after the data has been ingested, but before it is flushed to disk. The WALOGS, when enabled, protect against this data loss, by first writing out incoming data to a WALOG. The WALOG is more efficient than creating RFiles, because it does not contain sorted data or indexes.
It's just a playback file, so that in case of a failure, Mutations that the client believed had been ingested, aren't lost.

Putting the WALOG in memory defeats the purpose of the WALOG, but it can be disabled (per-table), if you care more about performance than protection against data loss. Don't disable it for the !METADATA table, though...

You can generate RFiles directly (perhaps using a M/R job), and bypass the WALOG, and bulk import them into Accumulo.

--
Christopher L Tubbs II
http://gravatar.com/ctubbsii


On Wed, Sep 25, 2013 at 4:39 PM, Slater, David M.
<Da...@jhuapl.edu> wrote:
> First, thank you all for the responses on my BatchWriter question, as 
> I was able to increase my ingestion rate by a large factor. I am now 
> hitting disk i/o limits, which is forcing me to look at reducing file 
> copying. My primary thoughts concerning this are reducing the hadoop 
> replication factor as well as reducing the number of major compactions.
>
> However, from what I understand about write ahead logs (in 1.4), even 
> if you remove all major compactions, all data will essentially be 
> written to disk
> twice: once to the WALOG in the local directory (HDFS is 1.5), then 
> from the WALOG to an RFile on HDFS. Is this understanding correct?
>
> I’m trying to understand what the primary reasons are for having the WALOG.
>
> Is there any way to write directly to an RFile from the In-Memory Map 
> (or have the WALOG in memory)?
>
>
>
> Thanks,
> David

Re: WALOG Design

[Christopher <ct...@apache.org>]

Mutations are written to WALOGS when they are inserted into a
TServer's in-memory map. The TServer's in-memory map gets flushed to
disk periodically, but there's a risk that the TServer will die after
the data has been ingested, but before it is flushed to disk. The
WALOGS, when enabled, protect against this data loss, by first writing
out incoming data to a WALOG. The WALOG is more efficient than
creating RFiles, because it does not contain sorted data or indexes.
It's just a playback file, so that in case of a failure, Mutations
that the client believed had been ingested, aren't lost.

Putting the WALOG in memory defeats the purpose of the WALOG, but it
can be disabled (per-table), if you care more about performance than
protection against data loss. Don't disable it for the !METADATA
table, though...

You can generate RFiles directly (perhaps using a M/R job), and bypass
the WALOG, and bulk import them into Accumulo.

--
Christopher L Tubbs II
http://gravatar.com/ctubbsii


On Wed, Sep 25, 2013 at 4:39 PM, Slater, David M.
<Da...@jhuapl.edu> wrote:
> First, thank you all for the responses on my BatchWriter question, as I was
> able to increase my ingestion rate by a large factor. I am now hitting disk
> i/o limits, which is forcing me to look at reducing file copying. My primary
> thoughts concerning this are reducing the hadoop replication factor as well
> as reducing the number of major compactions.
>
> However, from what I understand about write ahead logs (in 1.4), even if you
> remove all major compactions, all data will essentially be written to disk
> twice: once to the WALOG in the local directory (HDFS is 1.5), then from the
> WALOG to an RFile on HDFS. Is this understanding correct?
>
> I’m trying to understand what the primary reasons are for having the WALOG.
>
> Is there any way to write directly to an RFile from the In-Memory Map (or
> have the WALOG in memory)?
>
>
>
> Thanks,
> David