[
https://issues.apache.org/jira/browse/HBASE-2375?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13204701#comment-13204701
]
Jean-Daniel Cryans commented on HBASE-2375:
-------------------------------------------
A bunch of things changed since this jira was created:
- we now split based on the store size
- regions split at 1GB
- memstores flush at 128MB
- there's been a lot of work on tuning the store file selection algorithm
My understanding of this jira is that it aims at making the "out of the box
mass import" experience better. Now that we have bulk loads and pre-splitting
this use case is becoming less and less important... although we still see
people trying to benchmark it (hi hypertable).
I see three things we could do:
- Trigger splits after flushes, I hacked a patch and it works awesomely
- Have a lower split size for newly created tables. Hypertable does this with
a soft limit that gets doubled every time the table splits until it reaches the
normal split size
- Have multi-way splits (Todd's idea), so that if you have enough data that
you know you're going to be splitting after the current split then just spawn
as many daughters as you need.
I'm planning on just fixing the first bullet point in the context of this
jira. Maybe there's another stuff from the patch in this jira that we could fit
in.
> Make decision to split based on aggregate size of all StoreFiles and revisit
> related config params
> --------------------------------------------------------------------------------------------------
>
> Key: HBASE-2375
> URL: https://issues.apache.org/jira/browse/HBASE-2375
> Project: HBase
> Issue Type: Improvement
> Components: regionserver
> Affects Versions: 0.20.3
> Reporter: Jonathan Gray
> Assignee: Jonathan Gray
> Priority: Critical
> Labels: moved_from_0_20_5
> Attachments: HBASE-2375-v8.patch
>
>
> Currently we will make the decision to split a region when a single StoreFile
> in a single family exceeds the maximum region size. This issue is about
> changing the decision to split to be based on the aggregate size of all
> StoreFiles in a single family (but still not aggregating across families).
> This would move a check to split after flushes rather than after compactions.
> This issue should also deal with revisiting our default values for some
> related configuration parameters.
> The motivating factor for this change comes from watching the behavior of
> RegionServers during heavy write scenarios.
> Today the default behavior goes like this:
> - We fill up regions, and as long as you are not under global RS heap
> pressure, you will write out 64MB (hbase.hregion.memstore.flush.size)
> StoreFiles.
> - After we get 3 StoreFiles (hbase.hstore.compactionThreshold) we trigger a
> compaction on this region.
> - Compaction queues notwithstanding, this will create a 192MB file, not
> triggering a split based on max region size (hbase.hregion.max.filesize).
> - You'll then flush two more 64MB MemStores and hit the compactionThreshold
> and trigger a compaction.
> - You end up with 192 + 64 + 64 in a single compaction. This will create a
> single 320MB and will trigger a split.
> - While you are performing the compaction (which now writes out 64MB more
> than the split size, so is about 5X slower than the time it takes to do a
> single flush), you are still taking on additional writes into MemStore.
> - Compaction finishes, decision to split is made, region is closed. The
> region now has to flush whichever edits made it to MemStore while the
> compaction ran. This flushing, in our tests, is by far the dominating factor
> in how long data is unavailable during a split. We measured about 1 second
> to do the region closing, master assignment, reopening. Flushing could take
> 5-6 seconds, during which time the region is unavailable.
> - The daughter regions re-open on the same RS. Immediately when the
> StoreFiles are opened, a compaction is triggered across all of their
> StoreFiles because they contain references. Since we cannot currently split
> a split, we need to not hang on to these references for long.
> This described behavior is really bad because of how often we have to rewrite
> data onto HDFS. Imports are usually just IO bound as the RS waits to flush
> and compact. In the above example, the first cell to be inserted into this
> region ends up being written to HDFS 4 times (initial flush, first compaction
> w/ no split decision, second compaction w/ split decision, third compaction
> on daughter region). In addition, we leave a large window where we take on
> edits (during the second compaction of 320MB) and then must make the region
> unavailable as we flush it.
> If we increased the compactionThreshold to be 5 and determined splits based
> on aggregate size, the behavior becomes:
> - We fill up regions, and as long as you are not under global RS heap
> pressure, you will write out 64MB (hbase.hregion.memstore.flush.size)
> StoreFiles.
> - After each MemStore flush, we calculate the aggregate size of all
> StoreFiles. We can also check the compactionThreshold. For the first three
> flushes, both would not hit the limit. On the fourth flush, we would see
> total aggregate size = 256MB and determine to make a split.
> - Decision to split is made, region is closed. This time, the region just
> has to flush out whichever edits made it to the MemStore during the
> snapshot/flush of the previous MemStore. So this time window has shrunk by
> more than 75% as it was the time to write 64MB from memory not 320MB from
> aggregating 5 hdfs files. This will greatly reduce the time data is
> unavailable during splits.
> - The daughter regions re-open on the same RS. Immediately when the
> StoreFiles are opened, a compaction is triggered across all of their
> StoreFiles because they contain references. This would stay the same.
> In this example, we only write a given cell twice (instead of 4 times) while
> drastically reducing data unavailability during splits. On the original
> flush, and post-split to remove references. The other benefit of post-split
> compaction (which doesn't change) is that we then get good data locality as
> the resulting StoreFile will be written to the local DataNode. In another
> jira, we should deal with opening up one of the daughter regions on a
> different RS to distribute load better, but that's outside the scope of this
> one.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
