[
https://issues.apache.org/jira/browse/HBASE-4414?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13105515#comment-13105515
]
Gary Helmling commented on HBASE-4414:
--------------------------------------
So it looks like this patch isn't quite sufficient yet for splits to always
happen by max filesize.
The patch does work during active writes (when we're triggering memstore
flushes). In this case, regions get correctly split by size.
But following the initial write load, on an inactive cluster, the multiple
store files from flushes get compacted back into a single store file that can
then exceed max file size, in the worst case by multiple times. So we still
need some kind of split check following compaction to ensure we're not
exceeding max size.
> Region splits by size not being triggered in at least some cases
> ----------------------------------------------------------------
>
> Key: HBASE-4414
> URL: https://issues.apache.org/jira/browse/HBASE-4414
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.92.0
> Reporter: Gary Helmling
> Priority: Blocker
> Fix For: 0.92.0
>
> Attachments: HBASE-4414.patch
>
>
> We seem to have lost the triggering of region splits by size somewhere in
> trunk.
> Running a simple test to load data only:
> 1. create 'usertable', 'f1' in hbase shell
> 2. run a YCSB load of 10M records
> I wind up with a single region containing all records, around 13GB, despite
> max region size being configured to 640MB.
> {noformat}
> ip-10-160-217-155.us-west-1.compute.internal:8120 1316045713501
> requestsPerSecond=0, numberOfOnlineRegions=1, usedHeapMB=1544,
> maxHeapMB=2962
> usertable,,1316045755455.1e11a9f71072113258942e03dabaa468.
> numberOfStores=1, numberOfStorefiles=16,
> storefileUncompressedSizeMB=13611, storefileSizeMB=13621,
> compressionRatio=1.0007, memstoreSizeMB=50, storefileIndexSizeMB=0,
> readRequestsCount=0, writeRequestsCount=1930, rootIndexSizeKB=108,
> totalStaticIndexSizeKB=10511, totalStaticBloomSizeKB=0,
> totalCompactingKVs=3356000, currentCompactedKVs=3356000,
> compactionProgressPct=1.0
> {noformat}
> As best I can tell, the changes introduced in HBASE-3797 and HBASE-1476
> dropped some cases where we were triggering region splits when we didn't
> compact.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
