[
https://issues.apache.org/jira/browse/HBASE-15871?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15324260#comment-15324260
]
Jeongdae Kim commented on HBASE-15871:
--------------------------------------
I found that the flush operation does not always create new snapshot. if flush
failed by some reson, next flush operation will use the snapshot that created
by previous failed flush opeartion. in this case, there's no guarantee that in
memstore, all cells that were already committed before the second flush
operation, are flushed to HFile, because the snapshot to be flushed, was
already created by previous failed flush operation.
So, I changed some codes to ensure that the sequence id of HFile is always same
as the sequence id of the flush operation which created the snapshot.
> Memstore flush doesn't finish because of backwardseek() in memstore scanner.
> ----------------------------------------------------------------------------
>
> Key: HBASE-15871
> URL: https://issues.apache.org/jira/browse/HBASE-15871
> Project: HBase
> Issue Type: Bug
> Components: Scanners
> Affects Versions: 1.1.2
> Reporter: Jeongdae Kim
> Fix For: 1.1.2
>
> Attachments: HBASE-15871.branch-1.1.001.patch,
> HBASE-15871.branch-1.1.002.patch, HBASE-15871.branch-1.1.003.patch,
> memstore_backwardSeek().PNG
>
>
> Sometimes in our production hbase cluster, it takes a long time to finish
> memstore flush.( for about more than 30 minutes)
> the reason is that a memstore flusher thread calls
> StoreScanner.updateReaders(), waits for acquiring a lock that store scanner
> holds in StoreScanner.next() and backwardseek() in memstore scanner runs for
> a long time.
> I think that this condition could occur in reverse scan by the following
> process.
> 1) create a reversed store scanner by requesting a reverse scan.
> 2) flush a memstore in the same HStore.
> 3) puts a lot of cells in memstore and memstore is almost full.
> 4) call the reverse scanner.next() and re-create all scanners in this store
> because all scanners was already closed by 2)'s flush() and backwardseek()
> with store's lastTop for all new scanners.
> 5) in this status, memstore is almost full by 2) and all cells in memstore
> have sequenceID greater than this scanner's readPoint because of 2)'s
> flush(). this condition causes searching all cells in memstore, and
> seekToPreviousRow() repeatly seach cells that are already searched if a row
> has one column. (described this in more detail in a attached file.)
> 6) flush a memstore again in the same HStore, and wait until 4-5) process
> finished, to update store files in the same HStore after flusing.
> I searched HBase jira. and found a similar issue. (HBASE-14497) but,
> HBASE-14497's fix can't solve this issue because that fix just changed
> recursive call to loop.(and already applied to our HBase version)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
