[
https://issues.apache.org/jira/browse/HBASE-19095?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16224873#comment-16224873
]
Duo Zhang commented on HBASE-19095:
-----------------------------------
Thanks for the explanation. Skimmed the code again, much easier to understand
now.
There is a flatten operation which is used to make a segment immutable. And
then we can do merge/compact. In basic mode, we do merging, so we use
{{MemStoreMergerSegmentsIterator}}, where we use a KeyValueHeap directly to
skip the SQM logic. In eager mode, we do compaction, so we use
{{MemStoreCompactorSegmentsIterator }}, where we use a StoreScanner and will
drop cells.
Agree that we could provide a preInMemoryCompactScannerOpen method which allow
users to change the max versions and TTL. And we could also change the
StoreScanner field in {{MemStoreCompactorSegmentsIterator}} to InternalScanner,
and add a preInMemoryCompact method which allow users to do filtering.
And seems there is no selection so we do not need
(pre/post)InMemoryCompactSelection methods, so the only unsure method is
postInMemoryCompact. For compaction and flush we have a similar method, but I
do not think it is very useful for in memory compaction? Anyway, no harm to add
it I think.
Let me provide a patch.
Thanks.
> Add CP hooks in RegionObserver for in memory compaction
> -------------------------------------------------------
>
> Key: HBASE-19095
> URL: https://issues.apache.org/jira/browse/HBASE-19095
> Project: HBase
> Issue Type: Sub-task
> Components: Coprocessors
> Reporter: Duo Zhang
> Fix For: 2.0.0-beta-1
>
>
> This is a hole in our CP hooks.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
