[
https://issues.apache.org/jira/browse/HBASE-17924?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16019135#comment-16019135
]
stack commented on HBASE-17924:
-------------------------------
I like [~jerryhe]'s question. Not sure how sort makes a difference
("...conflicts faster..." implies concurrent threads are moving through the
batch in lock-step); and if we are going to sort batches, would be good to do
it across the board so we could lean on batches being sorted (say, skip taking
lock if second batch on a row); i.e. +1 on the semantic improvement. I'm in
this section at the mo. It is looking like our reeentrant read/write lock, the
basis of our current row locking, goes to hell if there are more than a few
read-then-modify operations in the mix (i.e. a bunch of increments on a row
also taking mutations). Thanks.
> Consider sorting the row order when processing multi() ops before taking
> rowlocks
> ---------------------------------------------------------------------------------
>
> Key: HBASE-17924
> URL: https://issues.apache.org/jira/browse/HBASE-17924
> Project: HBase
> Issue Type: Improvement
> Affects Versions: 2.0.0, 1.1.8
> Reporter: Andrew Purtell
> Assignee: Allan Yang
> Fix For: 2.0.0, 1.4.0
>
> Attachments: HBASE-17924.patch, HBASE-17924.v0.patch,
> HBASE-17924.v2.patch, HBASE-17924.v3.patch, HBASE-17924.v4.patch,
> HBASE-17924.v5.patch
>
>
> When processing a batch mutation, we take row locks in whatever order the
> mutations were added to the multi op by the client.
>
> {noformat}
> RSRpcServices#multi -> RSRpcServices#mutateRows -> HRegion#mutateRow ->
> HRegion#mutateRowsWithLocks -> HRegion#processRowsWithLocks
> {noformat}
> Or
> {noformat}
> RSRpcServices#multi -> RSRpcServices#doNonAtomicRegionMutation ->
> HRegion#get
> | HRegion#append
> | HRegion#increment
> | HRegionServer#doBatchOp -> HRegion#batchMutate ->
> HRegion#doMiniBatchMutation
> {noformat}
>
> multi() is fed by client APIs that accept a RowMutations object containing
> actions for multiple rows. The container for ops inside RowMutations is an
> ArrayList, which doesn't change the ordering of objects added to it. The
> protobuf implementation of the messages for multi ops do not reorder the list
> of actions. When processing multi ops we iterate over the actions in the
> order rehydrated from protobuf.
> We should discuss sorting the order of ops by row key when processing multi()
> ops before taking row locks. Does this make lock ordering more predictable
> for server side operations? Yes, but potentially surprising for the client,
> right? Is there any legitimate reason we should take locks out of row key
> sorted order because the client has structured the request as such?
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
