[
https://issues.apache.org/jira/browse/PHOENIX-4051?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16104289#comment-16104289
]
James Taylor commented on PHOENIX-4051:
---------------------------------------
I think this patch prevents out-of-order mutations for a few reasons:
# For DELETE and UPSERT SELECT, when the statement is compiled, Phoenix uses
the time stamp it got back from the server when it checks that it has the
latest metadata as the time stamp for the mutations that the statement
generates. For an UPSERT VALUES, the timestamp is gotten from the server at
commit time. This is done so that the timestamp is consistent across all rows
being written. However, this could make it to the RS in a different order. This
patch is essentially ignoring the timestamp from the client and setting it on
the RS instead.
# On the HBase side (and my knowledge is sketchy here), it seems like the code
you indicated will cause a region wide lock to be in place while the
preBatchMutate call is made:
{code}
lock(this.updatesLock.readLock(), numReadyToWrite);
locked = true;
{code}
By setting the time stamp here (really the first point we can set it), I don't
think the rows will change. I'm not sure what kind of row lock is gotten in
0.98, but it looks like only a read lock is gotten in 1.3. Does that mean that
the row can be modified with this lock in place?
It's seems that (1) is the biggest factor. Either way, it feels like a
workaround to me. The indexing code should generate the correct mutations when
they're out of order. I've filed PHOENIX-4052 for that.
> Prevent out-of-order updates for mutable index updates
> ------------------------------------------------------
>
> Key: PHOENIX-4051
> URL: https://issues.apache.org/jira/browse/PHOENIX-4051
> Project: Phoenix
> Issue Type: Bug
> Reporter: James Taylor
> Assignee: James Taylor
> Attachments: PHOENIX-4051_v1.patch
>
>
> Out-of-order processing of data rows during index maintenance causes mutable
> indexes to become out of sync with regard to the data table. Here's a simple
> example to illustrate the issue:
> # Assume table T(K,V) and index X(V,K).
> # Upsert T(A, 1) at t10. Index updates: Put X(1,A) at t10.
> # Upsert T(A, 3) at t30. Index updates: Delete X(1,A) at t29, Put X(3,A) at
> t30.
> # Upsert T(A,2) at t20. Index updates: Delete X(1,A) at t19, Put X(2,A) at
> t20, Delete X(2,A) at t29
> Ideally, we'd want to remove the Delete X(1,A) at t29 since this isn't
> correct in terms of timeline consistency, but we can't do that with HBase
> without support for deleting/undoing Delete markers.
> The above is not what is occurring. Instead, when T(A,2) comes in, the Put
> X(2,A) will occur at t20, but the Delete won't occur. This causes more index
> rows than data rows, essentially making it invalid.
> A quick fix is to reset the timestamp of the data table mutations to the
> current time within the preBatchMutate call, when the row is exclusively
> locked. This skirts the issue because then timestamps won't overlap.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
