[
https://issues.apache.org/jira/browse/HIVE-23725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17184154#comment-17184154
]
Peter Varga commented on HIVE-23725:
------------------------------------
[~kgyrtkirk] I will have to work on an upgraded version of this, because
releasing the locks when reexecuting causing problems when the degree of
concurrent update queries for the same partition is high. I will try to address
your comments there and invite you to review.
Few questions, so I can go for the right direction:
* I have seen that the other plugin implementation use hooks, but my problem
was that this exception is thrown between the compilation and execution phase.
Can I use the failure hook there, would the failure hook be called?
* The whole reexecution count misery was introduced, to make it possible for
the new plugin to have an individual config for rexec count that can be higher
than the HIVE_QUERY_MAX_REEXECUTION_COUNT, but still keep that config for every
other plugin. If you have an idea how to solve this properly, I happy to
implement it, because I do not like the current solution either
> ValidTxnManager snapshot outdating causing partial reads in merge insert
> ------------------------------------------------------------------------
>
> Key: HIVE-23725
> URL: https://issues.apache.org/jira/browse/HIVE-23725
> Project: Hive
> Issue Type: Bug
> Reporter: Peter Varga
> Assignee: Peter Varga
> Priority: Major
> Labels: pull-request-available
> Time Spent: 5h 40m
> Remaining Estimate: 0h
>
> When the ValidTxnManager invalidates the snapshot during merge insert and
> starts to read committed transactions that were not committed when the query
> compilation happened, it can cause partial read problems if the committed
> transaction created new partition in the source or target table.
> The solution should be not only fix the snapshot but also recompile the query
> and acquire the locks again.
> You could construct an example like this:
> 1. open and compile transaction 1 that merge inserts data from a partitioned
> source table that has a few partition.
> 2. Open, run and commit transaction 2 that inserts data to an old and a new
> partition to the source table.
> 3. Open, run and commit transaction 3 that inserts data to the target table
> of the merge statement, that will retrigger a snapshot generation in
> transaction 1.
> 4. Run transaction 1, the snapshot will be regenerated, and it will read
> partial data from transaction 2 breaking the ACID properties.
> Different setup.
> Switch the transaction order:
> 1. compile transaction 1 that inserts data to an old and a new partition of
> the source table.
> 2. compile transaction 2 that insert data to the target table
> 2. compile transaction 3 that merge inserts data from the source table to the
> target table
> 3. run and commit transaction 1
> 4. run and commit transaction 2
> 5. run transaction 3, since it cointains 1 and 2 in its snaphot the
> isValidTxnListState will be triggered and we do a partial read of the
> transaction 1 for the same reasons.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
