[
https://issues.apache.org/jira/browse/HIVE-21917?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Craig Condit updated HIVE-21917:
--------------------------------
Summary: COMPLETED_TXN_COMPONENTS table is never cleaned up unless
Compactor runs (was: COMPLETED_TXN_COMPONENTS table is never cleaned up unless
Compator runs)
> COMPLETED_TXN_COMPONENTS table is never cleaned up unless Compactor runs
> ------------------------------------------------------------------------
>
> Key: HIVE-21917
> URL: https://issues.apache.org/jira/browse/HIVE-21917
> Project: Hive
> Issue Type: Bug
> Components: Transactions
> Affects Versions: 3.1.0, 3.1.1
> Reporter: Craig Condit
> Priority: Major
>
> The Initiator thread in the metastore repeatedly loops over entries in the
> COMPLETED_TXN_COMPONENTS table to determine which partitions / tables might
> need to be compacted. However, entries are never removed from this table
> except by a completed Compactor run.
> In a cluster where most tables / partitions are write-once read-many, this
> results in stale entries in this table never being cleaned up. In a small
> test cluster, we have observed approximately 45k entries in this table
> (virtually equal to the number of partitions in the cluster) while < 100 of
> these tables have delta files at all. Since most of the tables will never get
> enough writes to trigger a compaction (and in fact have only ever been
> written to once), the initiator thread keeps trying to evaluate them on every
> loop.
> On this test cluster, it takes approximately 10 minutes to loop through all
> the entries and results in severe performance degradation on metastore
> operations. With the default run timing of 5 minutes, the initiator basically
> never stops running.
> On a production cluster with 2M partitions, this would be a non-starter.
> The initiator thread should proactively remove entries from
> COMPLETED_TXN_COMPONENTS when it determines that a compaction is not needed,
> so that they are not evaluated again on the next loop.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
