[
https://issues.apache.org/jira/browse/PHOENIX-7115?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17794438#comment-17794438
]
ASF GitHub Bot commented on PHOENIX-7115:
-----------------------------------------
jpisaac commented on code in PR #1748:
URL: https://github.com/apache/phoenix/pull/1748#discussion_r1419676088
##########
phoenix-core/src/main/java/org/apache/phoenix/coprocessor/MetaDataEndpointImpl.java:
##########
@@ -604,6 +608,8 @@ public static PName newPName(byte[] keyBuffer, int
keyOffset, int keyLength) {
private MetricsMetadataSource metricsSource;
private long metadataCacheInvalidationTimeoutMs;
+ private Connection invalidateMetadataCacheConnection = null;
Review Comment:
@shahrs87 As discussed offline, wondering whether this connection for
invalidation can be wrapped inside CQSI. It might be useful to provide an
interface through CQSI. This will be in line with our current packaging of
hConnections inside CQSI. Another advantage will be we can then use
PhoenixConnection/CQSI to invoke this from out-of-band tools and clients too.
To begin with, the invalidation API can be exposed only for server-side
connection and be a no-op for other clients/profiles.
> Create separate handler thread pool for invalidating server metadata cache
> --------------------------------------------------------------------------
>
> Key: PHOENIX-7115
> URL: https://issues.apache.org/jira/browse/PHOENIX-7115
> Project: Phoenix
> Issue Type: Sub-task
> Reporter: Rushabh Shah
> Assignee: Rushabh Shah
> Priority: Major
>
> MutableIndexFailureIT#testIndexWriteFailure is failing. See
> [this|https://ci-hadoop.apache.org/job/Phoenix/job/Phoenix-PreCommit-GitHub-PR/job/PR-1726/9/#showFailuresLink]
> build for reference.
> Currently all the invalidateServerMetadataCache invocations are handled by
> default RPC handler threads. We have 5 default handler threads configured in
> tests.
> This test makes sure that write to index table fails and since
> disableIndexOnWriteFailure is set to true, it will disable the index. We use
> around 10 different threads on the client side to write to base table and
> index table, so we use all the 5 handler threads to serve writes.
> Since writes to index fails, within the handler threads it will try to update
> the index state to DISABLE.
> On receiving the updateIndexState rpc, MetadataEndpointImpl will try to
> invalidate the server metadata cache on all the regionservers. The
> regionserver hosting index table doesn't have any available handler threads
> to serve invalidateServerMetadataCache requests. Hence the test fails.
> The root cause is all the read/write operations and invalidate server
> metadata cache operations share the same RPC handler pool. We need to have
> separate thread pool for
> invalidate server metadata cache operations
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
