[
https://issues.apache.org/jira/browse/HDDS-16417?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Wei-Chiu Chuang resolved HDDS-16417.
------------------------------------
Resolution: Duplicate
> InterSCMGrpcClient mis-applies ozone.scm.ha.grpc.deadline.interval
> (MILLISECONDS vs SECONDS)
> --------------------------------------------------------------------------------------------
>
> Key: HDDS-16417
> URL: https://issues.apache.org/jira/browse/HDDS-16417
> Project: Apache Ozone
> Issue Type: Bug
> Components: SCM HA
> Reporter: Wei-Chiu Chuang
> Priority: Major
> Labels: grpc, scm-ha
>
> h3. Problem
> {{ozone.scm.ha.grpc.deadline.interval}} is documented as the deadline for SCM
> DB checkpoint download over Inter-SCM gRPC (default {{30m}} in
> {{ozone-default.xml}}).
> {{InterSCMGrpcClient}} reads the config with {{getTimeDuration(...,
> TimeUnit.MILLISECONDS)}}, so {{30m}} becomes {{1_800_000}} (milliseconds).
> The gRPC stub deadline is then set with:
> {code}
> withDeadlineAfter(timeout, TimeUnit.SECONDS);
> {code}
> So the numeric value is interpreted as *seconds*, not milliseconds. With the
> default config this is ~1,800,000 seconds (~21 days), not 30 minutes. The
> documented deadline is not enforced.
> h3. Affected code
> *
> {{hadoop-hdds/server-scm/src/main/java/org/apache/hadoop/hdds/scm/ha/InterSCMGrpcClient.java}}
> * Config: {{ScmConfigKeys.OZONE_SCM_HA_GRPC_DEADLINE_INTERVAL}} /
> {{ozone-default.xml}}
> Used when a follower (or Recon) downloads the leader SCM RocksDB checkpoint
> during SCM HA install snapshot ({{SCMSnapshotProvider.getSCMDBSnapshot}}).
> h3. Expected behavior
> A {{30m}} setting should expire the client RPC after 30 minutes of wall-clock
> time (gRPC {{DEADLINE_EXCEEDED}}), allowing operators to tune large
> {{scm.db}} transfers.
> h3. Suggested fix
> Use {{TimeUnit.MILLISECONDS}} in {{withDeadlineAfter}}, or read the duration
> with {{TimeUnit.SECONDS}} and pass seconds consistently. Add a unit test that
> a small value (e.g. {{1s}}) fails a slow/hung download within expected bounds.
> h3. Impact
> Operators may believe snapshot downloads are capped at 30m while the
> effective gRPC deadline is orders of magnitude larger; troubleshooting
> snapshot loops may incorrectly rule out client deadline expiry.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
