Tanuj Khurana created PHOENIX-7990:
--------------------------------------
Summary: Convergent CAS race in setHAGroupStatusIfNeeded aborts
co-active RegionServers on stale-version loss
Key: PHOENIX-7990
URL: https://issues.apache.org/jira/browse/PHOENIX-7990
Project: Phoenix
Issue Type: Sub-task
Reporter: Tanuj Khurana
Assignee: Tanuj Khurana
When two or more active RegionServers react to the same peer-degrade event,
each independently attempts to move the shared HA-status znode from
ACTIVE_IN_SYNC → ACTIVE_NOT_IN_SYNC via an optimistic compare-and-swap (CAS).
The expected version for the CAS comes from the watch-lagged local cache. The
first writer wins; the remaining writers CAS against a now-stale version,
receive StaleHAGroupStoreRecordVersionException, and abort the RegionServer —
even though the shared status has already reached the exact value they intended
to write.
Unnecessary RegionServer aborts during a normal convergent reaction to a peer
degrade. The write is idempotent at the group level, so a peer having already
set the target state is a success condition, not a conflict — but the code
treated it as fatal.
Root cause
The validate → build → CAS body ran exactly once against a single cache-derived
(record, version) snapshot. Any lost CAS propagated as an exception with no
reconciliation against the actual current ZK state.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
