Tanuj Khurana created PHOENIX-7974:
--------------------------------------
Summary: Forwarder must preserve origin server identity in the
peer destination path to avoid same-timestamp file collisions wedging SYNC
re-entry
Key: PHOENIX-7974
URL: https://issues.apache.org/jira/browse/PHOENIX-7974
Project: Phoenix
Issue Type: Sub-task
Reporter: Tanuj Khurana
Assignee: Tanuj Khurana
When the fallback cluster forwards replication log files to the peer standby
cluster, ReplicationLogDiscoveryForwarder.processFile computed the peer
destination path as:
{code:java}
getWriterPath(sourceTs, forwardingServerName){code}
This keys the destination filename <ts>_<server>.plog on the forwarding
RegionServer's own name, discarding the origin writer's identity. The origin
RS's (server, timestamp) uniqueness invariant is therefore lost on the peer.
When two source log files share the same timestamp but originate from different
origin RegionServers, both collapse onto a single destination path. Because the
copy uses FileUtil.copy(..., overwrite=false), the second file is permanently
rejected with PathExistsException. processFile throws on every retry, so
processOneRandomFile never calls markCompleted, the source stays in
out_progress, and the HA group is held in SYNC_AND_FORWARD long after the data
has actually been replicated — SYNC re-entry is wedged indefinitely.
Root cause
The destination filename must reproduce the origin RS's (server, timestamp)
pair so it inherits that writer's existing uniqueness guarantee. Keying on the
forwarding RS instead breaks the invariant precisely for cross-RegionServer
same-timestamp files.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
