Nirdosh Kumar Yadav created HBASE-30433:
-------------------------------------------
Summary: Seed the master's flushedSequenceIdByRegion watermark
with the region's final flushed seqid on region CLOSE, so a subsequent WAL
split of a drained/crashed source RS recognizes already-durable edits instead
of writing orphaned recovered.edits.
Key: HBASE-30433
URL: https://issues.apache.org/jira/browse/HBASE-30433
Project: HBase
Issue Type: Bug
Components: proc-v2
Affects Versions: 2.6.7, 2.6.5, 3.0.0
Reporter: Nirdosh Kumar Yadav
Summary — HBASE-30335 seeds the master's flushedSequenceIdByRegion watermark
with openSeqNum at OPEN time so a later WAL split of a drained/crashed source
RS recognizes already-durable edits instead of writing orphaned
recovered.edits. This follow-up additionally seeds the same watermark at CLOSE
time, using the region's final flushed seqid, to tighten a narrow window the
OPEN-time seed alone doesn't cover.
Motivation (window flagged by [~Umeshkumar9414] on #8584):
1. A region is gracefully closed on the source RS (memstore flushed to the
close marker).
2. The source RS dies before the target RS finishes OPEN — so the OPEN-time
seed hasn't fired.
3. The source RS's SCP splits its WAL in that window, filtering against the
still-stale watermark, and writes a (harmless-but-present) recovered.edits file.
Seeding on CLOSE advances the watermark to the reported flushed seqid before
the source RS dies. It's additive, not a replacement: it does nothing for the
pure-crash / never-gracefully-closed path, so the OPEN-time seed must remain.
Both writers use the same monotonic merge(Math::max), so they compose safely.
Proposed change (no proto/RPC change — the RegionStateTransition message
already carries optional openSeqNum, and CLOSED is already routed through the
same master handler):
- RegionServer — CloseRegionHandler / UnassignRegionHandler pass
HRegion.getMaxFlushedSeqId() instead of NO_SEQNUM;
HRegionServer.createReportRegionStateTransitionRequest sets the wire openSeqNum
field for CLOSED too (when >= 0).
- Master — AssignmentManager CLOSED case seeds via
serverManager.reportRegionOpen(regionInfo, seqId).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
