Tanuj Khurana created PHOENIX-8058:
--------------------------------------
Summary: Skip re-shipping already-replicated WAL edits during
region recovery
Key: PHOENIX-8058
URL: https://issues.apache.org/jira/browse/PHOENIX-8058
Project: Phoenix
Issue Type: Sub-task
Reporter: Tanuj Khurana
Assignee: Tanuj Khurana
When a RegionServer crashes, the crash procedure splits its WAL. Each region
then replays its recovered edits on open, and IndexRegionObserver.preWALRestore
re-ships every replayed HA edit to the replication log, to cover edits that
committed locally but never shipped before the crash. Today:
- Every recovered edit is re-shipped, including the large majority that already
reached the peer before the crash. The standby has to apply them again, which
is harmless but wasted work.
- One append and one sync per edit. The region open waits on a replication sync
for every recovered edit, so recovery time grows with the number of edits
replayed, and the region stays unavailable meanwhile.
- One HA group lookup per edit. Each edit is checked against its HA group, and
a failure throws, which fails the region open. A peer problem during recovery
(for example during a failover) can therefore keep regions from opening, or
loop the open retries.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)