Tanuj Khurana created PHOENIX-8057:
--------------------------------------
Summary: Initialize ReplicationLogGroup outside the instance cache
and close it when init fails
Key: PHOENIX-8057
URL: https://issues.apache.org/jira/browse/PHOENIX-8057
Project: Phoenix
Issue Type: Sub-task
Reporter: Tanuj Khurana
Assignee: Tanuj Khurana
ReplicationLogGroup.getOrCreate runs init() inside INSTANCES.computeIfAbsent.
That has three consequences:
- Blocking under a map lock. init() does HDFS and ZooKeeper work: it creates
the log writer, resolves the HA group state, and may enter STORE_AND_FORWARD.
All of that runs while holding the ConcurrentHashMap bin lock, so other callers
for the same key, and for other keys in the same bin, block behind it.
- A failed init() leaks resources. When init() throws, the group never enters
the map and nothing closes it. Whatever it had already built is left running:
- the Disruptor consumer;
- the open log writer;
- the store-and-forward status task;
- the HA state listeners it registered.
- Failing inits repeat. Each blocked caller runs a failing init() again, so one
bad state, for example an unreachable peer HDFS during a failover, produces a
burst of repeated init attempts and leaks.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)