[
https://issues.apache.org/jira/browse/IGNITE-23753?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17904446#comment-17904446
]
Mikhail Efremov commented on IGNITE-23753:
------------------------------------------
The ticket is closed because in process there were found several important
issues and bugs inside {{ReplicaStateManager}} such as:
* Uncovered replica stopping states on table destruction and node stopping.
* Memory leak due to some states aren't removed from context map event after
replica stopping.
* Redundant state creation-deletion process on every replica under rebalance
(even) if a node won't create replica.
* There are no tests, only several strict in-code assertions without proper
description thus it's hard to make changes and fixes then into.
So, there a refactoring should be done, the corresponding epic IGNITE-23931 is
created
> Replica's state doesn't change on node stopping
> -----------------------------------------------
>
> Key: IGNITE-23753
> URL: https://issues.apache.org/jira/browse/IGNITE-23753
> Project: Ignite
> Issue Type: Bug
> Reporter: Mikhail Efremov
> Assignee: Mikhail Efremov
> Priority: Major
> Labels: ignite-3
> Time Spent: 10m
> Remaining Estimate: 0h
>
> *Description*
> There is a possible situation while an Ignite node was being stopped
> {{TableManager}} desroyed all partitions and replicas already, but
> {{ReplicaManager}} isn't even took a busy lock and e.g.
> {{ReplicaStateManager}} after IGNITE-22036 may attemts to get already removed
> replica future. For now we may ignore such {{null}} value, but it's clear
> that the picture is wrong: we already deleted a replica, but it's state isn't
> {{STOPPED}}.
> Also we may found a spagetti-like method calls between {{TableManager}} and
> {{ReplicaManager}} on a replica's stopping process: {{TableManager}} calls
> {{ReplicaManager#weakReplicaStop}} that call the given lambda that actually
> {{TableManager#stopAndDestroyPartition}} that at the end calls
> {{ReplicaManager#stopReplicaInternal}}. Probably we should separate this code.
> *Motivation*
> In the future we may have a logic that get replicas and we might get an
> unconsistent replica's state on a node's stopping process.
> *Definition of Done*
> * New {{WeakStopReason.SHUTDOWN}} is added and is used for node stopping
> situation in {{TableManager}}.
> * In case of {{SHUTDOWN}} reason we must set {{STOPPED}} replica state and
> stop the replica immediately.
> * (optional) Consider an opportunity to refactor
> {{TableManager#stopAndDestroyPartition}} in a looser couple way.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
