[
https://issues.apache.org/jira/browse/IGNITE-22315?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17856838#comment-17856838
]
Mikhail Efremov commented on IGNITE-22315:
------------------------------------------
Actual PR link: https://github.com/apache/ignite-3/pull/3956
> Make raft-client starting only once and only with raft-client and replica
> together
> ----------------------------------------------------------------------------------
>
> Key: IGNITE-22315
> URL: https://issues.apache.org/jira/browse/IGNITE-22315
> Project: Ignite
> Issue Type: Improvement
> Reporter: Mikhail Efremov
> Assignee: Mikhail Efremov
> Priority: Major
> Labels: ignite-3
> Time Spent: 10m
> Remaining Estimate: 0h
>
> *Description*
> The goal is to make sure that when partition starting on an ignite-node, if
> the node in assignments, then there should be only ones raft-node,
> raft-client and replica and no any started entities otherwise.
> *Motivation*
> Now on the one hand in {{TableManager}} there is a path when raft-node and
> replica didn't start, but raft-client is:
> In {{TableManager#startPartitionAndStartClient}} there is a place where
> {{startGroupFut}} returns {{{}false{}}}:\{{ localMemberAssignment}} is
> {{null}} or {{PartitionReplicatorNodeRecovery#shouldStartGroup}} returns
> {{{}false{}}}. Then, raft-node and replica will not be started, but
> regardless of {{startGroupFut}} result raft client will be started:
> {code:java}
> startGroupFut
> .thenComposeAsync(v -> inBusyLock(busyLock, () -> {
> TableRaftService tableRaftService =
> table.internalTable().tableRaftService();
> try {
> // Return existing service if it's already started.
> return completedFuture(
> (TopologyAwareRaftGroupService)
> tableRaftService.partitionRaftGroupService(replicaGrpId.partitionId())
> );
> } catch (IgniteInternalException e) {
> // We use "IgniteInternalException" in accordance with the
> javadoc of "partitionRaftGroupService" method.
> try {
> // TODO IGNITE-19614 This procedure takes 10 seconds if
> there's no majority online.
> return raftMgr
> .startRaftGroupService(replicaGrpId,
> newConfiguration, raftGroupServiceFactory, raftCommandsMarshaller);
> } catch (NodeStoppingException ex) {
> return failedFuture(ex);
> }
> }
> }), ioExecutor)
> .thenAcceptAsync(updatedRaftGroupService -> inBusyLock(busyLock, ()
> -> {
> ((InternalTableImpl) internalTbl).tableRaftService()
> .updateInternalTableRaftGroupService(partId,
> updatedRaftGroupService);
> boolean startedRaftNode = startGroupFut.join();
> if (localMemberAssignment == null || !startedRaftNode ||
> replicaMgr.isReplicaStarted(replicaGrpId)) {
> return;
> } {code}
> the code shows that {{v}} argument ({{{}startGroupFut{}}}'s result) in the
> lambda doesn't affect on raft-client's starting, and also the client is put
> into {{TableRaftService}} then regardless of replica's starting.
> On the other hand, there is a place that rely on raft-client creation on
> every node with table's ID in {{TableManager#tables}} map: inside
> {{TableManager#handleChangePendingAssignmentEvent}} there are two calls:
> # the same name method
> {{{}TableManager#handleChangePendingAssignmentEvent{}}}{{{{}}{}}}
> # {{TableManager#changePeersOnRebalance}}
> Both are asking for internal table's
> {{{}tableRaftService().partitionRaftGroupService(partId){}}}, but there no
> any checks about is the local node is suitable for raft-entities and replica
> starting or they are already started.
>
> Then, when we would try to fix instant raft-client starting it will lead to
> instant {{IgniteInternalException}} with _"No such partition P in table T"_
> from {{{}TableRaftService#{}}}{{{}partitionRaftGroupService{}}}. This is a
> problem that should be solved.
> {color:#383838}*Definition of done*{color}
> # {color:#383838}Raft-client must be started only together with raft-node
> and replica or shouldn't be started otherwise.{color}
> # {color:#383838}{{TableRaftService}} shouldn't be requested for raft-client
> if the local node isn't in a raft group.{color}
> # {color:#383838}Some tests may rely on described behavior, then failed
> tests should be investigated and, probably fixed.{color}
> # {color:#383838}New tests that check single starting of raft-node,
> raft-client and replica together are required{color}
> {color:#0f54d6} {color}
> {color:#383838} {color}
> {color:#00855f} {color}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
