[
https://issues.apache.org/jira/browse/RATIS-2055?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17838365#comment-17838365
]
Tsz-wo Sze commented on RATIS-2055:
-----------------------------------
The compatibility problem is easy to fix in Ratis (although it is a bug in
Ozone since SCM uses notifyTermIndexUpdated(..) to set the leader is ready. The
correct way is to use notifyLeaderReady() as [~Symious] mentioned.) We may
move down the notifyTermIndexUpdated(..) call as below. Would you like to
submit a pull request?
{code}
+++
b/ratis-server/src/main/java/org/apache/ratis/server/impl/RaftServerImpl.java
@@ -1805,9 +1805,6 @@ class RaftServerImpl implements RaftServer.Division,
CompletableFuture<Message>
applyLogToStateMachine(ReferenceCountedObject<LogEntryProto> nextRef)
throws RaftLogIOException {
LogEntryProto next = nextRef.get();
- if (!next.hasStateMachineLogEntry()) {
- stateMachine.event().notifyTermIndexUpdated(next.getTerm(),
next.getIndex());
- }
if (next.hasConfigurationEntry()) {
// the reply should have already been set. only need to record
@@ -1830,6 +1827,7 @@ class RaftServerImpl implements RaftServer.Division,
throw new RaftLogIOException(e);
}
}
+ stateMachine.event().notifyTermIndexUpdated(next.getTerm(),
next.getIndex());
return null;
}
{code}
> Changed logic of LeaderState.isReady impacts application
> --------------------------------------------------------
>
> Key: RATIS-2055
> URL: https://issues.apache.org/jira/browse/RATIS-2055
> Project: Ratis
> Issue Type: Bug
> Components: Leader
> Affects Versions: 3.1.0
> Reporter: Duong
> Priority: Major
>
> RATIS-2019 changes the logic of `LeaderState.isReady`. Before, isReady itself
> performs the logic to check if the leader state is ready. After the change,
> the logic is moved to `LeaderState.checkReady` and `isReady` now only reads
> the result.
> This change impacts client applications that rely on `isReady` check. For
> example, Ozone's SCMStateMachine [listens to notifyTermIndexUpdated
> |https://github.com/apache/ozone/blob/5ff79613e546a09f23b2c3bae61d62fb757eaf70/hadoop-hdds/server-scm/src/main/java/org/apache/hadoop/hdds/scm/ha/SCMStateMachine.java#L359]
> and checks if the leader state is ready. However, because the
> `leader.checkReady` is called after the `notifyTermIndexUpdated`,
> SCMStateMachine never sees leader ready state in the intercepted
> `notifyTermIndexUpdated` events.
> {code:java|title=RaftServerImpl.applyLogToStateMachine}
> CompletableFuture<Message>
> applyLogToStateMachine(ReferenceCountedObject<LogEntryProto> nextRef)
> throws RaftLogIOException {
> LogEntryProto next = nextRef.get();
> if (!next.hasStateMachineLogEntry()) {
> stateMachine.event().notifyTermIndexUpdated(next.getTerm(),
> next.getIndex());
> }
> if (next.hasConfigurationEntry()) {
> // the reply should have already been set. only need to record
> // the new conf in the metadata file and notify the StateMachine.
> state.writeRaftConfiguration(next);
> stateMachine.event().notifyConfigurationChanged(next.getTerm(),
> next.getIndex(), next.getConfigurationEntry());
> role.getLeaderState().ifPresent(leader -> leader.checkReady(next));
> ....{code}
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
