[
https://issues.apache.org/jira/browse/ZOOKEEPER-4419?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kezhu Wang reassigned ZOOKEEPER-4419:
-------------------------------------
Assignee: Haoze Wu
> The potential exception in Learner$LeaderConnector.connectToLeader may cause
> unnecessary re-election or service delay
> ---------------------------------------------------------------------------------------------------------------------
>
> Key: ZOOKEEPER-4419
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-4419
> Project: ZooKeeper
> Issue Type: Bug
> Components: server
> Affects Versions: 3.6.2
> Reporter: Haoze Wu
> Assignee: Haoze Wu
> Priority: Major
> Labels: pull-request-available
> Time Spent: 40m
> Remaining Estimate: 0h
>
> We were doing some testing in ZooKeeper 3.6.2, and found that sometimes a
> follower may take longer time (about 10s) to join the quorum. In our cluster,
> usually we expect that the leader and followers are ready to provide service
> in 1 or 2 seconds.
> *Analysis*
> We did some investigations and found that the root cause is the exception
> thrown from `Learner.createSocket`
> ([https://github.com/apache/zookeeper/blob/release-3.6.2/zookeeper-server/src/main/java/org/apache/zookeeper/server/quorum/Learner.java#L410-L423]).
> `Learner.createSocket` is used in `Learner$LeaderConnector.connectToLeader`.
> The code snippet
> ([https://github.com/apache/zookeeper/blob/release-3.6.2/zookeeper-server/src/main/java/org/apache/zookeeper/server/quorum/Learner.java#L344-L407])
> is as follows:
> {code:java}
> private Socket connectToLeader() throws IOException, X509Exception,
> InterruptedException {
> Socket sock = createSocket(); // line 345
> // ...
> int remainingTimeout;
> long startNanoTime = nanoTime();
> for (int tries = 0; tries < 5 && socket.get() == null; tries++) {
> try {
> remainingTimeout = connectTimeout - (int) ((nanoTime() -
> startNanoTime) / 1_000_000);
> if (remainingTimeout <= 0) {
> LOG.error(...);
> throw new IOException("connectToLeader exceeded on
> retries.");
> }
> sockConnect(sock, address, Math.min(connectTimeout,
> remainingTimeout));
> // ...
> break;
> } catch (IOException e) {
> remainingTimeout = connectTimeout - (int) ((nanoTime() -
> startNanoTime) / 1_000_000);
> if (remainingTimeout <= leaderConnectDelayDuringRetryMs) {
> LOG.error(...);
> throw e;
> } else if (tries >= 4) {
> LOG.error(...);
> throw e;
> } else {
> LOG.warn(...);
> sock = createSocket(); // line 400
> }
> }
> Thread.sleep(leaderConnectDelayDuringRetryMs);
> }
> return sock;
> }
> {code}
> The retry (5 times) in this code implies that sometimes exceptions are
> expected to appear, and the program should proceed to try again. However,
> `Learner.createSocket` is invoked in either line 345 or line 400. And it may
> also throw exceptions. But it’s not within the try block. Therefore, this
> exception will turn to `Follower.followLeader` (the caller of
> `Learner$LeaderConnector.connectToLeader`). And then this server will change
> from FOLLOWING state to LOOKING state.
> `Learner.createSocket` may throw exceptions in line 417 or 421:
> {code:java}
> /**
> * Creating a simple or and SSL socket.
> * This can be overridden in tests to fake already connected sockets for
> connectToLeader.
> */
> protected Socket createSocket() throws X509Exception, IOException {
> Socket sock;
> if (self.isSslQuorum()) {
> sock = self.getX509Util().createSSLSocket(); // line 417
> } else {
> sock = new Socket();
> }
> sock.setSoTimeout(self.tickTime * self.initLimit); // line 421
> return sock;
> } {code}
> The reason for throwing the exception may be
> [https://stackoverflow.com/questions/22423063/java-exception-on-sslsocket-creation]
> or
> [https://stackoverflow.com/questions/7064788/when-does-java-net-socket-setsotimeout-throws-socketexception.]
> We reproduced this bug by intentionally injecting an IOException right before
> the invocation of `Learner.createSocket` in line 345. We only grant one
> single injection in a testing run of a ZooKeeper cluster (3 nodes). Our
> analysis proves to be true. In terms of the symptom, sometimes the affected
> follower takes more time to join the quorum; sometimes the leader election is
> also affected, and the whole cluster does the re-election, meaning that every
> node is affected.
> *Fix*
> We propose that the `Learner.createSocket` should be moved to the try-catch
> block. We have created a pull request.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
