[
https://issues.apache.org/jira/browse/ZOOKEEPER-2849?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16256099#comment-16256099
]
ASF GitHub Bot commented on ZOOKEEPER-2849:
-------------------------------------------
Github user afine commented on a diff in the pull request:
https://github.com/apache/zookeeper/pull/419#discussion_r151553425
--- Diff:
src/java/main/org/apache/zookeeper/server/quorum/ExponentialBackoffStrategy.java
---
@@ -0,0 +1,192 @@
+package org.apache.zookeeper.server.quorum;
+
+/**
+ * A {@link BackoffStrategy} that increases the wait time between each
+ * interval up to the configured maximum wait time.
+ */
+public class ExponentialBackoffStrategy implements BackoffStrategy {
+
+ // Sensible default values to use if not set by the user
+ private static final long DEFAULT_INITIAL_BACKOFF_MILLIS = 500L; //
0.5s
+ private static final long DEFAULT_MAX_BACKOFF_MILLIS = 30_000L; // 30s
+ private static final long DEFAULT_MAX_ELAPSED_MILLIS = 5 * 60_000L; //
10m
+ private static final double DEFAULT_BACKOFF_MULTIPLE = 1.5;
+
+ // internal values per instance
+ private final long initialBackoffMillis;
+ private final long maxBackoffMillis;
+ private final long maxElapsedMillis;
+ private final double backoffMultiple;
+
+ // internal state
+ private long nextWait;
+ private long totalElapsed;
+ private final boolean limitBackoffMillis;
+ private final boolean checkElapsedTime;
+
+ /**
+ * Construct a new instance.
+ * @param builder the Builder to use for configuring this
BackoffStrategy
+ */
+ private ExponentialBackoffStrategy(Builder builder) {
+ this.initialBackoffMillis = builder.initialBackoffMillis;
+ this.maxBackoffMillis = builder.maxBackoffMillis;
+ this.maxElapsedMillis = builder.maxElapsedMillis;
+ this.backoffMultiple = builder.backoffMultiple;
+
+ if(maxBackoffMillis == -1) {
+ limitBackoffMillis = false;
+ } else {
+ limitBackoffMillis = true;
+ }
+
+ if(maxElapsedMillis == -1) {
+ checkElapsedTime = false;
+ } else {
+ checkElapsedTime = true;
+ }
+
+ reset();
+ }
+
+
+ @Override
+ public long nextWaitMillis() throws IllegalStateException {
+ // check if we have exceeded the allowed maximum elapsed time
+ if(checkElapsedTime && totalElapsed > maxElapsedMillis) {
+ return BackoffStrategy.STOP;
+ }
+
+ long waitMillis = nextWait;
+
+ // calculate the next wait milliseconds
+ nextWait = Math.round(nextWait * backoffMultiple);
+
+ // don't exceed the allowed maximum wait milliseconds
+ // if a maximum was configured
+ if(limitBackoffMillis && nextWait > maxBackoffMillis) {
+ nextWait = maxBackoffMillis;
+ }
+
+ // track total elapsed time, even if we don't wait we have to
assume
+ // that some amount of time passed outside of the wait or we'll
never
+ // hit the elapsed time limit
+ totalElapsed += waitMillis != 0 ? waitMillis : 1L;
+ return waitMillis;
+ }
+
+ @Override
+ public void reset() {
+ nextWait = this.initialBackoffMillis;
+ totalElapsed = 0;
+ }
+
+ /**
+ *
+ * @return a new {@link Builder} instance.
+ */
+ public static Builder builder() {
+ return new Builder();
+ }
+
+ /**
+ * Builder for instances of {@link ExponentialBackoffStrategy}.
+ */
+ public static final class Builder {
--- End diff --
I'm not sure a builder is the best way to handle this. Since I think it
would be nice to have much of this be user configurable, perhaps we can just
pull the config from system properties?
> Quorum port binding needs exponential back-off retry
> ----------------------------------------------------
>
> Key: ZOOKEEPER-2849
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2849
> Project: ZooKeeper
> Issue Type: Improvement
> Components: quorum
> Affects Versions: 3.4.6, 3.5.3
> Reporter: Brian Lininger
> Assignee: Brian Lininger
> Priority: Minor
>
> Recently we upgraded the AWS instance type we use for running out ZooKeeper
> nodes, and by doing so we're intermittently hitting an issue where ZooKeeper
> cannot bind to the server election port because the IP is incorrect. This is
> due to name resolution in Route53 not being in sync when ZooKeeper starts on
> the more powerful EC2 instances. Currently in QuorumCnxManager.Listener, we
> only attempt to bind 3 times with a 1s sleep between retries, which is not
> long enough.
> I'm proposing to change this to follow an exponential back-off type strategy
> where each failed attempt causes a longer sleep between retry attempts. This
> would allow for Zookeeper to gracefully recover when the host is
> misconfigured, and subsequently corrected, without requiring the process to
> be restarted while also minimizing the impact to the running instance.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
