HDDS-4622: Use singe server raft cluster in MiniOzoneCluster. #1738
Conversation
GlenGeng-awx
force-pushed
the
HDDS-4622
branch
from
d32258e to
5079c33
Compare
|
O seems like we are more or less doing the similar change :) #1733 |
| if (Strings.isNullOrEmpty(storageDir)) { | ||
| storageDir = ServerUtils.getDefaultRatisDirectory(conf); | ||
| File metaDirPath = ServerUtils.getOzoneMetaDirPath(conf); | ||
| storageDir = (new File(metaDirPath, "scm-ha")).getPath(); |
There was a problem hiding this comment.
I think it's prefered to use SCMHA Configuration to set the storageDir
We can fix this issue together, and you can take away my changes if needed. |
| public static StorageContainerManager getScmSimple(OzoneConfiguration conf) | ||
| throws IOException, AuthenticationException { | ||
| SCMConfigurator configurator = new SCMConfigurator(); | ||
| configurator.setSCMHAManager(MockSCMHAManager.getInstance(true)); |
There was a problem hiding this comment.
You will need this line:
conf.setBoolean(ScmConfigKeys.OZONE_SCM_HA_ENABLE_KEY, true);
SCMHAManagerImpl will return isLeader immediately without enabling this HA:
if (!SCMHAUtils.isSCMHAEnabled(conf)) {
// When SCM HA is not enabled, the current SCM is always the leader.
return Optional.of((long)0);
}
There was a problem hiding this comment.
So currently the cluster.waitForClusterToBeReady returns quickly even the SCM has not become leader yet.
I am guessing CI can still pass become you also have changed the rpc timeout parameters. So either the cluster wait for ready or the parameter change can solve the not leader problem.
I am not sure if we need both. Maybe we only need to keep cluster.waitForClusterToBeReady?
| Rpc.setRequestTimeout(properties, TimeDuration.valueOf( | ||
| conf.getRatisRequestTimeout(), TimeUnit.MILLISECONDS)); | ||
| Rpc.setTimeoutMin(properties, TimeDuration.valueOf( | ||
| conf.getRatisRequestMinTimeout(), TimeUnit.MILLISECONDS)); |
There was a problem hiding this comment.
I think you don't need to update these parameters. The cluster.waitForClusterToBeReady is supposed to wait for SCM become leader thus solve the not leader exception.
There was a problem hiding this comment.
You are right, we only need the wait leader in cluster.waitForClusterToBeReady. The problem here is, we've set the min/max timeout un-properly. I will create a separate jira for his fix.
| lock.readLock().lock(); | ||
| try { | ||
| return stateManager | ||
| return stateManager |
There was a problem hiding this comment.
so the dead lock solution seems is removing the read write lock? Is there an explanation why we can remove locks here?
There was a problem hiding this comment.
I will create a jira for this issue, can give the explanation of this solution there.
2 participants
What changes were proposed in this pull request?
Use
SCMHAManagerImplandSCMRatisServerImplinstead ofMockSCMHAManagerandMockRatisServerinMiniOzoneCluster.The
MockRatisServerandMockSCMHAManagercan not fully test the HA code.What is the link to the Apache JIRA
https://issues.apache.org/jira/browse/HDDS-4622
How was this patch tested?
CI