[
https://issues.apache.org/jira/browse/HDFS-16890?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17686119#comment-17686119
]
ASF GitHub Bot commented on HDFS-16890:
---------------------------------------
omalley commented on code in PR #5298:
URL: https://github.com/apache/hadoop/pull/5298#discussion_r1100744979
##########
hadoop-hdfs-project/hadoop-hdfs-rbf/src/main/java/org/apache/hadoop/hdfs/server/federation/router/RouterRpcClient.java:
##########
@@ -1730,4 +1755,30 @@ private static boolean isReadCall(Method method) {
}
return !method.getAnnotationsByType(ReadOnly.class)[0].activeOnly();
}
+
+ /**
+ * Checks and sets last refresh time for a namespace's stateId.
+ * Returns true if refresh time is newer than threshold.
+ * Otherwise, return false and call should be handled by active namenode.
+ * @param nsId namespaceID
+ */
+ @VisibleForTesting
+ boolean isNamespaceStateIdFresh(String nsId) {
+ if (activeNNStateIdRefreshPeriodMs < 0) {
+ return true;
+ }
+
+ long currentTimeMs = Time.monotonicNow();
+ LongAccumulator latestRefreshTimeMs = lastActiveNNRefreshTimes
+ .computeIfAbsent(nsId, key -> new LongAccumulator(Math::max, 0));
+
+ return ((currentTimeMs - latestRefreshTimeMs.get()) <=
activeNNStateIdRefreshPeriodMs);
+ }
+
+ private void refreshTimeOfLastCallToActiveNameNode(String namespaceId) {
+ LongAccumulator latestRefreshTimeMs = lastActiveNNRefreshTimes
Review Comment:
I'd suggest putting this common code into a method like:
LongAccumulator getLastCallToActive(String nsId) { ... }
Then the other lines could mostly be inline:
return currentTimeMs - getLastCallToActive(nsId).get() <=
activeNNStateIdRefreshPeriodMs;
and
getLastCallToActive(nsId).accumulate(Time.monotonicNow());
##########
hadoop-hdfs-project/hadoop-hdfs-rbf/src/main/java/org/apache/hadoop/hdfs/server/federation/router/RouterRpcClient.java:
##########
@@ -1702,10 +1724,12 @@ private List<? extends FederationNamenodeContext>
getOrderedNamenodes(String nsI
boolean isObserverRead) throws IOException {
final List<? extends FederationNamenodeContext> namenodes;
- if (RouterStateIdContext.getClientStateIdFromCurrentCall(nsId) >
Long.MIN_VALUE) {
+ if (isNamespaceStateIdFresh(nsId)
+ && (RouterStateIdContext.getClientStateIdFromCurrentCall(nsId) >
Long.MIN_VALUE)) {
namenodes = namenodeResolver.getNamenodesForNameserviceId(nsId,
isObserverRead);
} else {
namenodes = namenodeResolver.getNamenodesForNameserviceId(nsId, false);
+ refreshTimeOfLastCallToActiveNameNode(nsId);
Review Comment:
Shouldn't this be updated on whether we went to the active? In other words,
if isObserverRead is false, this should be updated also.
> RBF: Add period state refresh to keep router state near active namenode's
> -------------------------------------------------------------------------
>
> Key: HDFS-16890
> URL: https://issues.apache.org/jira/browse/HDFS-16890
> Project: Hadoop HDFS
> Issue Type: Task
> Reporter: Simbarashe Dzinamarira
> Assignee: Simbarashe Dzinamarira
> Priority: Major
> Labels: pull-request-available
>
> When using the ObserverReadProxyProvider, clients can setÂ
> *dfs.client.failover.observer.auto-msync-period...* to periodically get the
> Active namenode's state. When using routers without the
> ObserverReadProxyProvider, this periodic update is lost.
> In a busy cluster, the Router constantly gets updated with the active
> namenode's state when
> # There is a write operation.
> # There is an operation (read/write) from a new clients.
> However, in the scenario when there are no new clients and no write
> operations, the state kept in the router can lag behind the active's. The
> router does update its state with responses from the Observer, but the
> observer may be lagging behind too.
> We should have a periodic refresh in the router to serve a similar role as
> *dfs.client.failover.observer.auto-msync-period*
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
