[
https://issues.apache.org/jira/browse/IGNITE-20199?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Mirza Aliev updated IGNITE-20199:
---------------------------------
Description:
Seems that the current logic in
{{DistributionZoneRebalanceEngine#onUpdateReplicas}} is not correct in terms of
futures chaining. Currently we block configuration notification thread until
all partitions would updates theirs rebalance assignments keys in metastorage.
{code:java}
private CompletableFuture<?>
onUpdateReplicas(ConfigurationNotificationEvent<Integer> replicasCtx) {
...
...
return distributionZoneManager.dataNodes(replicasCtx.storageRevision(),
zoneCfg.zoneId())
.thenCompose(dataNodes -> {
...
for (TableView tableCfg : tableViews) {
...
CompletableFuture<?>[] partitionFutures =
RebalanceUtil.triggerAllTablePartitionsRebalance(...);
tableFutures.add(allOf(partitionFutures));
}
return
allOf(tableFutures.toArray(CompletableFuture[]::new));
});
...
} {code}
As a solution, we could just return completed future in the
{{DistributionZoneRebalanceEngine#onUpdateReplicas}} after we started
asynchronous logic of updating rebalance assignmnets.
was:
After IGNITE-19506 was merged, it added the regression.
Previously we haven't returned future, that updates rebalance pending keys, in
{{{}DistributionZoneRebalanceEngine#onUpdateReplicas{}}}. It was done because
otherwise we would block configuration notification thread until all partitions
would updates theirs rebalance assignments keys in metastorage.
Unfortunately, IGNITE-19506 has brought this blockage
{code:java}
private CompletableFuture<?>
onUpdateReplicas(ConfigurationNotificationEvent<Integer> replicasCtx) {
...
...
return distributionZoneManager.dataNodes(replicasCtx.storageRevision(),
zoneCfg.zoneId())
.thenCompose(dataNodes -> {
...
for (TableView tableCfg : tableViews) {
...
CompletableFuture<?>[] partitionFutures =
RebalanceUtil.triggerAllTablePartitionsRebalance(...);
tableFutures.add(allOf(partitionFutures));
}
return
allOf(tableFutures.toArray(CompletableFuture[]::new));
});
...
} {code}
> Do not chain updating rebalance assignments future in
> DistributionZoneRebalanceEngine#onUpdateReplicas
> -------------------------------------------------------------------------------------------------------
>
> Key: IGNITE-20199
> URL: https://issues.apache.org/jira/browse/IGNITE-20199
> Project: Ignite
> Issue Type: Bug
> Reporter: Mirza Aliev
> Priority: Major
> Labels: ignite-3
>
> Seems that the current logic in
> {{DistributionZoneRebalanceEngine#onUpdateReplicas}} is not correct in terms
> of futures chaining. Currently we block configuration notification thread
> until all partitions would updates theirs rebalance assignments keys in
> metastorage.
>
> {code:java}
> private CompletableFuture<?>
> onUpdateReplicas(ConfigurationNotificationEvent<Integer> replicasCtx) {
> ...
> ...
> return
> distributionZoneManager.dataNodes(replicasCtx.storageRevision(),
> zoneCfg.zoneId())
> .thenCompose(dataNodes -> {
> ...
> for (TableView tableCfg : tableViews) {
> ...
> CompletableFuture<?>[] partitionFutures =
> RebalanceUtil.triggerAllTablePartitionsRebalance(...);
> tableFutures.add(allOf(partitionFutures));
> }
> return
> allOf(tableFutures.toArray(CompletableFuture[]::new));
> });
> ...
> } {code}
> As a solution, we could just return completed future in the
> {{DistributionZoneRebalanceEngine#onUpdateReplicas}} after we started
> asynchronous logic of updating rebalance assignmnets.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
