[
https://issues.apache.org/jira/browse/IGNITE-20317?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Mirza Aliev updated IGNITE-20317:
---------------------------------
Epic Link: IGNITE-20611 (was: IGNITE-20166)
> Meta storage invokes are not completed when events are handled in DZM
> ----------------------------------------------------------------------
>
> Key: IGNITE-20317
> URL: https://issues.apache.org/jira/browse/IGNITE-20317
> Project: Ignite
> Issue Type: Bug
> Reporter: Sergey Uttsel
> Assignee: Mirza Aliev
> Priority: Major
> Labels: ignite-3
>
> h3. *Motivation*
> There are meta storage invokes in DistributionZoneManager in zone's
> lifecycle. The futures of these invokes are ignored, so after the lifecycle
> method is completed actually not all its actions are completed. Therefore
> several invokes for example on createZone and alterZone can be reordered.
> Currently it does the meta storage invokes in:
> # ZonesConfigurationListener#onCreate to init a zone.
> # ZonesConfigurationListener#onDelete to clean up the zone data.
> # DistributionZoneManager#onUpdateFilter to save data nodes in the meta
> storage.
> # DistributionZoneManager#onUpdateScaleUp
> # DistributionZoneManager#onUpdateScaleDown
> -DistributionZoneRebalanceEngine#onUpdateReplicas to apdate assignment on
> replicas update.-
> -LogicalTopologyEventListener to update logical topology.-
> -DistributionZoneRebalanceEngine#createDistributionZonesDataNodesListener
> watch listener to update pending assignments.-
> h3. *Definition of Done*
> Need to ensure event handling linearization. All immediate data nodes
> recalculation must be returned to the event handler.
> h3. *Implementation Notes*
> * ZonesConfigurationListener#onCreate, ZonesConfigurationListener#onDelete,
> DistributionZoneManager#onUpdateFilter and
> DistributionZoneRebalanceEngine#onUpdateReplicas are invoked in configuration
> listeners. So we can just return the ms invoke future from these methods
> and it ensure, that this invoke will be completed within the current event
> handling.
> * We cannnot return future from LogicalTopologyEventListener's methods. We
> can ignore these futures. It has drawback: we can skip the topology update
> # topology=[A,B], dataNodes=[A,B], scaleUp=0, scaleDown=100
> # Node C was joined to the topology and left quickly and ms invokes to update
> topology entry was reordered.
> # data nodes was not updated immediately to [A,B,C].
> We think that we can ignore this bug because eventually it doesn't break the
> consistency of the date node. For this purpose we need to change the invoke
> condition:
> `value(zonesLogicalTopologyVersionKey()).lt(longToBytes(newTopology.version()))`
> instead of
> `value(zonesLogicalTopologyVersionKey()).eq(longToBytes(newTopology.version()
> - 1))`
> * Need to return ms invoke futures from WatchListener#onUpdate method of the
> data nodes listener.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
