goaymode created IGNITE-25727:
---------------------------------

             Summary: Lock Contention and OOM Risks in 
GridDhtPartitionTopologyImpl and RingMessageWorker
                 Key: IGNITE-25727
                 URL: https://issues.apache.org/jira/browse/IGNITE-25727
             Project: Ignite
          Issue Type: Improvement
          Components: cache, messaging
    Affects Versions: 2.16
            Reporter: goaymode


There is a potential for performance degradation and OOM issues in 
[GridDhtPartitionTopologyImpl.nodes0.|https://github.com/apache/ignite/blob/fb76e9cdd324a8f43009451a5bc81b0806047d53/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/distributed/dht/topology/GridDhtPartitionTopologyImpl.java#L1203]
 Every topology change invalidates cached mappings and forces full iterations 
over node2part, making the cost scale with the number of retained nodes. Also 
since nodes0 is called once per partition, the total cost grows linearly with 
the number of partitions. Each call also creates a new 
[ArrayList|https://github.com/apache/ignite/blob/fb76e9cdd324a8f43009451a5bc81b0806047d53/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/distributed/dht/topology/GridDhtPartitionTopologyImpl.java#L1243]
 of nodes increasing memory pressure if there are many large concurrent 
allocations. All of these run under a read lock, increasing the chance of 
contention and delays during high load or frequent topology changes. 

[ServerImpl$RingMessageWorker.checkFailedNodesList|https://github.com/apache/ignite/blob/fb76e9cdd324a8f43009451a5bc81b0806047d53/modules/core/src/main/java/org/apache/ignite/spi/discovery/tcp/ServerImpl.java#L6251]
 also has a similar performance risk under high load or frequent invocation. A 
synchronized block is being used used in combination with iteration over the 
failedNodes and failedNodesMsgSent collections. As these collections grow 
especially when many nodes fail—each invocation becomes more expensive and 
increases the risk of contention and synchronization bottlenecks.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to