Jyothsna Konisa created CASSANALYTICS-184:
---------------------------------------------
Summary: CdcState.ReplicaCountSerializer map-size overflow
corrupts persisted CDC state
Key: CASSANALYTICS-184
URL: https://issues.apache.org/jira/browse/CASSANALYTICS-184
Project: Apache Cassandra Analytics
Issue Type: Bug
Reporter: Jyothsna Konisa
CdcState.ReplicaCountSerializer serializes the number of entries in the CDC
watermarker's replica-count map (numUpdates — the count of distinct mutations
awaiting sufficient replica acknowledgment) using a signed short (max 32,767),
this silently overflows once the map grows past that sizeOnce corrupted. The
bad size gets persisted to Cassandra via SidecarStatePersister and re-read on
every subsequent restart, causing a permanent crash loop.
message="Failed to start CDC consumers, cleaning up resources"
exception=java.lang.IllegalArgumentException: Illegal initial capacity: -4366
at java.util.HashMap.<init>
at
org.apache.cassandra.cdc.state.CdcState$ReplicaCountSerializer.read(CdcState.java:404)
at
org.apache.cassandra.cdc.state.CdcState$Serializer.read(CdcState.java:380)
at org.apache.cassandra.cdc.Cdc.start(Cdc.java:137)
-4366 is the short-overflow wraparound of a map size that grew past 32,767
under sustained backpressure (CDC write-rejection storms inflating the
watermarker as mutations piled up waiting for replica quorum). The corrupted
size persists indefinitely — each restart re-reads the same broken state and
crashes identically, with no self-heal.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
