ChenXi created KAFKA-21481:
------------------------------
Summary: ListTransactions can return transactions from
no-longer-owned partitions
Key: KAFKA-21481
URL: https://issues.apache.org/jira/browse/KAFKA-21481
Project: Kafka
Issue Type: Bug
Components: core
Reporter: ChenXi
h1. Problem and reproduction
TransactionStateManager.listTransactionStates iterates the entire
transactionMetadataCache and returns Errors.NONE. The cache entry is removed
only
by TransactionCoordinator.onResignation, which runs after the replica has
already
become a follower. During that window, a broker can return transactions for a
__transaction_state partition it no longer leads.
Affected source in trunk:
TransactionStateManager.scala:331-392 (listTransactionStates), especially :380:
{code:scala}
transactionMetadataCache.foreachEntry { (_, cache) =>
// partition id discarded; no leadership check
cache.metadataPerTransactionalId.forEach { ... }
}
{code}
TransactionCoordinator.scala:299-310 only checks the process-wide isActive flag:
{code:scala}
if (!isActive.get()) {
... COORDINATOR_NOT_AVAILABLE
} else {
txnManager.listTransactionStates(...)
}
{code}
The leadership helper exists but is used only by the point-lookup path:
{code:scala}
// TransactionStateManager.scala:403-405
private def isCoordinatorFor(partitionId: Int, coordinatorEpoch: Int): Boolean
= ...
// TransactionStateManager.scala:423-424 (getAndMaybeAddTransactionState only)
case Some(cacheEntry) if !isCoordinatorFor(partitionId,
cacheEntry.coordinatorEpoch) =>
Left(Errors.NOT_COORDINATOR)
{code}
Cache removal still happens in onResignation after leadership loss:
TransactionCoordinator.scala:497-505.
Reproduction:
1. Load a transaction metadata cache entry for partition P at coordinator epoch
E.
2. Make the local replica a follower for P, or move its leader epoch away from
E,
while keeping the cache entry present (the pre-onResignation state).
3. Call ListTransactions on that broker.
4. The response has error NONE and still includes P's transactional IDs.
Expected:
P is skipped.
This can be covered by a unit test in TransactionStateManagerTest using the
existing partitionLeadership mock.
h1. Relationship to KAFKA-21181 and difference
This is a follow-up, not a duplicate. KAFKA-21181 added isCoordinatorFor and
applied it only to getAndMaybeAddTransactionState, returning NOT_COORDINATOR for
point lookups, EndTxn, AddPartitions and InitProducerId.
The listTransactionStates path was not updated. It still uses
foreachEntry
{ (_, cache) => ... }
, loses the partition id and never calls
isCoordinatorFor.
The expected semantics differ deliberately: point requests should return
NOT_COORDINATOR so the client rediscovers the coordinator; ListTransactions is a
cluster-wide administrative aggregation and should skip stale partitions while
continuing to return transactions from partitions the broker still owns.
*Proposed fix:* call isCoordinatorFor before reading each cache entry in
listTransactionStates, add regression tests, and document that no wire/API/KIP
change is required.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)