KAFKA-19164: Keep track of groups when deleting transactional offsets

[squah-confluent]

When deleting pending transactional offsets, we must preserve the list
of groups associated with the producer ID, otherwise we cannot clean up
the list of pending transactions for the group once the transaction is
committed or aborted.

[squah-confluent]

We use offsetsByGroup to clean up the list of open transactions per group.

kafka/group-coordinator/src/main/java/org/apache/kafka/coordinator/group/OffsetMetadataManager.java

Lines 1046 to 1054 in 6e26ec0

	pendingOffsets.offsetsByGroup.keySet().forEach(groupId -> {
	TimelineHashSet<Long> openTransactions = openTransactionsByGroup.get(groupId);
	if (openTransactions != null) {
	openTransactions.remove(producerId);
	if (openTransactions.isEmpty()) {
	openTransactionsByGroup.remove(groupId);
	}
	}
	});

If we do not maintain an entry here, the clean up will not happen. Later, we will not delete the group when we expire its last offset.

kafka/group-coordinator/src/main/java/org/apache/kafka/coordinator/group/OffsetMetadataManager.java

Lines 887 to 888 in 6e26ec0

	// We don't want to remove the group if there are ongoing transactions.
	return allOffsetsExpired.get() && !openTransactionsByGroup.containsKey(groupId);

[jolshan]

I guess we could also just remove from the openTransactionsByGroup here as well? Not a strong opinion either way. If we do it this way we should also clean it up when we remove from openTransactionByGroup? I don't think we want to preserve it forever?

[squah-confluent]

I guess we could. It depends on whether we are trying to track whether a group has an open transaction with offset commits that will take effect when committed or whether a group has an open transaction at all.

We certainly shouldn't delete groups with uncommitted offsets, since they could be committed afterwards and would then not get cleaned up. I'm just not sure about groups with deleted uncommitted offsets that are a no-op when committed.

@dajac What should we be doing in the second case?

[squah-confluent]

Closing this since #19497 fixes the bug.