yashmayya commented on code in PR #14005:
URL: https://github.com/apache/kafka/pull/14005#discussion_r1263619000
##########
connect/mirror/src/test/java/org/apache/kafka/connect/mirror/integration/MirrorConnectorsIntegrationExactlyOnceTest.java:
##########
@@ -46,4 +51,49 @@ public void startClusters() throws Exception {
super.startClusters();
}
+ @Override
+ @Test
+ public void testReplication() throws Exception {
+ super.testReplication();
+
+ // Augment the base replication test case with some extra testing of
the offset management
+ // API introduced in KIP-875
+ // We do this only when exactly-once support is enabled in order to
avoid having to worry about
+ // zombie tasks producing duplicate records and/or writing stale
offsets to the offsets topic
+
+ String backupTopic1 = remoteTopicName("test-topic-1",
PRIMARY_CLUSTER_ALIAS);
+ String backupTopic2 = remoteTopicName("test-topic-2",
PRIMARY_CLUSTER_ALIAS);
+
+ // Explicitly move back to offset 0
+ // Note that the connector treats the offset as the last-consumed
offset,
+ // so it will start reading the topic partition from offset 1 when it
resumes
+ alterMirrorMakerSourceConnectorOffsets(backup, n -> 0L,
"test-topic-1");
+ // Reset the offsets for test-topic-2
+ resetSomeMirrorMakerSourceConnectorOffsets(backup, "test-topic-2");
+ resumeMirrorMakerConnectors(backup, MirrorSourceConnector.class);
+
+ int expectedRecordsTopic1 = NUM_RECORDS_PRODUCED +
((NUM_RECORDS_PER_PARTITION - 1) * NUM_PARTITIONS);
+ assertEquals(expectedRecordsTopic1,
backup.kafka().consume(expectedRecordsTopic1, RECORD_TRANSFER_DURATION_MS,
backupTopic1).count(),
+ "Records were not re-replicated to backup cluster after
altering offsets.");
+ int expectedRecordsTopic2 = NUM_RECORDS_PER_PARTITION * 2;
+ assertEquals(expectedRecordsTopic2,
backup.kafka().consume(expectedRecordsTopic2, RECORD_TRANSFER_DURATION_MS,
backupTopic2).count(),
+ "New topic was not re-replicated to backup cluster after
altering offsets.");
+
+ @SuppressWarnings({"unchecked", "rawtypes"})
+ Class<? extends Connector>[] connectorsToReset =
CONNECTOR_LIST.toArray(new Class[0]);
+ // Resetting the offsets for the heartbeat and checkpoint connectors
doesn't have any effect
+ // on their behavior, but users may want to wipe offsets from them to
prevent the offsets topic
+ // from growing infinitely. So, we include them in the list of
connectors to reset as a sanity check
Review Comment:
> (As noted in the comment here) Users may want to tombstone offsets for
these connectors to prevent the offsets topic from growing infinitely
Since the set of source partitions is limited here, shouldn't log compaction
be good enough?
> We have no way of knowing if the set of offsets given to us is the total
set or not, so we can't choose to only allow total resets instead of partial
resets
Yeah, that makes sense.
> Since partial resets become possible, it seems reasonable to allow users
to "undo" a partial reset by re-submitting an offset for a given partition
If these offsets are never read back and actually used, why would users want
to "undo" partial or complete resets?
> Also, I think the UnsupportedOperationException case was more intended for
when offsets are managed externally, not for when it doesn't seem to make sense
to modify them (because, e.g., they're never read back by the connector or its
tasks).
>
> So with all this in mind, I figured it'd be best to allow offsets to be
modified by users, but only if they match the format of the offsets that the
connector's tasks already emit. Thoughts?
Hm, I still feel like if offset modification isn't really doing anything
here it might be more intuitive to simply disallow it rather than informing
users that the offsets have been modified successfully but with no actual
side-effect. However, I don't think too many users are going to be attempting
to modify the offsets for `MirrorCheckpointConnector` /
`MirrorHeartbeatConnector` and expecting any change in their behavior so I'm
not too fussed about this either way.
> As an aside... I'm wondering if we should abandon all kinds of validation
when the proposed partition/offset pair is a tombstone. Just in case the
offsets topic somehow ends up with garbage in it, it'd be nice to allow users
to clean it up via the REST API, and if there's no existing partition/offset
pair in the offsets topic for the given partition already, then emitting a
tombstone effectively becomes a no-op
Hm this is an interesting one. I think if the offsets topic ends up with
garbage in it through an external producer, users could always clean it up
using the same external producer. If we don't let users write garbage using the
offsets REST API, maybe we shouldn't let them clean up garbage using the REST
API either? Overall, I think I'm more in favor of only allowing standard use
cases with the offsets management REST APIs to reduce the potential number of
footguns for users.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: jira-unsubscr...@kafka.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
