The jmh LRUCacheBenchmark will exhibit an int overflow when run on a fast machine:
```
java.lang.ArrayIndexOutOfBoundsException: Index -3648 out of bounds for length 10000
at org.apache.kafka.jmh.cache.LRUCacheBenchmark.testCachePerformance(LRUCacheBenchmark.java:70)
at org.apache.kafka.jmh.cache.generated.LRUCacheBenchmark_testCachePerformance_jmhTest.testCachePerformance_thrpt_jmhStub(LRUCacheBenchmark_testCachePerformance_jmhTest.java:119)
at org.apache.kafka.jmh.cache.generated.LRUCacheBenchmark_testCachePerformance_jmhTest.testCachePerformance_Throughput(LRUCacheBenchmark_testCachePerformance_jmhTest.java:83)
```
Reviewers: Jason Gustafson <jason@confluent.io>
As part of commit 4d1ee26a136997d31dbd6ddca07e09b34c41c77d streams
version 2.3.0 test jar was added, but there was a simple typo in the
path that specified the version.
`ducker-ak up` was failing because of that. Fixed that.
Reviewers: Guozhang Wang <wangguoz@gmail.com>
Reviewers: cpettitt-confluent <53191309+cpettitt-confluent@users.noreply.github.com>, A. Sophie Blee-Goldman <sophie@confluent.io>, Guozhang Wang <wangguoz@gmail.com>
Currently on commit streams will attempt to delete offsets from repartition topics. However, if a topology does not have any repartition topics, then the recordsToDelete map will be empty.
This PR adds a check that the recordsToDelete is not empty before executing the AdminClient#deleteRecords() method.
Reviewers: A. Sophie Blee-Goldman <sophie@confluent.io>, Guozhang Wang <wangguoz@gmail.com>
- Call `assertNoNonDaemonThreads` in test method instead of tear down method
to avoid situation where parent's class tear down is not invoked.
- Pass the thread prefix in tests that call `assertNoNonDaemonThreads` so that it
works correctly.
- Rename `verifyNonDaemonThreadsStatus` to `assertNoNonDaemonThreads` to
make it clear that it may throw.
Reviewers: Anna Povzner <anna@confluent.io>, Jason Gustafson <jason@confluent.io>, Ismael Juma <ismael@juma.me.uk>
This patch fixes the quota system test whose JMX tool relies on the existence
of these metrics.
Reviewers: Guozhang Wang <wangguoz@gmail.com>, Nikhil Bhatia <nikhil@confluent.io>, Tu V. Tran <tuvtran97@gmail.com>, Ismael Juma <ismael@juma.me.uk>
Splits the existing StickyAssignor logic into an AbstractStickyAssignor class, which is extended by the existing (eager) StickyAssignor and by the new CooperativeStickyAssignor which supports incremental cooperative rebalancing.
There is no actual change to the logic -- most methods from StickyAssignor were moved to AbstractStickyAssignor to be shared with CooperativeStickyAssignor, and the abstract MemberData memberData(Subscription) method converts the Subscription to the embedded list of owned partitions for each assignor.
The "generation" logic is left in, however this is always Optional.empty() for the CooperativeStickyAssignor as onPartitionsLost should always be called when a generation is missed.
Reviewers: Jason Gustafson <jason@confluent.io>, Guozhang Wang <wangguoz@gmail.com>
Make sure to show the message key, even when the message value is null.
This changes the output of one of the tools. Is the output of the tool considered a public API? Does this need a discussion or a KIP?
Testing: Ran the tool on a compacted topic. Previously, the tool did not show any message keys for tombstone messages (messages where the value is null). Now, the tool shows message keys.
Reviewers: Mickael Maison <mimaison@users.noreply.github.com>, Guozhang Wang <wangguoz@gmail.com>
When sending bad records, the Trogdor task will fail if the final record produced is bad. Instead we should catch the exception to allow the task to finish since sending bad records is a valid use case.
Reviewers: Tu V. Tran <tuvtran97@gmail.com>, Guozhang Wang <wangguoz@gmail.com>
RocksDB metrics are added to the Kafka metrics. For each segmented state store only
one set of metrics is exposed rather than one set of metrics for each segment.
The metrics are not computed yet.
Reviewers: John Roesler <john@confluent.io>, Guozhang Wang <guozhang@confluent.io>
Make offsets immutable to users of RecordCollector.offsets. Fix up an
existing case where offsets could be modified in this way. Add a simple
test to verify offsets cannot be changed externally.
Reviewers: Bruno Cadonna <bruno@confluent.io>, Guozhang Wang <guozhang@confluent.io>, Matthias J. Sax <matthias@confluent.io>
Prior to this change an NPE is raised when calling AssignedTasks.close
under the following conditions:
1. EOS is enabled
2. The task was in a suspended state
The cause for the NPE is that when a clean close is requested for a
StreamTask the StreamTask tries to commit. However, in the suspended
state there is no producer so ultimately an NPE is thrown for the
contained RecordCollector in flush.
The fix put forth in this commit is to have AssignedTasks call
closeSuspended when it knows the underlying StreamTask is suspended.
Note also that this test is quite involved. I could have just tested
that AssignedTasks calls closeSuspended when appropriate, but that is
testing, IMO, a detail of the implementation and doesn't actually verify
we reproduced the original problem as it was described. I feel much more
confident that we are reproducing the behavior - and we can test exactly
the conditions that lead to it - when testing across AssignedTasks and
StreamTask. I believe this is an additional support for the argument of
eventually consolidating the state split across classes.
Reviewers: Matthias J. Sax <matthias@confluent.io>, Guozhang Wang <wangguoz@gmail.com>
I've spent quite a bit of time on trying to discover the root cause, with no luck so far. I have been able to reproduce it locally by running the following 100 times:
```
./gradlew connect:runtime:clean connect:runtime:test --tests org.apache.kafka.connect.integration.RebalanceSourceConnectorsIntegrationTest
```
The `testReconfigConnector` test failed 28% of the time and the others failed 0%. This issue and KAFKA-8661 suggest that `testDeleteConnector` and `testStartTwoConnectors` are also flaky, though I've not seen those tests fail locally.
Because this flakiness is causing issues for the rest of the project, I'm going to temporarily ignore several of the flaky ITs while I continue to investigate:
* `RebalanceSourceConnectorsIntegrationTest.testReconfigConnector`
* `RebalanceSourceConnectorsIntegrationTest.testDeleteConnector`
* `RebalanceSourceConnectorsIntegrationTest.testStartTwoConnectors`
**This should be backported to the `2.3` branch, which is when these integration tests were first added.**
Author: Randall Hauch <rhauch@gmail.com>
Reviewers: Ismael Juma
Closes#7237 from rhauch/kafka-8391-temporary
Changed Connect's `WorkerSourceTask` to capture non-retriable exceptions from the `producer.send(...)` (e.g., authentication or authorization errors) and to fail the connector task when such an error is encountered. Modified the existing unit tests to verify this functionality.
Note that most producer errors are retriable, and Connect will (by default) set up each producer with 1 max in-flight message and infinite retries. This change only affects non-retriable errors.
Move the error code resetting logic from the onPartitionsRevoked callback into the streamthread directly after we've decided to rejoin the group, since onPartitionsRevoked are not guaranteed to be triggered.
Ran system tests on the originally failed StreamsUpgradeTest 10 times and passed.
Reviewers: A. Sophie Blee-Goldman <sophie@confluent.io>, Jun Rao <junrao@gmail.com>
This is the implementation for [KIP-503](https://cwiki.apache.org/confluence/display/KAFKA/KIP-503%3A+Add+metric+for+number+of+topics+marked+for+deletion)
When deleting a large number of topics, the Controller can get quite bogged down. One problem with this is the lack of visibility into the progress of the Controller. We can look into the ZK path for topics marked for deletion, but in a production environment this is inconvenient. This PR adds a JMX metric `kafka.controller:type=KafkaController,name=TopicsToDeleteCount` to make it easier to see how many topics are being deleted.
Reviewers: Stanislav Kozlovski <stanislav@confluent.io>, Jun Rao <junrao@gmail.com>, Jason Gustafson <jason@confluent.io>
Author: asutosh936 <asutosh.pandya@hotmail.com>
Reviewers: Mickael Maison <mickael.maison@gmail.com>, Manikumar Reddy <manikumar.reddy@gmail.com>, Vahid Hashemian <vahid.hashemian@gmail.com>, Jason Gustafson <jason@confluent.io>
Closes#7141 from asutosh936/KAFKA-8698
In a KTable context, we should not pass null into a user-supplied serde.
Testing: I verified that the change to the test results in test failures without the patch.
Reviewers: Matthias J. Sax <matthias@confluent.io>, Bill Bejeck <bill@confluent.io>, Guozhang Wang <wangguoz@gmail.com>,
This patch fixes a bug in the handling of MESSAGE_TOO_LARGE errors. The large batch is split, the smaller batches are re-added to the accumulator, and the batch is deallocated, but it was not removed from the list of in-flight batches. When the batch was eventually expired from the in-flight batches, the producer would try to deallocate it a second time, causing an error. This patch changes the behavior to correctly remove the batch from the list of in-flight requests.
Reviewers: Luke Stephenson, Ismael Juma <ismael@juma.me.uk>, Jason Gustafson <jason@confluent.io>
`testProduceAfterLogDirFailureOnLeader` currently disables producer retries in
order to catch and validate the exception thrown by a failure, and then tries to
produce successfully once the leadership changes. This second produce can
intermittently fail, causing test flakiness. This patch splits these validations
into two tests in order to allow retries for the produce request after the
leadership change.
Reviewers: Colin P. Mccabe <cmccabe@confluent.io>, Ismael Juma <ismael@juma.me.uk>
If retrieving metadata during `poll(Duration)` repeatedly takes longer than the
poll timeout, we don't make progress and `waitUntilTrue` eventually times out
causing the test to fail. This behaviour differs from the older `poll(long)` that
would block until metadata retrieval had completed. The flakiness was likely
introduced when we switched from the latter to the former.
Reviewers: Ismael Juma <ismael@juma.me.uk>
Reverts the TreeMap -> ConcurrentSkipListMap change that caused a performance regression in 2.3, and fixes the ConcurrentModificationException by copying (just) the key set to iterate over
Reviewers: Bill Bejeck <bill@confluent.io>, Guozhang Wang <wangguoz@gmail.com>
Refactors the DescribeDelegationToken to use the generated RPC classes.
Author: Viktor Somogyi-Vass <viktorsomogyi@gmail.com>
Author: Viktor Somogyi <viktorsomogyi@gmail.com>
Reviewers: Mickael Maison <mickael.maison@gmail.com>, Manikumar Reddy <manikumar.reddy@gmail.com>
Closes#7154 from viktorsomogyi/refactor-describe-dt
This patch updates ConsumerGroupCommandTest.scala to use the maximum possible number of AdminClient retries. The test runs will still be bounded by the request timeout. This address flakiness in tests such as testResetOffsetsNotExistingGroup and testResetOffsetsExistingTopic, which was caused by group coordinators being intermittently unavailable.
Reviewers: Ismael Juma <ismael@juma.me.uk>, Jason Gustafson <jason@confluent.io>
Update the javadoc of SockerServer#startup(). SocketServer#startProcessors() does not exist any more and it has been replaced by SocketServer#startDataPlaneProcessors() and SocketServer#startControlPlaneProcessor().
Reviewers: Jason Gustafson <jason@confluent.io>
Credit to @lbradstreet for profiling the producer with a large number of partitions.
Cache `topicMetadata`, `brokers` and `controller` in the `MetadataResponse`
the first time it's needed avoid unnecessary recomputation. We were previously
computing`brokersMap` 4 times per partition in one code path that was invoked from
multiple places. This is a regression introduced via a42f16f980 and first released
in 2.3.0.
The `Cluster` constructor became significantly more allocation heavy due to
2c44e77e2f, first released in 2.2.0. Replaced `merge` calls with more verbose,
but more efficient code. Added a test to verify that the returned collections are
unmodifiable.
Add `topicAuthorizedOperations` and `clusterAuthorizedOperations` to
`MetadataResponse` and remove `data()` method.
Reviewers: Manikumar Reddy <manikumar.reddy@gmail.com>, Lucas Bradstreet <lucasbradstreet@gmail.com>, Colin P. McCabe <cmccabe@confluent.io>, Stanislav Kozlovski <stanislav_kozlovski@outlook.com>, Justine Olshan <jolshan@confluent.io>
This patch stores `clusterId` in the `meta.properties` file. During startup, the broker checks that it joins the correct cluster and fails fast otherwise.
The `meta.properties' is versioned. I have decided to not bump the version because 1) the clusterId is null anyway if not present in the file; and 2) bumping it means that rolling back to a previous version won't work.
I have refactored the way the metadata is read and written as it was strongly coupled with the brokerId bits. Now, the metadata is read independently during the startup and used to 1) check the clusterId and 2) get or generate the brokerId (as before).
Reviewers: Stanislav Kozlovski <stanislav_kozlovski@outlook.com>, Jason Gustafson <jason@confluent.io>
Add the AlterPartitionReassignments and ListPartitionReassignments APIs. Also remove an unused methodlength suppression for KafkaAdminClient.
Reviewers: Colin P. McCabe <cmccabe@apache.org>, Viktor Somogyi <viktorsomogyi@gmail.com>
Kafka ships with default ZK configuration. With the upgrade to ZK 3.5, our defaults include running ZK's AdminServer on port 8080. This is an unfortunate default as it tends to cause conflicts.
I suggest we default to disable ZK's AdminServer in the default ZK configs that we ship. Users who want to use AdminServer can enable it and set the port to something that works for them. Realistically, in most production environments, a different ZK server will be used anyway. So this is mostly to save new users who are trying Kafka on their own machine from running into accidental and frustrating port conflicts.
Author: Gwen Shapira <gwen@confluent.io>
Reviewers: Ismael Juma
Closes#7203 from gwenshap/zk_disable_adminserver
The RecordAccumulator ready calls `leaderFor` unnecessarily when the ProducerBatch
queue is empty. When producing to many partitions, the queue is often empty and the
`leaderFor` call can be expensive in comparison. Remove the unnecessary call.
Reviewers: Ismael Juma <ismael@juma.me.uk>
Reviewers: A. Sophie Blee-Goldman <sophie@confluent.io>, Bill Bejeck <bill@confluent.io>, Richard Yu <yohan.richard.yu@gmail.com>, Guozhang Wang <guozhang@confluent.io>
Iterator#remove has a default implementation that throws UnsupportedOperatorException so there's no need to override it with the same thing.
Should be cherry-picked back to whenever we switched to Java 8
Reviewers: Bill Bejeck <bill@confluent.io>, Matthias J. Sax <matthias@confluent.io>, Guozhang Wang <guozhang@confluent.io>
InMemoryKeyValueStore uses ConcurrentSkipListMap#size which takes linear time as it iterates over the entire map. We should just track size ourselves for approximateNumEntries
Reviewers: Guozhang Wang <wangguoz@gmail.com>, Matthias J. Sax <mjsax@apache.org>
This adds a basic system test that enables rack-aware brokers with the rack-aware replica selector for fetch from followers (KIP-392). The test asserts that the follower was read from at least once and that all the messages that were produced were successfully consumed.
Reviewers: Jason Gustafson <jason@confluent.io>
Added the ability for the connector handles and task handles, which are used by the monitorable source and sink connectors used to verify the functionality of the Connect framework, to record the number of times the connector and tasks have each been started, and to allow a test to obtain a `RestartLatch` that can be used to block until the connectors and/or tasks have been restarted a specified number of types.
Typically, a test will get the `ConnectorHandle` for a connector, and call the `ConnectorHandle.expectedRestarts(int)` method with the expected number of times that the connector and/or tasks will be restarted, and will hold onto the resulting `RestartLatch`. The test will then change the connector (or otherwise cause the connector to restart) one or more times as desired, and then call `RestartLatch.await(long, TimeUnit)` to block the test up to a specified duration for the connector and all tasks to be started the specified number of times.
This commit also increases several of the maximum wait times used in other integration tests. It doesn’t hurt to potentially wait longer, since most test runs will not need to wait the maximum amount of time anyway. However, in the rare cases that do need that extra time, waiting a bit more is fine if we can reduce the flakiness and minimize test failures that happened to time out too early.
Unit tests were added for the new `RestartLatch` and `StopAndStartCounter` utility classes. This PR only affects the tests and does not affect any runtime code or API.
**This should be merged on `trunk` and backported to the `2.3.x` branch.**
Author: Randall Hauch <rhauch@gmail.com>
Reviewers: Konstantine Karantasis, Arjun Satish
Closes#7019 from rhauch/kafka-8391
When calling readLogToEnd(), the KafkaBasedLog worker thread should catch TimeoutException and log a warning, which can occur if brokers are unavailable, otherwise the worker thread terminates.
Includes an enhancement to MockConsumer that allows simulating exceptions not just when polling but also when querying for offsets, which is necessary for testing the fix.
Author: Paul Whalen <pgwhalen@gmail.com>
Reviewers: Randall Hauch <rhauch@gmail.com>, Arjun Satish <arjun@confluent.io>, Ryanne Dolan <ryannedolan@gmail.com>
Corrected the AbstractHerder to correctly identify task configs that contain variables for externalized secrets. The original method incorrectly used `matcher.matches()` instead of `matcher.find()`. The former method expects the entire string to match the regex, whereas the second one can find a pattern anywhere within the input string (which fits this use case more correctly).
Added unit tests to cover various cases of a config with externalized secrets, and updated system tests to cover case where config value contains additional characters besides secret that requires regex pattern to be found anywhere in the string (as opposed to complete match).
Author: Arjun Satish <arjun@confluent.io>
Reviewer: Randall Hauch <rhauch@gmail.com>
Lucas Bradstreet
Vikas Singh
Bruno Cadonna
Bill Bejeck
Anastasia Vela
Stanislav Kozlovski
A. Sophie Blee-Goldman
James Cheng
Robert Yokota
cpettitt-confluent
saisandeep
Konstantine Karantasis
Randall Hauch
Chris Egerton
Guozhang Wang
David Arthur
asutosh936
John Roesler
Bob Barrett
Andy Coates
Matthias J. Sax
Brian Bushree
Lee Dongjin
Cheng Pan
Justine Olshan
Viktor Somogyi-Vass
Jeff Huang
Mickael Maison
David Jacot
Ismael Juma
Stanislav Kozlovski
Gwen Shapira
Lucas Bradstreet
Paul
Arjun Satish