When a Connect distributed worker starts up talking with broker versions 0.10.1.0 and later, it will use the AdminClient to look for the internal topics and attempt to create them if they are missing. Although the AdminClient was added in 0.11.0.0, the AdminClient uses APIs to create topics that existed in 0.10.1.0 and later. This feature works as expected when Connect uses a broker version 0.10.1.0 or later.
However, when a Connect distributed worker starts up using a broker older than 0.10.1.0, the AdminClient is not able to find the required APIs and thus will throw an UnsupportedVersionException. Unfortunately, this exception is not caught and instead causes the Connect worker to fail even when the topics already exist.
This change handles the UnsupportedVersionException by logging a debug message and doing nothing. The existing producer logic will get information about the topics, which will cause the broker to create them if they don’t exist and broker auto-creation of topics is enabled. This is the same behavior that existed prior to 0.11.0.0, and so this change restores that behavior for brokers older than 0.10.1.0.
This change also adds a system test that verifies Connect works with a variety of brokers and is able to run source and sink connectors. The test verifies that Connect can read from the internal topics when the connectors are restarted.
Author: Randall Hauch <rhauch@gmail.com>
Reviewers: Ewen Cheslack-Postava <ewen@confluent.io>
Closes#3641 from rhauch/kafka-5704
ewencp would be great to cherry-pick this back into 0.11.x if possible
Author: Xavier Léauté <xavier@confluent.io>
Reviewers: Ewen Cheslack-Postava <ewen@confluent.io>
Closes#3645 from xvrl/system-test-cluster-id
Added handling of _DUCKTAPE_OPTIONS (mainly for enabling debugging)
Author: Paolo Patierno <ppatierno@live.com>
Reviewers: Ewen Cheslack-Postava <ewen@confluent.io>
Closes#3578 from ppatierno/kafka-5643
Support a --custom-ducktape flag which allows developers to install
their own versions of ducktape into Docker images. This is helpful for
ducktape development.
Author: Colin P. Mccabe <cmccabe@confluent.io>
Reviewers: Ewen Cheslack-Postava <me@ewencp.org>, Ismael Juma <ismael@juma.me.uk>
Closes#3539 from cmccabe/KAFKA-5602
The bump from 0.11.1.0-SNAPSHOT to 1.0.0-SNAPSHOT broke a couple
of system tests:
* TestVerifiableProducer.test_simple_run
* KafkaVersionTest.test_multi_version
Author: Ismael Juma <ismael@juma.me.uk>
Reviewers: Damian Guy <damian.guy@gmail.com>
Closes#3587 from ijuma/fix-_kafka_jar_versions
Author: Dong Lin <lindong28@gmail.com>
Reviewers: Jiangjie Qin <becket.qin@gmail.com>, Jun Rao <junrao@gmail.com>, Ismael Juma <ismael@juma.me.uk>, Onur Karaman <okaraman@linkedin.com>
Closes#2929 from lindong28/KAFKA-4763
Waiting for the first line of output was added in KAFKA-2527 when JmxMixin was originally added as a heuristic to
determine when the process was ready. We've since determined this is not good enough given JmxTool's limitations
and now include a separate, more reliable check before starting JmxTool. This check is also dangerous since a
consumer that is started before data is available in the topic, it won't output anything to stdout and only logs
errors to a separate log file. This means we may have a long delay between starting the process and starting JMX
monitoring.
Since we have a more reliable check for liveness via JMX now (and in cases that need it, partition assignment
metrics via JMX), we should no longer need to wait for the first line of output.
Author: Ewen Cheslack-Postava <me@ewencp.org>
Reviewers: Ismael Juma <ismael@juma.me.uk>, Apurva Mehta <apurva@confluent.io>
Closes#3447 from ewencp/dont-wait-first-line-console-consumer
Author: Matthias J. Sax <matthias@confluent.io>
Reviewers: Eno Thereska <eno.thereska@gmail.com>, Damian Guy <damian.guy@gmail.com>
Closes#3497 from mjsax/disable-flaky-system-tests
Update system tests to make use of the newly released 0.11 version.
Add on to https://github.com/apache/kafka/pull/3454
Author: Eno Thereska <eno.thereska@gmail.com>
Reviewers: Damian Guy <damian.guy@gmail.com>
Closes#3479 from enothereska/minor-compatibility-tests
Author: Ismael Juma <ismael@juma.me.uk>
Reviewers: Eno Thereska <eno.thereska@gmail.com>, Ewen Cheslack-Postava <me@ewencp.org>
Closes#3454 from ijuma/test-upgrades-from-0.11.0.x
Author: Matthias J. Sax <matthias@confluent.io>
Reviewers: Guozhang Wang <wangguoz@gmail.com>
Closes#3310 from mjsax/kafka-5362-add-eos-system-tests-for-streams-api
-Tests for rolling upgrades for a streams app (keeping broker config fixed)
-Tests for rolling upgrades of brokers (keeping streams app config fixed)
Author: Eno Thereska <eno.thereska@gmail.com>
Reviewers: Matthias J. Sax <matthias@confluent.io>, Damian Guy <damian.guy@gmail.com>
Closes#3411 from enothereska/KAFKA-5487-upgrade-test-streams
Author: Matthias J. Sax <matthias@confluent.io>
Reviewers: Joseph Rea <jrea@users.noreply.github.com>, Ewen Cheslack-Postava <ewen@confluent.io>
Closes#2575 from mjsax/minor-update-system-test-readme
Increased the timeout from 30sec to 60sec. When running the system tests with packaged Kafka, Connect workers can take about 30seconds to start.
Author: Randall Hauch <rhauch@gmail.com>
Reviewers: Ewen Cheslack-Postava <ewen@confluent.io>
Closes#3344 from rhauch/KAFKA-5450
Author: Jason Gustafson <jason@confluent.io>
Reviewers: Apurva Mehta <apurva@confluent.io>, Ismael Juma <ismael@juma.me.uk>
Closes#3340 from hachikuji/add-random-aborts-to-system-test
When the message copier hangs (like when there is a bug in the client), it ignores the sigterm and doesn't shut down. this leaves the cluster in an unclean state causing future tests to fail.
In this patch we always send SIGKILL when cleaning the node if the process isn't already dead. This is consistent with the other services.
Author: Apurva Mehta <apurva@confluent.io>
Reviewers: Jason Gustafson <jason@confluent.io>
Closes#3308 from apurvam/KAFKA-5437-force-kill-message-copier-on-cleanup
Author: Matthias J. Sax <matthias@confluent.io>
Reviewers: Damian Guy <damian.guy@gmail.com>, Bill Bejeck <bill@confluent.io>, Guozhang Wang <wangguoz@gmail.com>
Closes#3201 from mjsax/kafka-5362-add-eos-system-tests-for-streams-api
We need this to debug most issues with the transactions system test.
Author: Apurva Mehta <apurva@confluent.io>
Reviewers: Jason Gustafson <jason@confluent.io>
Closes#3261 from apurvam/MINOR-set-log-level-for-producer-to-trace-for-transactions-test
The log_level parameter is used in system tests in kafka.py. However the log4j template accepted that parameter in only one place. This led to a large number of DEBUG lines printed even when the intention was to capture only INFO lines. Led to huge log files. Thanks to ijuma for noticing this.
Author: Eno Thereska <eno.thereska@gmail.com>
Reviewers: Ismael Juma <ismael@juma.me.uk>
Closes#3247 from enothereska/minor-log4j-template-fix
This currently fails in multiple ways. One of which is most likely KAFKA-5355, where the concurrent consumer reads duplicates.
During broker bounces, the concurrent consumer misses messages completely. This is another bug.
Author: Apurva Mehta <apurva@confluent.io>
Reviewers: Jason Gustafson <jason@confluent.io>
Closes#3217 from apurvam/KAFKA-5366-add-concurrent-reads-to-transactions-system-test
Also update message format tests now that we have a third message
format.
Finally, set group.initial.rebalance.delay.ms=100.
Author: Ismael Juma <ismael@juma.me.uk>
Reviewers: Ewen Cheslack-Postava <ewen@confluent.io>, Jason Gustafson <jason@confluent.io>
Closes#2701 from ijuma/update-upgrade-tests-for-0.11
It avoids the need to handle protocol downgrades and it's safe (i.e. it will never cause
the auto creation of topics).
Author: Colin P. Mccabe <cmccabe@confluent.io>
Reviewers: Ismael Juma <ismael@juma.me.uk>
Closes#3220 from ijuma/kafka-5374-admin-client-metadata
- add broker compatibility system tests
Author: Matthias J. Sax <matthias@confluent.io>
Reviewers: Damian Guy, Eno Thereska, Guozhang Wang
Closes#2974 from mjsax/kafka-4923-add-eos-to-streams-add-broker-check-and-system-test
## Fixes race condition in TestVerifiableProducer sanity test:
The test starts a producer, waits for at least 5 acks, and then
logs in to the worker to grep for the producer process to figure
out what version it is running.
The problem was that the producer was set up to produce 1000 messages
at a rate of 1000 msgs/s and then exit. This means it will have a
typical runtime slightly above 1 second.
Logging in to the vagrant instance might take longer than that thus
resulting in the process grep to fail, failing the test.
This commit doesn't really fix the issue - a proper fix would be to tell
the producer to stick around until explicitly killed - but it increases
the chances of the test passing, at the expense of a slightly longer
runtime.
## Improves error reporting when is_version() fails
Author: Magnus Edenhill <magnus@edenhill.se>
Reviewers: Apurva Mehta <apurva@confluent.io>, Ewen Cheslack-Postava <ewen@confluent.io>
Closes#2765 from edenhill/trunk
The backing store for offsets, status, and configs now attempts to use the new AdminClient to look up the internal topics and create them if they don’t yet exist. If the necessary APIs are not available in the connected broker, the stores fall back to the old behavior of relying upon auto-created topics. Kafka Connect requires a minimum of Apache Kafka 0.10.0.1-cp1, and the AdminClient can work with all versions since 0.10.0.0.
All three of Connect’s internal topics are created as compacted topics, and new distributed worker configuration properties control the replication factor for all three topics and the number of partitions for the offsets and status topics; the config topic requires a single partition and does not allow it to be set via configuration. All of these new configuration properties have sensible defaults, meaning users can upgrade without having to change any of the existing configurations. In most situations, existing Connect deployments will have already created the storage topics prior to upgrading.
The replication factor defaults to 3, so anyone running Kafka clusters with fewer nodes than 3 will receive an error unless they explicitly set the replication factor for the three internal topics. This is actually desired behavior, since it signals the users that they should be aware they are not using sufficient replication for production use.
The integration tests use a cluster with a single broker, so they were changed to explicitly specify a replication factor of 1 and a single partition.
The `KafkaAdminClientTest` was refactored to extract a utility for setting up a `KafkaAdminClient` with a `MockClient` for unit tests.
Author: Randall Hauch <rhauch@gmail.com>
Reviewers: Ewen Cheslack-Postava <ewen@confluent.io>
Closes#2984 from rhauch/kafka-4667
Author: Jason Gustafson <jason@confluent.io>
Reviewers: Apurva Mehta <apurva@confluent.io>, Ismael Juma <ismael@juma.me.uk>
Closes#2745 from hachikuji/add-replication-testcase-for-compression
Author: Matthias J. Sax <matthias@confluent.io>
Reviewers: Eno Thereska <eno@confluent.io>, Ismael Juma <ismael@juma.me.uk>
Closes#2897 from mjsax/KAFKA-4564-follow-up
Author: Matthias J. Sax <matthias@confluent.io>
Reviewers: Ismael Juma, Eno Thereska, Matthias J. Sax, Guozhang Wang
Closes#2837 from mjsax/kafka-4564-fail-fast-test-stream-compatibility
Author: Matthias J. Sax <matthias@confluent.io>
Reviewers: Magnus Edenhill, Eno Thereska, Damian Guy, Guozhang Wang
Closes#2836 from mjsax/minor-broker-comp-test
This PR replaces https://github.com/apache/kafka/pull/2743 (just raising from Confluent repo)
This PR describes the addition of Partition Level Leader Epochs to messages in Kafka as a mechanism for fixing some known issues in the replication protocol. Full details can be found here:
[KIP-101 Reference](https://cwiki.apache.org/confluence/display/KAFKA/KIP-101+-+Alter+Replication+Protocol+to+use+Leader+Epoch+rather+than+High+Watermark+for+Truncation)
*The key elements are*:
- Epochs are stamped on messages as they enter the leader.
- Epochs are tracked in both leader and follower in a new checkpoint file.
- A new API allows followers to retrieve the leader's latest offset for a particular epoch.
- The logic for truncating the log, when a replica becomes a follower, has been moved from Partition into the ReplicaFetcherThread
- When partitions are added to the ReplicaFetcherThread they are added in an initialising state. Initialising partitions request leader epochs and then truncate their logs appropriately.
This test provides a good overview of the workflow `EpochDrivenReplicationProtocolAcceptanceTest.shouldFollowLeaderEpochBasicWorkflow()`
The corrupted log use case is covered by the test
`EpochDrivenReplicationProtocolAcceptanceTest.offsetsShouldNotGoBackwards()`
Remaining work: There is a do list here: https://docs.google.com/document/d/1edmMo70MfHEZH9x38OQfTWsHr7UGTvg-NOxeFhOeRew/edit?usp=sharing
Author: Ben Stopford <benstopford@gmail.com>
Author: Jun Rao <junrao@gmail.com>
Reviewers: Ismael Juma <ismael@juma.me.uk>, Jun Rao <junrao@gmail.com>
Closes#2808 from benstopford/kip-101-v2
Several fixes for handling broker failures:
- default replication value for internal topics is now 3 in test itself (not in streams code, that will require a KIP.
- streams producer waits for acks from all replicas in test itself (not in streams code, that will require a KIP.
- backoff time for streams client to try again after a failure to contact controller.
- fix bug related to state store locks (this helps in multi-threaded scenarios)
- fix related to catching exceptions property for network errors.
- system test for all the above
Author: Eno Thereska <eno@confluent.io>
Author: Eno Thereska <eno.thereska@gmail.com>
Reviewers: Matthias J. Sax <matthias@confluent.io>, Damian Guy <damian.guy@gmail.com>, Guozhang Wang <wangguoz@gmail.com>, Dan Norwood <norwood@confluent.io>, Ismael Juma <ismael@juma.me.uk>, Ewen Cheslack-Postava <ewen@confluent.io>
Closes#2719 from enothereska/KAFKA-4916-broker-bounce-test
This is from the KIP-98 proposal.
The main points of discussion surround the correctness logic, particularly the Log class where incoming entries are validated and duplicates are dropped, and also the producer error handling to ensure that the semantics are sound from the users point of view.
There is some subtlety in the idempotent producer semantics. This patch only guarantees idempotent production upto the point where an error has to be returned to the user. Once we hit a such a non-recoverable error, we can no longer guarantee message ordering nor idempotence without additional logic at the application level.
In particular, if an application wants guaranteed message order without duplicates, then it needs to do the following in the error callback:
1. Close the producer so that no queued batches are sent. This is important for guaranteeing ordering.
2. Read the tail of the log to inspect the last message committed. This is important for avoiding duplicates.
Author: Apurva Mehta <apurva@confluent.io>
Author: hachikuji <jason@confluent.io>
Author: Apurva Mehta <apurva.1618@gmail.com>
Author: Guozhang Wang <wangguoz@gmail.com>
Author: fpj <fpj@apache.org>
Author: Jason Gustafson <jason@confluent.io>
Reviewers: Jason Gustafson <jason@confluent.io>, Ismael Juma <ismael@juma.me.uk>, Jun Rao <junrao@gmail.com>
Closes#2735 from apurvam/exactly-once-idempotent-producer
Colin P. Mccabe
Eno Thereska
Xavier Léauté
Randall Hauch
Paolo Patierno
Ismael Juma
Dong Lin
Ewen Cheslack-Postava
Matthias J. Sax
Jason Gustafson
Apurva Mehta
Magnus Edenhill
Konstantine Karantasis
Ben Stopford
Eno Thereska