Implementation for lazy down-conversion in a chunked manner for efficient memory usage during down-conversion. This pull request is mainly to get initial feedback on the direction of the patch. The patch includes all the main components from KIP-283.
Reviewers: Jason Gustafson <jason@confluent.io>
This implements KIP-219, where a broker returns a response with throttle time on
quota violation immediately after processing the corresponding request. After
the response is sent out, the broker will keep the channel muted until the
throttle time is over. Also, on receiving a response with throttle time, client
will block outgoing communication to the broker for the specified throttle time.
See PR 4830, 5064 and 5094 for all the review history
Author: Jon Lee <jonlee@jonlee-ld1.linkedin.biz>
Reviewers: Jun Rao <junrao@gmail.com>, Rajini Sivaram <rajinisivaram@googlemail.com>, Ismael Juma <ismael@juma.me.uk>, Dong Lin <lindong28@gmail.com>
Closes#5064 from jonlee2/kip-219
KafkaController currently writes reassignment znode once for every partition that has been successfully reassigned. This is unnecessary and controller should be able to update reassignment znode once to remove all partitions that have been reassigned from the reassignment znode.
Author: Dong Lin <dolin@linkedin.com>
Reviewers: Ismael Juma <ismael@juma.me.uk>
Closes#4659 from lindong28/KAFKA-6617
Refresh metadata if broker connection fails so that new calls are sent only to nodes that are alive and requests to controller are sent to the new controller if controller changes due to broker failure. Also reassign calls that could not be sent.
Reviewers: Dong Lin <lindong28@gmail.com>, Jason Gustafson <jason@confluent.io>
Document cases where `IllegalStateException` is raised when attempting an invalid operation on an unassigned partition. Also change `position()` to raise `IllegalStateException` when called on an unassigned partition for consistency.
Add the new stricter-timeout version of `poll` proposed in KIP-266.
The pre-existing variant `poll(long timeout)` would block indefinitely for metadata
updates if they were needed, then it would issue a fetch and poll for `timeout` ms
for new records. The initial indefinite metadata block caused applications to become
stuck when the brokers became unavailable. The existence of the timeout parameter
made the indefinite block especially unintuitive.
This PR adds `poll(Duration timeout)` with the semantics:
1. iff a metadata update is needed:
1. send (asynchronous) metadata requests
2. poll for metadata responses (counts against timeout)
- if no response within timeout, **return an empty collection immediately**
2. if there is fetch data available, **return it immediately**
3. if there is no fetch request in flight, send fetch requests
4. poll for fetch responses (counts against timeout)
- if no response within timeout, **return an empty collection** (leaving async fetch request for the next poll)
- if we get a response, **return the response**
The old method, `poll(long timeout)` is deprecated, but we do not change its semantics, so it remains:
1. iff a metadata update is needed:
1. send (asynchronous) metadata requests
2. poll for metadata responses *indefinitely until we get it*
2. if there is fetch data available, **return it immediately**
3. if there is no fetch request in flight, send fetch requests
4. poll for fetch responses (counts against timeout)
- if no response within timeout, **return an empty collection** (leaving async fetch request for the next poll)
- if we get a response, **return the response**
One notable usage is prohibited by the new `poll`: previously, you could call `poll(0)` to block for metadata updates, for example to initialize the client, supposedly without fetching records. Note, though, that this behavior is not according to any contract, and there is no guarantee that `poll(0)` won't return records the first time it's called. Therefore, it has always been unsafe to ignore the response.
This KIP adds the following functionality related to SASL/OAUTHBEARER:
1) Allow clients (both brokers when SASL/OAUTHBEARER is the inter-broker protocol as well as non-broker clients) to flexibly retrieve an access token from an OAuth 2 authorization server based on the declaration of a custom login CallbackHandler implementation and have that access token transparently and automatically transmitted to a broker for authentication.
2) Allow brokers to flexibly validate provided access tokens when a client establishes a connection based on the declaration of a custom SASL Server CallbackHandler implementation.
3) Provide implementations of the above retrieval and validation features based on an unsecured JSON Web Token that function out-of-the-box with minimal configuration required (i.e. implementations of the two types of callback handlers mentioned above will be used by default with no need to explicitly declare them).
4) Allow clients (both brokers when SASL/OAUTHBEARER is the inter-broker protocol as well as non-broker clients) to transparently retrieve a new access token in the background before the existing access token expires in case the client has to open new connections.
consumer offset path in zookeeper should be /consumers/${group}/offsets/${topic}/${partition} instead of /consumers/${group}/offset/${topic}/${partition}. Added `s` to the word `offset`.
Reviewers: Ismael Juma <ismael@juma.me.uk>, Manikumar Reddy O <manikumar.reddy@gmail.com>, Jun Rao <junrao@gmail.com>
The Signal classes are not available in the compile classpath
if --release is used so we use reflection as a workaround.
As part of that moved the code to Java and added a simple
unit test.
Also disabled the signal handler if the IBM JDK is being used
due to KAFKA-6918.
Manually tested shutdown via ctrl+c and verified that
the message is printed.
* Removed Scala producers, request classes, kafka.tools.ProducerPerformance, encoders,
tests.
* Updated ConsoleProducer to remove Scala producer support (removed `BaseProducer`
and several options that are not used by the Java producer).
* Updated a few Scala consumer tests to use the new producer (including a minor
refactor of `produceMessages` methods in `TestUtils`).
* Updated `ClientUtils.fetchTopicMetadata` to use `SimpleConsumer` instead of
`SyncProducer`.
* Removed `TestKafkaAppender` as it looks useless and it defined an `Encoder`.
* Minor import clean-ups
No new tests added since behaviour should remain the same after these changes.
Author: Ismael Juma <ismael@juma.me.uk>
Reviewers: Manikumar Reddy O <manikumar.reddy@gmail.com>, Dong Lin <lindong28@gmail.com>
Closes#5045 from ijuma/kafka-6921-remove-old-producer
Fix the check, add unit test to verify the change, update `DynamicBrokerReconfigurationTest` to avoid dynamic keystore update in tests which are not expected to update keystores.
We no longer need them since we now require Java 8.
Author: Ismael Juma <ismael@juma.me.uk>
Reviewers: Andras Beni <andrasbeni@cloudera.com>, Manikumar Reddy O <manikumar.reddy@gmail.com>, Dong Lin <lindong28@gmail.com>
Closes#5049 from ijuma/remove-base64
* The consumer groups API should expose group state and coordinator information. This information is needed by administrative tools and scripts that access consume groups.
* The partition assignment will be empty when the group is rebalancing. Fix an issue where the adminclient attempted to deserialize this empty buffer.
* Remove nulls from the API and make all collections immutable.
* DescribeConsumerGroupsResult#all should return a result as expected, rather than Void
* Fix exception text for GroupIdNotFoundException, GroupNotEmptyException. It was being filled in as "The group id The group id does not exist was not found" and similar.
Reviewers: Attila Sasvari <asasvari@apache.org>, Andras Beni <andrasbeni@cloudera.com>, Dong Lin <lindong28@gmail.com>, Jason Gustafson <jason@confluent.io>
This patch adds a few metrics that are useful for monitoring controller health. See KIP-237 for more detail.
Author: Dong Lin <lindong28@gmail.com>
Reviewers: Jun Rao <junrao@gmail.com>
Closes#4392 from lindong28/KAFKA-3473
Removed usage of deprecated AdminClient from StreamsResetter
No additional tests are required.
Matthias J. Sax <matthias@confluent.io>, Guozhang Wang <wangguoz@gmail.com>
Also include a few clean-ups:
* Method/variable/parameter renames to make them consistent with
the class name
* Return `ApiVersion` from `minSupportedFor`
* Use `values` to remove some code duplication
* Reduce duplication in `ApiVersion` by introducing the `shortVersion`
method and building the versions map programatically
* Avoid unnecessary `regex` in `ApiVersion.apply`
* Added scaladoc to a few methods
Some of these were originally discussed in:
https://github.com/apache/kafka/pull/4583#pullrequestreview-98089400
Added a test for `ApiVersion.shortVersion`. Relying on existing tests
for the rest since there is no change in behaviour.
Reviewers: Jason Gustafson <jason@confluent.io>
Implementation of KIP-279 as described here: https://cwiki.apache.org/confluence/display/KAFKA/KIP-279%3A+Fix+log+divergence+between+leader+and+follower+after+fast+leader+fail+over
In summary:
- Added leader_epoch to OFFSET_FOR_LEADER_EPOCH_RESPONSE
- Leader replies with the pair( largest epoch less than or equal to the requested epoch, the end offset of this epoch)
- If Follower does not know about the leader epoch that leader replies with, it truncates to the end offset of largest leader epoch less than leader epoch that leader replied with, and sends another OffsetForLeaderEpoch request. That request contains the largest leader epoch less than leader epoch that leader replied with.
Reviewers: Dong Lin <lindong28@gmail.com>, Jun Rao <junrao@gmail.com>
Grow buffers in log cleaner to hold one message set after sanity check even if message set is bigger than max.message.bytes.
Reviewers: Jason Gustafson <jason@confluent.io>, Ismael Juma <ismael@juma.me.uk>, Jun Rao <junrao@gmail.com>
Fixes a deadlock between the controller's beforeInitializingSession callback which holds the zookeeper client initialization lock while awaiting completion of an asynchronous event which itself depends on the same lock.
Also catch and log callback exceptions to ensure the ZooKeeper reconnection takes place.
Finally, configure KafkaScheduler in ZooKeeperClient to have at least 1 thread.
Added tests that fail or hang without the changes in this PR.
Reviewers: Jun Rao <junrao@gmail.com>, Ismael Juma <ismael@juma.me.uk>
The leader must explicitly check if requested leader epoch is undefined, and return undefined offset so that the follower can fall back to truncating to high watermark. Otherwise, if the leader also is not tracking leader epochs, it may return its LEO, which will the follower to truncate to the incorrect offset.
Log cleaner grows buffers when result.messagesRead is zero. This contains the number of filtered messages read from source which can be zero when transactions are used because batches may be discarded. Log cleaner incorrectly assumes that messages were not read because the buffer was too small and attempts to double the buffer size unnecessarily, failing with an exception if the buffer is already max.message.bytes. Additional check for discarded batches has been added to avoid growing buffers when batches are discarded.
Reviewers: Ismael Juma <ismael@juma.me.uk>, Jason Gustafson <jason@confluent.io>
Enable dynamic update of default unclean leader election config of brokers. A new controller event has been added to process unclean leader election when the config is enabled dynamically.
Reviewers: Dong Lin <lindong28@gmail.com>, Manikumar Reddy <manikumar.reddy@gmail.com>
Prevent exception thrown by metric reporters to impact request processing and other reporters.
Co-authored-by: Mickael Maison <mickael.maison@gmail.com>
Co-authored-by: Edoardo Comar <ecomar@uk.ibm.com>
Reviewers: Rajini Sivaram <rajinisivaram@googlemail.com>
Added unit tests for ReplicaAlterLogDirsThread. Mostly focused on unit tests for truncating logic.
Fixed ReplicaAlterLogDirsThread.buildLeaderEpochRequest() to use future replica's latest epoch (not the latest epoch of replica it is fetching from). This follows the logic that offset for leader epoch request should be based on leader epoch of the follower (in this case it's the future local replica).
Also fixed PartitionFetchState constructor that takes offset and delay. The code ignored the delay parameter and used 0 for the delay. This constructor is used only by another constructor which passes delay = 0, which luckily works.
Author: Anna Povzner <anna@confluent.io>
Reviewers: Dong Lin <lindong28@gmail.com>
Closes#4918 from apovzner/kafka-6795
Currently if the client sends a produce request or a fetch request to a broker which isn't a replica,
we return UNKNOWN_TOPIC_OR_PARTITION. This is a bit surprising to see when the topic actually
exists. It would be better to return NOT_LEADER to avoid confusion. Clients typically handle both errors by refreshing metadata and retrying, so changing this should not cause any change in behavior on the client. This case can be hit following a partition reassignment after the leader is moved and the local replica is deleted.
To validate the current behavior and the fix, I've added integration tests for the fetch and produce APIs.
Start processing client connections only after completing KafkaServer initialization to ensure that credentials are loaded from ZK into cache before authentications are processed. Acceptors are started earlier so that bound port is known for registering in ZK.
Reviewers: Manikumar Reddy <manikumar.reddy@gmail.com>, Jun Rao <junrao@gmail.com>, Ismael Juma <ismael@juma.me.uk>
A partially deleted topic can end up with some partitions having no leadership info.
For the partially deleted topic, a new controller should be able to finish the topic deletion
by transitioning the rogue partition's replicas to OfflineReplica state.
This patch adds logic to transition replicas to OfflineReplica state whose partitions have
no leadership info.
Added a new test method to cover the partially deleted topic case.
Reviewers: Jun Rao <junrao@gmail.com>
Dhruvil Shah
Jon Lee
Dong Lin
Rajini Sivaram
Manikumar Reddy O
John Roesler
Chia-Ping Tsai
Ron Dagostino
maytals
yaphet
Jun Rao
Ismael Juma
Sasaki Toru
Jason Gustafson
Colin Patrick McCabe
Dong Lin
fedosov-alexander
Anna Povzner
Fedor Bobin
Guozhang Wang
Mickael Maison
Patrik Erdes
gitlw
Allen Wang