When the number of partitions is not divisible by the number of members,
some members will end up with one more partition than others.
Previously, we required these to be the members at the start of the
iteration order, which meant that partitions could be reassigned even
when the previous assignment was already balanced.
Allow any member to have the extra partition, so that we do not move
partitions around when the previous assignment is already balanced.
Before the PR
```
Benchmark (assignmentType) (assignorType) (isRackAware) (memberCount) (partitionsToMemberRatio) (subscriptionType) (topicCount) Mode Cnt Score Error Units
ServerSideAssignorBenchmark.doAssignment FULL RANGE false 10000 50 HOMOGENEOUS 1000 avgt 2 26.175 ms/op
ServerSideAssignorBenchmark.doAssignment FULL RANGE false 10000 50 HETEROGENEOUS 1000 avgt 2 123.955 ms/op
ServerSideAssignorBenchmark.doAssignment INCREMENTAL RANGE false 10000 50 HOMOGENEOUS 1000 avgt 2 24.408 ms/op
ServerSideAssignorBenchmark.doAssignment INCREMENTAL RANGE false 10000 50 HETEROGENEOUS 1000 avgt 2 114.873 ms/op
```
After the PR
```
Benchmark (assignmentType) (assignorType) (isRackAware) (memberCount) (partitionsToMemberRatio) (subscriptionType) (topicCount) Mode Cnt Score Error Units
ServerSideAssignorBenchmark.doAssignment FULL RANGE false 10000 50 HOMOGENEOUS 1000 avgt 2 24.259 ms/op
ServerSideAssignorBenchmark.doAssignment FULL RANGE false 10000 50 HETEROGENEOUS 1000 avgt 2 118.513 ms/op
ServerSideAssignorBenchmark.doAssignment INCREMENTAL RANGE false 10000 50 HOMOGENEOUS 1000 avgt 2 24.636 ms/op
ServerSideAssignorBenchmark.doAssignment INCREMENTAL RANGE false 10000 50 HETEROGENEOUS 1000 avgt 2 115.503 ms/op
```
Reviewers: David Jacot <djacot@confluent.io>
This PR creates a listener for `SharePartitionManager` to listen to any
changes in `ShareVersion` feature. In case, there is a toggle, we need
to change the attributes in `SharePartitionManager` accordingly.
Reviewers: Andrew Schofield <aschofield@confluent.io>
Improve docs for retries config, mainly to clarify the expected
behaviour on retries=0 Also remove unused funcs and fix typo.
Reviewers: Matthias J. Sax <matthias@confluent.io>, Ming-Yen Chung
<mingyen066@gmail.com>, PoAn Yang <payang@apache.org>
* Currently even if a user topic is deleted, its related records are not
deleted with respect to subscribed share groups from the GC and the SC.
* The event of topic delete is propagated from the
BrokerMetadataPublisher to the coordinators via the
`onPartitionsDeleted` method. This PR leverages this method to issue
cleanup calls to the GC and SC respectively.
* To prevent chaining of futures in the GC, we issue async calls to both
GC and SC independently and the methods take care of the respective
cleanups unaware of the other.
* This method is more efficient and transcends issues related to
timeouts/broker restarts resulting in chained future execution issues.
Reviewers: Andrew Schofield <aschofield@confluent.io>
This patch extends the OffsetFetch API to support topic ids. From
version 10 of the API, topic ids must be used.
The patch only contains the server side changes and it keeps the version
10 as unstable for now. We will mark the version as stable when the
client side changes are merged in.
Reviewers: TengYao Chi <frankvicky@apache.org>, Lianet Magrans <lmagrans@confluent.io>
Updated default locks config for share groups from 200 to 2000. The
increase in the limit is a result from tests which showed that with
default maxFetchRecords of 500 from client and 200 as locks limit, there
can't be parallel fetch for same partition. Also the tests resulted that
sharing a partition to an index of 3-4 is easily achievable, hence
raised the limit to 4 times of default limit of maxFetchRecords (500).
Reviewers: Andrew Schofield <aschofield@confluent.io>
When a group has pending transactional offsets but no committed offsets,
we can accidentally delete it while cleaning up expired offsets. Add a
check to avoid this case.
Reviewers: David Jacot <djacot@confluent.io>
As mentioned in
https://github.com/apache/kafka/pull/19378#pullrequestreview-2775598123,
the error messages for a 4.1 share consumer could be clearer for the
different cases for when it cannot successfully join a share group.
This PR uses different error messages for the different cases such as
out-of-date cluster or share groups just not enabled.
Reviewers: Apoorv Mittal <apoorvmittal10@gmail.com>
* In ConsoleShareConsumerTest, add `@SuppressWarnings("unchecked")`
annotation in method shouldUpgradeDeliveryCount
* In ListConsumerGroupOffsetsHandlerTest, add generic parameters to
HashSet constructors
* In TopicsImageTest, add explicit generic type to Collections.EMPTY_MAP
to fix raw type usage
Reviewers: Ken Huang <s7133700@gmail.com>, TengYao Chi
<kitingiao@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
At the retry limit binaryExponentialElectionBackoffMs it becomes
statistically likely that the exponential backoff returned
electionBackoffMaxMs. This is an issue as multiple replicas can get
stuck starting elections at the same cadence.
This change fixes that by added a random jitter to the max election
backoff.
Reviewers: José Armando García Sancio <jsancio@apache.org>, TaiJuWu
<tjwu1217@gmail.com>, Yung <yungyung7654321@gmail.com>
As a result of KAFKA-18905 the reassign test will often have test
failures which are unrelated to the actual reassignment of partitions.
This failure is mentioned in KAFKA-9199.
Quote from KAFKA-9199: "This issue popped up in the reassignment system
test. It ultimately caused the test to fail because the producer was
stuck retrying the duplicate batch repeatedly until ultimately giving
up."
Disabling the idempotent producer circumvents this issue and allows the
reassignment system tests to succeed reliably. The reassignment test
still check that produce batches were not lost.
Reviewers: José Armando García Sancio <jsancio@apache.org>
As part of KIP-1043, `Admin.listConsumerGroups()` and variants have been
deprecated. This is because there are now 4 types of group and listing
has been consolidated under `Admin.listGroups()`. This PR adds the
deprecation information to the upgrade documentation.
Reviewers: Lianet Magrans <lmagrans@confluent.io>
### Motivation
While investigating “events skipped in group
rebalancing” ([spring‑projects/spring‑kafka#3703](https://github.com/spring-projects/spring-kafka/issues/3703))
I discovered a race
condition between
- the main poll/commit thread, and
- the consumer‑coordinator heartbeat thread.
If the main thread enters
`ConsumerCoordinator.sendOffsetCommitRequest()` while the heartbeat
thread is finishing a rebalance (`SyncGroupResponseHandler.handle()`),
the group state transitions in the following order:
```
COMPLETING_REBALANCE → (race window) → STABLE
```
Because we read the state twice without a lock:
1. `generationIfStable()` returns `null` (state still
`COMPLETING_REBALANCE`),
2. the heartbeat thread flips the state to `STABLE`,
3. the main thread re‑checks with `rebalanceInProgress()` and wrongly
decides that a rebalance is still active,
4. a spurious `CommitFailedException` is returned even though the commit
could succeed.
For more details, please refer to sequence diagram below. <img
width="1494" alt="image"
src="https://github.com/user-attachments/assets/90f19af5-5e2d-4566-aece-ef764df2d89c"
/>
### Impact
- The exception is semantically wrong: the consumer is in a stable
group, but reports failure.
- Frameworks and applications that rely on the semantics of
`CommitFailedException` and `RetryableCommitException` (for example
`Spring Kafka`) take the wrong code path, which can ultimately skip the
events and break “at‑most‑once” guarantees.
### Fix
We enlarge the synchronized block in
`ConsumerCoordinator.sendOffsetCommitRequest()` so that the consumer
group state is examined atomically with respect to the heartbeat thread:
### Jira
https://issues.apache.org/jira/browse/KAFKA-19242
https: //github.com/spring-projects/spring-kafka/issues/3703
Signed-off-by: chickenchickenlove <ojt90902@naver.com>
Reviewers: David Jacot <david.jacot@gmail.com>
When fetching stable offsets in the group coordinator, we iterate over
all requested partitions. For each partition, we iterate over the
group's ongoing transactions to check if there is a pending
transactional offset commit for that partition.
This can get slow when there are a large number of partitions and a
large number of pending transactions. Instead, maintain a list of
pending transactions per partition to speed up lookups.
Reviewers: Shaan, Dongnuo Lyu <dlyu@confluent.io>, Chia-Ping Tsai <chia7712@gmail.com>, David Jaco <djacot@confluent.io>
Currently, the quorum uses kraft by default, so there's no need to
specify it explicitly.
For kip932 and isShareGroupTest, they are no longer used after #19542 .
Reviewers: PoAn Yang <payang@apache.org>, Ken Huang
<s7133700@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
The fallback mechanism for `remote.log.manager.copier.thread.pool.size`
and `remote.log.manager.expiration.thread.pool.size` defaulting to
`remote.log.manager.thread.pool.size` was introduced in KIP-950. This
approach was abandoned in KIP-1030, where default values were changed
from -1 to 10, and a configuration validator enforcing a minimum value
of 1 was added. As a result, this commit removes the fallback mechanism
from `RemoteLogManagerConfig.java` to align with the new defaults and
validation.
Reviewers: Chia-Ping Tsai <chia7712@gmail.com>
Migrates the `TxnTransitMetadata` class from scala to java, moving it
from to the `transaction-coordinator` module.
Reviewers: PoAn Yang <payang@apache.org>, Nick Guo
<lansg0504@gmail.com>, Ken Huang <s7133700@gmail.com>, Jhen-Yung Hsu
<jhenyunghsu@gmail.com>, TaiJuWu <tjwu1217@gmail.com>, Chia-Ping Tsai
<chia7712@gmail.com>
jira: https://issues.apache.org/jira/browse/KAFKA-19068
`RecordsIterator#decodeControlRecord` do the type check and then
`ControlRecord` constructor does that again.
we should add a static method to ControlRecord to create `ControlRecord`
with type check, and then `ControlRecord` constructor should be changed
to private to ensure all instance is created by the static method.
Reviewers: PoAn Yang <payang@apache.org>, Chia-Ping Tsai
<chia7712@gmail.com>
If directory id is `Uuid.ZERO_UUID`, the command don't print the result.
Reviewers: Ken Huang <s7133700@gmail.com>, TengYao Chi
<frankvicky@apache.org>, Chia-Ping Tsai <chia7712@gmail.com>
Add tests to check whether the results returned by the API
`createTopics` and `describeConfigs` contain internal configurations.
Reviewers: PoAn Yang <payang@apache.org>, Ken Huang
<s7133700@gmail.com>, TengYao Chi <frankvicky@apache.org>, TaiJuWu
<tjwu1217@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
Retaining a reference to ```AbstractConfig``` introduced coupling and
potential inconsistencies with dynamic config updates. This change
simplifies ```TransactionLogConfig``` into a POJO by removing the
internal ```AbstractConfig``` field, and aligns with feedback from
#19439
Reviewers: PoAn Yang <payang@apache.org>, TengYao Chi
<frankvicky@apache.org>, Chia-Ping Tsai <chia7712@gmail.com>
[As discussed in the mailing
list](https://lists.apache.org/thread/m03mpkm93737kk6d1nd6fbv9wdgsrhv9),
the broker only fetches remote data for ONE partition in a given
FetchRequest. In other words, if a consumer sends a FetchRequest
requesting 50 topic-partitions, and each partition's requested offset is
not stored locally - the broker will fetch and respond with just one
partition's worth of data from the remote store, and the rest will be
empty.
Given our defaults for total fetch response is 50 MiB and per partition
is 1 MiB, this can limit throughput. This patch documents the behavior
in 3 configs - `fetch.max.bytes`, `max.partition.fetch.bytes` and
`remote.fetch.max.wait.ms`
Reviewers: Luke Chen <showuon@gmail.com>, Kamal Chandraprakash
<kamal.chandraprakash@gmail.com>, Satish Duggana <satishd@apache.org>
@xijiu's invitation is timeout, so we have to remove the name, then
re-add it in a new commit.
see https://issues.apache.org/jira/browse/INFRA-26796
Reviewers: PoAn Yang <payang@apache.org>, xijiu <422766572@qq.com>,
Jhen-Yung Hsu <jhenyunghsu@gmail.com>, Ken Huang <s7133700@gmail.com>,
TengYao Chi <frankvicky@apache.org>
- A couple of newly added tests were found to be flaky in
`AuthorizerIntegrationTest.scala`.
- `testShareGroupDescribeWithGroupDescribeAndTopicDescribeAcl` and
`testShareGroupDescribeWithoutGroupDescribeAcl`. These tests pass
locally, so could not replicate the failure.
- But logs from develocity indicated that the test fails when the
following condition happens :
When the background error event arrives after the consumer had
unsubscribed, then these events are processed in the
`handleCompletedAcknowledgements` method and the exception from the
event is thrown, preventing `close()` to complete.
- We need to handle this race condition where we might get the
background event after unsubscribe and before processing the callbacks.
- PR fixes this by ignoring the exceptions in the background queue when
the `handleCompletedAcknowledgements` method is called during `close()`.
This ensures `close()` completes successfully.
- Have added a unit test which mimics the race condition as well.
Reviewers: Andrew Schofield <aschofield@confluent.io>
Admin.listConsumerGroups() was able to use the early versions of
ListGroups RPC with the version used dependent upon the filters the user
specified. Admin.listGroups(ListGroupsOptions.forConsumerGroups())
inadvertently required ListGroups v5 because it always set a types
filter. This patch handles the UnsupportedVersionException and winds
back the complexity of the request unless the user has specified filters
which demand a higher version.
It also adds ListGroupsOptions.forShareGroups() and forStreamsGroups().
The usability of Admin.listGroups() is much improved as a result.
Reviewers: Chia-Ping Tsai <chia7712@gmail.com>, PoAn Yang
<payang@apache.org>
Update docs to note index.interval.bytes sets entry frequency for offset index and, conditionally, time index. Improve clarity and readability of index.interval.bytes description.
Reviewers: Luke Chen <showuon@gmail.com>
When performing perf tests, we can specify a payload using the
"--payloadFile" flag. This file is utilized during the load/performance
testing process. This causes the entire file to get loaded into a String
and split using the delimiter. However, if the file is large, it may
result in NegativeArraySizeException error.
Moving the file loading logic to Scanner which doesn't have this issue.
Reviewers: José Armando García Sancio <jsancio@apache.org>, Ken Huang
<s7133700@gmail.com>, Zhe Guang <zheguang.zhao@alumni.brown.edu>
@xijiu's invitation is timeout, so we have to remove the name, then
re-add it in a new commit.
see https://issues.apache.org/jira/browse/INFRA-26796
Reviewers: Chih-Yuan Chien <joshua2519@gmail.com>, Kuan-Po Tseng
<brandboat@gmail.com>, PoAn Yang <payang@apache.org>, yunchi
<yunchipang@gmail.com>, TengYao Chi <kitingiao@gmail.com>, Hong-Yi Chen
<apalan60@gmail.com>, Bolin Lin <linbolin1230@gmail.com>, Shih-Yuan Lin
<shmily7829@gmail.com>, Mirai1129 <minecraftmiku831@gmail.com>
This PR moves SchedulerTest to server module and rewrite it with java.
Please also check updated import control config!
Reviewers: Ken Huang <s7133700@gmail.com>, Chia-Ping Tsai
<chia7712@gmail.com>
While working on https://github.com/apache/kafka/pull/19515, I came to
the conclusion that the OffsetFetchResponse is quite messy and overall
too complicated. This patch rationalize the constructors.
OffsetFetchResponse has a single constructor accepting the
OffsetFetchResponseData. A builder is introduced to handle the down
conversion. This will also simplify adding the topic ids. All the
changes are mechanical, replacing data structures by others.
Reviewers: Lianet Magrans <lmagrans@confluent.io>
The PR adds `slice` API in `Records.java` and further implementation in
`MemoryRecords`. With the addition of ShareFetch and it's support to
read from TieredStorage, where ShareFetch might acquire subset of fetch
batches and TieredStorage emits MemoryRecords, hence a slice API is
needed for MemoryRecords as well to limit the bytes transferred (if
subset batches are acquired).
MemoryRecords are sliced using `duplicate` and `slice` API of
ByteBuffer, which are backed by the original buffer itself hence no-copy
is created rather position, limit and offset are changed as per the new
position and length.
Reviewers: Andrew Schofield <aschofield@confluent.io>, Jun Rao
<junrao@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
The current ElectionWasClean checks if the new leader is in the previous
ISR. However, there is a corner case in the partition reassignment. The
partition reassignment can change the partition replicas. If the new
preferred leader (the first one in the new replicas) is the last one to
join ISR, this preferred leader will be elected in the same partition
change.
For example: In the previous state, the partition is Leader: 0,
Replicas (2,1,0), ISR (1,0), Adding(2), removing(0). Then replica 2
joins the ISR. The new partition would be like: Leader: 2, Replicas
(2,1), ISR(1,2). The new leader 2 is not in the previous ISR (1,0) but
it is still a clean election.
Reviewers: Jun Rao <junrao@gmail.com>
- Add new DS to wrap the member assignment (containing topic Ids, names
and partitions), to easily access the data as needed. This will be used
in following PR to integrate assignment with topic IDs into the
subscription state.
- Improve logging on the client assignment/reconciliation path
No changes in logic.
Reviewers: TengYao Chi <frankvicky@apache.org>, Andrew Schofield
<aschofield@confluent.io>
Change:
`public List<Optional<? extends RequestManager>> entries();`
to:
`public List<RequestManager> entries();`
and clean up the callers.
Reviewers: TengYao Chi <kitingiao@gmail.com>, Andrew Schofield
<aschofield@confluent.io>, Chia-Ping Tsai <chia7712@gmail.com>
This PR adds an additional test case to `FormatterTest` that checks that
formatting with `--standalone` and then formatting again with
`--standalone --ignore-formatted` is indeed a no-op.
Reviewers: José Armando García Sancio <jsancio@apache.org>
Currently in the tests in ShareFetchAcknowledgeRequestTest, subsequent
share fetch / share acknowledge requests creates a new socket everytime,
even when the requests are sent by the same member. In reality, a single
share consumer clisnet will reuse the same socket for all the share
related requests in its lifetime. This PR changes the behaviour in the
tests to align with reality and reuse the same socket for all requests
by the same share group member.
Reviewers: Apoorv Mittal <apoorvmittal10@gmail.com>
`ConstantBrokerOrActiveKController` was introduced in #14399, to provide
a mechanism for selecting the least loaded broker or the active
controller when using `bootstrap.controllers`.
Usage was removed in #18002, after `alterConfigs` was deprecated in
Kafka 2.4.0.
Reviewers: PoAn Yang <payang@apache.org>, Jhen-Yung Hsu
<jhenyunghsu@gmail.com>, Ken Huang <s7133700@gmail.com>, TengYao Chi
<kitingiao@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
### About
Added code to handle share partition fetch lock cleanly in
`DelayedShareFetch` to avoid a member incorrectly releasing a share
partition's fetch lock
### Testing
The code has been tested with the help of unit tests and integration
tests.
Reviewers: Andrew Schofield <aschofield@confluent.io>, Apoorv Mittal <apoorvmittal10@gmail.com>
New system test for KIP-1071.
Standby replicas need to be enabled via `kafka-configs.sh`.
Reviewers: Bill Bejeck <bill@confluent.io>, Matthias J. Sax
<matthias@confluent.io>
Reviewers: Colin P. McCabe <cmccabe@apache.org>, Kirk True <kirk@kirktrue.pro>, Jhen-Yung Hsu <jhenyunghsu@gmail.com>, Ian McDonald <ian_mcdonald@rocketmail.com>
Log a message when reading a batch that is larger than the currently
allocated batch.
Reviewers: Colin Patrick McCabe <cmccabe@apache.org>, PoAn Yang
<payang@apache.org>
When the persister is responding to a read share-group state summary
request, it has no way of including the leader epoch in its response,
even though it has the information to hand. This means that the leader
epoch information is not initialised in the admin client operation to
list share group offsets, and this then means that the information
cannot be displayed in kafka-share-groups.sh.
Reviewers: Apoorv Mittal <apoorvmittal10@gmail.com>, Sushant Mahajan
<smahajan@confluent.io>
- Removed unused fields and methods in clients tests
- Fixed IDEA code inspection warnings
Reviewers: Ken Huang <s7133700@gmail.com>, PoAn Yang
<payang@apache.org>, Andrew Schofield <aschofield@confluent.io>,
Chia-Ping Tsai <chia7712@gmail.com>, TengYao Chi
<frankvicky@apache.org>
1. Move `ForwardingManagerMetrics` and `ForwardingManagerMetricsTest` to
server module.
2. Rewrite them in Java.
Reviewers: Chia-Ping Tsai <chia7712@gmail.com>
rewrite `MetricsDuringTopicCreationDeletionTest` to `ClusterTest` infra
and move it to clients-integration-tests module.
Reviewers: PoAn Yang <payang@apache.org>, Ken Huang
<s7133700@gmail.com>, Jhen-Yung Hsu <jhenyunghsu@gmail.com>, Chia-Ping
Tsai <chia7712@gmail.com>
Sean Quah
Abhinav Dixit
Lianet Magrans
Sushant Mahajan
David Jacot
Apoorv Mittal
Andrew Schofield
Bolin Lin
Alyssa Huang
Jonah Hooper
ChickenchickenLove
Matthias J. Sax
Ming-Yen Chung
Kuan-Po Tseng
Yunchi Pang
Nick Guo
PoAn Yang
xijiu
Hong-Yi Chen
Stanislav Kozlovski
Chia-Ping Tsai
Shivsundar R
ShihYuan Lin
Manoj
Uladzislau Blok
Calvin Liu
Kirk True
Kevin Wu
Chirag Wadhwa
Lucas Brutschy
Lan Ding
José Armando García Sancio
Dmitry Werner