1) When the local.retention.ms/bytes is set to -2, we didn't replace it with the server-side retention.ms/bytes config, so the -2 local retention won't take effect.
2) When setting retention.ms/bytes to -2, we can notice this log message:
```
Deleting segment LogSegment(baseOffset=10045, size=1037087, lastModifiedTime=1724040653922, largestRecordTimestamp=1724040653835) due to local log retention size -2 breach. Local log size after deletion will be 13435280. (kafka.log.UnifiedLog) [kafka-scheduler-6]
```
This is not helpful for users. We should replace -2 with real retention value when logging.
Reviewers: Chia-Ping Tsai <chia7712@gmail.com>
The below remote log configs can be configured dynamically:
1. remote.log.manager.copy.max.bytes.per.second
2. remote.log.manager.fetch.max.bytes.per.second and
3. remote.log.index.file.cache.total.size.bytes
If those values are configured dynamically, then during the broker restart, it ensures the dynamic values are loaded instead of the static values from the config.
Reviewers: Chia-Ping Tsai <chia7712@gmail.com>, Satish Duggana <satishd@apache.org>, Luke Chen <showuon@gmail.com>
A patch for KAFKA-15046 got rid of fsync on LeaderEpochFileCache#truncateFromStart/End for performance reason, but it turned out this could cause corrupted leader-epoch checkpoint file on ungraceful OS shutdown, i.e. OS shuts down in the middle when kernel is writing dirty pages back to the device.
To address this problem, this PR makes below changes: (1) Revert LeaderEpochCheckpoint#write to always fsync
(2) truncateFromStart/End now call LeaderEpochCheckpoint#write asynchronously on scheduler thread
(3) UnifiedLog#maybeCreateLeaderEpochCache now loads epoch entries from checkpoint file only when current cache is absent
Reviewers: Jun Rao <junrao@gmail.com>
- Use SystemTime instead of MockTime when time is not mocked
- Use static assertions to reduce the line length
- Fold the lines if it exceeds the limit
- rename tp0 to tpId0 when it refers to TopicIdPartition
Reviewers: Kuan-Po (Cooper) Tseng <brandboat@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
Improve consistency and correctness for user-provided timeouts at the Consumer network request layer, per the Java client Consumer timeouts design (https://cwiki.apache.org/confluence/display/KAFKA/Java+client+Consumer+timeouts). While the changes introduced in KAFKA-15974 enforce timeouts at the Consumer's event layer, this change enforces timeouts at the network request layer.
The changes mostly fit into the following areas:
1. Create shared code and idioms so timeout handling logic is consistent across current and future RequestManager implementations
2. Use deadlineMs instead of expirationMs, expirationTimeoutMs, retryExpirationTimeMs, timeoutMs, etc.
3. Update "preemptive pruning" to remove expired requests that have had at least one attempt
Reviewers: Lianet Magrans <lianetmr@gmail.com>, Bruno Cadonna <cadonna@apache.org>
add more unit tests to LogSegments and do some small refactor in LogSegments.java
Reviewers: Luke Chen <showuon@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
Added the implementation of the quota manager that will be used to throttle copy and fetch requests from the remote storage. Reference KIP-956
Reviewers: Luke Chen <showuon@gmail.com>, Kamal Chandraprakash <kchandraprakash@uber.com>, Jun Rao <junrao@gmail.com>
Skip using stream when expiring the producer ID. This can improve the performance significantly when the count is high.
Before
Benchmark (numProducerIds) Mode Cnt Score Error Units
ProducerStateManagerBench.testDeleteExpiringIds 10000 avgt 3 101.253 ± 28.031 us/op
ProducerStateManagerBench.testDeleteExpiringIds 100000 avgt 3 2297.219 ± 1690.486 us/op
ProducerStateManagerBench.testDeleteExpiringIds 1000000 avgt 3 30688.865 ± 16348.768 us/op
After
Benchmark (numProducerIds) Mode Cnt Score Error Units
ProducerStateManagerBench.testDeleteExpiringIds 10000 avgt 3 39.122 ± 1.151 us/op
ProducerStateManagerBench.testDeleteExpiringIds 100000 avgt 3 464.363 ± 98.857 us/op
ProducerStateManagerBench.testDeleteExpiringIds 1000000 avgt 3 5731.169 ± 674.380 us/op
Also, made a change to the JMH testing which excludes the producer ID populating from the testing.
Reviewers: Artem Livshits <alivshits@confluent.io>, Justine Olshan <jolshan@confluent.io>
After JBOD is supported in KRaft, we should also enable JBOD support in tiered storage. Unit tests and Integration tests are also added.
Reviewers: Satish Duggana <satishd@apache.org>, Kamal Chandraprakash <kamal.chandraprakash@gmail.com>, Igor Soarez <soarez@apple.com>, Mickael Maison <mickael.maison@gmail.com>
Don't throw OFFSET_OUT_OF_RANGE error when converting the offset to metadata, and next time the leader should increment the high watermark by itself after receiving fetch requests from followers. This can happen when checkpoint files are missing and being elected as a leader.
Reviewers: Luke Chen <showuon@gmail.com>, Jun Rao <junrao@apache.org>
1. Replace TopicBasedRemoteLogMetadataManagerWrapperWithHarness with RemoteLogMetadataManagerTestUtils#builder in RemoteLogMetadataManagerTest.
2. Use ClusterTestExtention for RemoteLogMetadataManagerTest.
Signed-off-by: PoAn Yang <payang@apache.org>
Reviewers: Luke Chen <showuon@gmail.com>
TopicBasedRemoteLogMetadataManagerMultipleSubscriptionsTest has a race when it sets RemoteLogMetadataTopicPartitioner using the setter.
This change fixes the race condition by passing the RemoteLogMetadataTopicPartitioner instance in a Function<Integer, RemoteLogMetaedataTopicPartitioner> which is used in configure() in TopicBasedRemoteLogMetadataManager.
It also improves the waitingFor condition by spying on RemotePartitionMetadataStore and awaiting on Phasers to ensure ConsumerManager makes progress before performing assertions.
Reviewers: Kamal Chandraprakash <kamal.chandraprakash@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
Some tests in TopicBasedRemoteLogMetadataManagerTest flake because waitUntilConsumerCatchesUp may break early before consumer manager has caught up with all the events.
This PR adds an expected offsets for leader/follower metadataOffset partitions and ensures we wait for the offset to be at least equal to the argument to avoid flakyness.
Reviewers: Satish Duggana <satishd@apache.org>, Kamal Chandraprakash <kamal.chandraprakash@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
If there are some logs to be deleted during the log dir movement, we'll send for a scheduler to do the deletion later.
However, when the log dir movement completed, the future log is renamed, the async log deletion will fail with no file existed error.
Signed-off-by: PoAn Yang <payang@apache.org>
Reviewers: Luke Chen <showuon@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>, SoontaekLim <soontaek.lim@neya.kr>, Johnny Hsu <johnnyhsu@fb.com>
RemoteLogMetadataSerde references RemoteLogMetadataTransform in a Raw
form. Given that the class is parametrized we should make use of it.
Signed-off-by: Josep Prat <josep.prat@aiven.io>
Reviewers: Matthew de Detrich <matthew.dedetrich@aiven.io>, Mickael Maison <mickael.maison@gmail.com>
We do iterate the records to find the offsetOfMaxTimestamp instead of returning the cached one when handling ListOffsetsRequest.MAX_TIMESTAMP, since it is hard to align all paths to get correct offsetOfMaxTimestamp. The known paths are shown below.
1. convertAndAssignOffsetsNonCompressed -> we CAN get correct offsetOfMaxTimestamp when validating all records
2. assignOffsetsNonCompressed -> ditto
3. validateMessagesAndAssignOffsetsCompressed -> ditto
4. validateMessagesAndAssignOffsetsCompressed#buildRecordsAndAssignOffsets -> ditto
5. appendAsFollow#append#analyzeAndValidateRecords -> we CAN'T get correct offsetOfMaxTimestamp as iterating all records is expensive when fetching records from leader
6. LogSegment#recover -> ditto
Reviewers: Jun Rao <junrao@gmail.com>
- Use `Empty` instead of 'none' when referring to `Optional` values.
- `Headers.lastHeader` returns `null` when no header is found.
- Fix minor spelling mistakes.
Reviewers: Chia-Ping Tsai <chia7712@gmail.com>
- Fix the verifying logic. If it's LOG_APPEND_TIME, we choose the offset of the first record. Else, we choose the record with the maxTimeStamp.
- rename the shallowOffsetOfMaxTimestamp to offsetOfMaxTimestamp
Reviewers: Jun Rao <junrao@gmail.com>, Luke Chen <showuon@gmail.com>, Chia-Ping Tsai <chia7712@gmail.com>
Mickael Maison
Kuan-Po Tseng
PoAn Yang
Kamal Chandraprakash
Okada Haruki
Chia Chuan Yu
Murali Basani
Kirk True
Anatoly Popov
Abhijeet Kumar
Calvin Liu
Luke Chen
Chia-Ping Tsai
Gaurav Narula
Linu Shibu
Omnia Ibrahim
Cheng-Kai, Zhang
Josep Prat
Alok Thatikunta
Erik van Oosten
Johnny Hsu