This change fixes a performance regression due to follower last seen highwatermark
handling introduced in 23beeea. maybeUpdateHwAndSendResponse is expensive for
brokers with high partition counts, as it requires a partition and a replica lookup for every
partition being fetched. This refactor moves the last seen watermark update into the follower
fetch state update where we have already looked up the partition and replica.
I've seen cases where maybeUpdateHwAndSendResponse is responsible 8% of CPU usage, not including the responseCallback call that is part of it.
I have benchmarked this change with `UpdateFollowerFetchStateBenchmark` and it adds 5ns
of overhead to Partition.updateFollowerFetchState, which is a rounding error compared to the
current overhead of maybeUpdateHwAndSendResponse.
Reviewers: David Arthur <mumrah@gmail.com>, Jason Gustafson <jason@confluent.io>, Ismael Juma <ismael@juma.me.uk>
Replaced UpdateMetadata{Request, Response}, LeaderAndIsr{Request, Response}
and StopReplica{Request, Response} with the automated protocol classes.
Updated the JSON schema for the 3 request types to be more consistent and
less strict (if needed to avoid duplication).
The general approach is to avoid generating new collections in the request
classes. Normalization happens in the constructor to make this possible. Builders
still have to group by topic to maintain the external ungrouped view.
Introduced new tests for LeaderAndIsrRequest and UpdateMetadataRequest to
verify that the new logic is correct.
A few other clean-ups/fixes in code that was touched due to these changes:
* KAFKA-8956: Refactor DelayedCreatePartitions#updateWaiting to avoid modifying
collection in foreach.
* Avoid unnecessary allocation for state change trace logging if trace logging is not enabled
* Use `toBuffer` instead of `toList`, `toIndexedSeq` or `toSeq` as it generally performs
better and it matches the performance characteristics of `java.util.ArrayList`. This is
particularly important when passing such instances to Java code.
* Minor refactoring for clarity and readability.
* Removed usage of deprecated `/:`, unused imports and unnecessary `var`s.
* Include exception in `AdminClientIntegrationTest` failure message.
* Move StopReplicaRequest verification in `AuthorizerIntegrationTest` to the end
to match the comment.
Reviewers: Colin Patrick McCabe <cmccabe@apache.org>
This PR makes two changes to code in the ReplicaManager.updateFollowerFetchState path, which is in the hot path for follower fetches. Although calling ReplicaManager.updateFollowerFetch state is inexpensive on its own, it is called once for each partition every time a follower fetch occurs.
1. updateFollowerFetchState no longer calls maybeExpandIsr when the follower is already in the ISR. This avoid repeated expansion checks.
2. Partition.maybeIncrementLeaderHW is also in the hot path for ReplicaManager.updateFollowerFetchState. Partition.maybeIncrementLeaderHW calls Partition.remoteReplicas four times each iteration, and it performs a toSet conversion. maybeIncrementLeaderHW now avoids generating any intermediate collections when updating the HWM.
**Benchmark results for Partition.updateFollowerFetchState on a r5.xlarge:**
Old:
```
1288.633 ±(99.9%) 1.170 ns/op [Average]
(min, avg, max) = (1287.343, 1288.633, 1290.398), stdev = 1.037
CI (99.9%): [1287.463, 1289.802] (assumes normal distribution)
```
New (when follower fetch offset is updated):
```
261.727 ±(99.9%) 0.122 ns/op [Average]
(min, avg, max) = (261.565, 261.727, 261.937), stdev = 0.114
CI (99.9%): [261.605, 261.848] (assumes normal distribution)
```
New (when follower fetch offset is the same):
```
68.484 ±(99.9%) 0.025 ns/op [Average]
(min, avg, max) = (68.446, 68.484, 68.520), stdev = 0.023
CI (99.9%): [68.460, 68.509] (assumes normal distribution)
```
Reviewers: Ismael Juma <ismael@juma.me.uk>, Jason Gustafson <jason@confluent.io>
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>
* KAFKA-8106:Reducing the allocation and copying of ByteBuffer when logValidator do validation.
* KAFKA-8106:Reducing the allocation and copying of ByteBuffer when logValidator do validation.
* github comments
* use batch.skipKeyValueIterator
* cleanups
* no need to skip kv for uncompressed iterator
* checkstyle fixes
* fix findbugs
* adding unit tests
* reuse decompression buffer; and using streaming iterator
* checkstyle
* add unit tests
* remove reusing buffer supplier
* fix unit tests
* add unit tests
* use streaming iterator
* minor refactoring
* rename
* github comments
* github comments
* reuse buffer at DefaultRecord caller
* some further optimization
* major refactoring
* further refactoring
* update comment
* github comments
* minor fix
* add jmh benchmarks
* update jmh
* github comments
* minor fix
* github comments
Verified that the https links work.
I didn't update the license header in this PR since that touches
so many files. Will file a separate one for that.
Reviewers: Manikumar Reddy <manikumar.reddy@gmail.com>
- Use Xlint:all with 3 exclusions (filed KAFKA-7613 to remove the exclusions)
- Use the same javac options when compiling tests (seems accidental that
we didn't do this before)
- Replaced several deprecated method calls with non-deprecated ones:
- `KafkaConsumer.poll(long)` and `KafkaConsumer.close(long)`
- `Class.newInstance` and `new Integer/Long` (deprecated since Java 9)
- `scala.Console` (deprecated in Scala 2.11)
- `PartitionData` taking a timestamp (one of them seemingly a bug)
- `JsonMappingException` single parameter constructor
- Fix unnecessary usage of raw types in several places.
- Add @SuppressWarnings for deprecations, unchecked and switch fallthrough in
several places.
- Scala clean-ups (var -> val, ETA expansion warnings, avoid reflective calls)
- Use lambdas to simplify code in a few places
- Add @SafeVarargs, fix varargs usage and remove unnecessary `Utils.mkList` method
Reviewers: Matthias J. Sax <mjsax@apache.org>, Manikumar Reddy <manikumar.reddy@gmail.com>, Randall Hauch <rhauch@gmail.com>, Bill Bejeck <bill@confluent.io>, Stanislav Kozlovski <stanislav_kozlovski@outlook.com>
Fixes two issues with the JMH benchmark example:
* Trivial: The output should be in `ops/ms` for readability
reasons (it's in the millions of operations per second)
* Important: The benchmark is not actually measuring the
LRU Cache performance as most of the time in each run is
wasted on concatenating `key + counter` as well as
`value + counter`. Fixed by pre-generating 10k K-V pairs
(100x the cache capacity) and iterating over them. This
brings the performance up by a factor of more than 5 on
a standard 4 core i7 (`~6k/ms` before goes to `~35k/ms`).
Author: Armin Braun <me@obrown.io>
Reviewers: Bill Bejeck <bbejeck@gmail.com>, Guozhang Wang <wangguoz@gmail.com>, Ismael Juma <ismael@juma.me.uk>
Closes#2903 from original-brownbear/fix-jmh-example
…h-benchmarks/jmh.sh
Author: bbejeck <bbejeck@gmail.com>
Reviewers: Ismael Juma <ismael@juma.me.uk>, Guozhang Wang <wangguoz@gmail.com>
Closes#2654 from bbejeck/KAFKA-3989_follow_up
More specifically, fix the case where a compressed V0 or V1 message is
larger than the producer batch size.
Author: Jason Gustafson <jason@confluent.io>
Reviewers: Apurva Mehta <apurva@confluent.io>, Ismael Juma <ismael@juma.me.uk>
Closes#3356 from hachikuji/KAFKA-5456
I included a JMH benchmark and the results follow. The
implementation in this PR takes no more than 1/10th
of the time when compared to trunk. I also included
results for an alternative implementation that is a little
slower than the one in the PR.
Trunk:
```text
TopicBenchmark.testValidate topic avgt 15 134.107 ± 3.956 ns/op
TopicBenchmark.testValidate longer-topic-name avgt 15 316.241 ± 13.379 ns/op
TopicBenchmark.testValidate very-long-topic-name_with_more_text avgt 15 636.026 ± 30.272 ns/op
```
Implementation in the PR:
```text
TopicBenchmark.testValidate topic avgt 15 13.153 ± 0.383 ns/op
TopicBenchmark.testValidate longer-topic-name avgt 15 26.139 ± 0.896 ns/op
TopicBenchmark.testValidate very-long-topic-name.with_more_text avgt 15 44.829 ± 1.390 ns/op
```
Alternative implementation where boolean validChar = Character.isLetterOrDigit(c) || c == '.' || c == '_' || c == '-';
```text
TopicBenchmark.testValidate topic avgt 15 18.883 ± 1.044 ns/op
TopicBenchmark.testValidate longer-topic-name avgt 15 36.696 ± 1.220 ns/op
TopicBenchmark.testValidate very-long-topic-name_with_more_text avgt 15 65.956 ± 0.669 ns/op
```
Author: Ismael Juma <ismael@juma.me.uk>
Reviewers: Guozhang Wang <wangguoz@gmail.com>
Closes#3234 from ijuma/optimise-topic-is-valid
The JMH benchmark included shows that the redundant
volatile write causes the constructor of `ProducerRecord`
to take more than 50% longer:
ProducerRecordBenchmark.constructorBenchmark avgt 15 24.136 ± 1.458 ns/op (before)
ProducerRecordBenchmark.constructorBenchmark avgt 15 14.904 ± 0.231 ns/op (after)
Author: Ismael Juma <ismael@juma.me.uk>
Reviewers: Jason Gustafson <jason@confluent.io>
Closes#3233 from ijuma/remove-volatile-write-in-records-header-constructor
- reuse decompression buffers in consumer Fetcher
- switch lz4 input stream to operate directly on ByteBuffers
- avoids performance impact of catching exceptions when reaching the end of legacy record batches
- more tests with both compressible / incompressible data, multiple
blocks, and various other combinations to increase code coverage
- fixes bug that would cause exception instead of invalid block size
for invalid incompressible blocks
- fixes bug if incompressible flag is set on end frame block size
Overall this improves LZ4 decompression performance by up to 40x for small batches.
Most improvements are seen for batches of size 1 with messages on the order of ~100B.
We see at least 2x improvements for for batch sizes of < 10 messages, containing messages < 10kB
This patch also yields 2-4x improvements on v1 small single message batches for other compression types.
Full benchmark results can be found here
https://gist.github.com/xvrl/05132e0643513df4adf842288be86efd
Author: Xavier Léauté <xavier@confluent.io>
Author: Ismael Juma <ismael@juma.me.uk>
Reviewers: Jason Gustafson <jason@confluent.io>, Ismael Juma <ismael@juma.me.uk>
Closes#2967 from xvrl/kafka-5150
Lucas Bradstreet
Guozhang Wang
Ismael Juma
Lifei Chen
Armin Braun
Bill Bejeck
Jason Gustafson
Xavier Léauté