-
Notifications
You must be signed in to change notification settings - Fork 896
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Improve group and flush add-responses after journal sync #3848
Improve group and flush add-responses after journal sync #3848
Conversation
…r prepareSendResponseV2. 2.Make maxPendingResponsesSize is more reasonable.
780da22
to
20e3db7
Compare
Codecov Report
@@ Coverage Diff @@
## master #3848 +/- ##
============================================
- Coverage 68.30% 68.28% -0.02%
- Complexity 6787 6789 +2
============================================
Files 474 474
Lines 41112 41103 -9
Branches 5256 5254 -2
============================================
- Hits 28080 28068 -12
- Misses 10782 10787 +5
+ Partials 2250 2248 -2
Flags with carried forward coverage won't be shown. Click here to find out more.
... and 24 files with indirect coverage changes 📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
rerun failure checks |
@@ -495,6 +504,7 @@ public void run() { | |||
|
|||
journalStats.getForceWriteQueueSize().addCount(-requestsCount); | |||
|
|||
Set<BookieRequestHandler> writeHandlers = new HashSet<>(); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Using the set has its own overhead too. In this case it will allocate a Node
object for the linked list for each entry added here.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Perhaps we could use ObjectHashSet
from carrotsearch https://github.com/carrotsearch/hppc
It's already used in Pulsar too.
We should also reuse the same set each time.
@@ -503,17 +513,26 @@ public void run() { | |||
// responses | |||
for (int i = 0; i < requestsCount; i++) { | |||
ForceWriteRequest req = localRequests.get(i); | |||
req.getForceWriteWaiters().forEach(ele -> { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
req.process()
is already iterating over all the entries for each ForceWriteRequest
. It would be better to reuse the same loop.
Additionally using forEach
has 2 problems:
- allocates an iterator, unlike doing
for (int i =0; i... .)
- the lambda is also capturing variables and thus it will allocate a temporary object
public synchronized void flushPendingResponse() { | ||
if (pendingSendResponses != null) { | ||
maxPendingResponsesSize = (int) Math.max( | ||
maxPendingResponsesSize * 0.9 + 0.1 * pendingSendResponses.readableBytes(), |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This looks like it's going to take quite a bit of iterations to expand the cluster size. While in the meantime we have to keep expanding the ByteBuf
, which will incur in multiple allocations and memory copies
@@ -149,6 +149,10 @@ | |||
<scope>runtime</scope> | |||
<optional>true</optional> | |||
</dependency> | |||
<dependency> | |||
<groupId>com.carrotsearch</groupId> | |||
<artifactId>hppc</artifactId> |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We should also add to license file
bookkeeper-server/src/main/java/org/apache/bookkeeper/bookie/Journal.java
Outdated
Show resolved
Hide resolved
…ournal.java address the comments. Co-authored-by: Matteo Merli <mmerli@apache.org>
rerun failure checks |
rerun failure checks |
rerun failure checks |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Good work!
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Good job!
* Fix memory leak issue of reading small entries (apache#3844) * Make read entry request recyclable (apache#3842) * Make read entry request recyclable * Move recycle to finally block * Fix test and comments * Fix test * Avoid unnecessary force write. (apache#3847) * Avoid unnecessary force write. * code clean. * fix style * Correct the running job name for the test group (apache#3851) --- ### Motivation The running tests job name doesn't match the tests. Correct the job name. * add timeout for two flaky timeout tests (apache#3855) * add V2 protocal and warmupMessages support for benchMark (apache#3856) * disable trimStackTrack for code-coverage profile (apache#3854) * Fix bkperf log directory not found (apache#3858) ### Motivation When using the bkperf command `bin/bkperf journal append -j data -n 100000000 --sync true` to test the BookKeeper journal performance, it failed with the following exception ``` [0.002s][error][logging] Error opening log file '/Users/hangc/Downloads/tmp/tc/batch/ta/bookkeeper-all-4.16.0-SNAPSHOT/logs/bkperf-gc.log': No such file or directory [0.002s][error][logging] Initialization of output 'file=/Users/hangc/Downloads/tmp/tc/batch/ta/bookkeeper-all-4.16.0-SNAPSHOT/logs/bkperf-gc.log' using options 'filecount=5,filesize=64m' failed. Invalid -Xlog option '-Xlog:gc=info:file=/Users/hangc/Downloads/tmp/tc/batch/ta/bookkeeper-all-4.16.0-SNAPSHOT/logs/bkperf-gc.log::filecount=5,filesize=64m', see error log for details. Error: Could not create the Java Virtual Machine. Error: A fatal exception has occurred. Program will exit. ``` The root cause is that the `logs` directory was not created. ### Modifications Create the `logs` directory before bkperf started. * [improve] Fix indexDirs upgrade failed (apache#3762) * fix indexDirs upgrade failed * Bump checkstyle-plugin from 3.1.2 to 3.2.1 (apache#3850) * [Flaky] Fix flaky test in testRaceGuavaEvictAndReleaseBeforeRetain (apache#3857) * Fix flaky test in testRaceGuavaEvictAndReleaseBeforeRetain * format code * Fix NPE in BenchThroughputLatency (apache#3859) * Update website to 4.15.4 (apache#3862) --- ### Motivation Update website to 4.15.4 * change rocksDB config level_compaction_dynamic_level_bytes to CFOptions (apache#3860) ### Motivation After PR apache#3056 , Bookkeeper set `level_compaction_dynamic_level_bytes=true` as `TableOptions` in `entry_location_rocksdb.conf.default` , which will cause `level_compaction_dynamic_level_bytes` lose efficacy and will cause rocksDB .sst file compact sort chaos when update bookie release. As RocksDB conf, `level_compaction_dynamic_level_bytes` need set as `CFOptions` https://github.com/facebook/rocksdb/blob/master/examples/rocksdb_option_file_example.ini <img width="703" alt="image" src="https://user-images.githubusercontent.com/84127069/224640399-d5481fe5-7b75-4229-ac06-3d280aa9ae6d.png"> <img width="240" alt="image" src="https://user-images.githubusercontent.com/84127069/224640621-737d0a42-4e01-4f38-bd5a-862a93bc4b32.png"> ### Changes 1. Change `level_compaction_dynamic_level_bytes=true` from `TableOptions` to `CFOptions` in `entry_location_rocksdb.conf.default` ; * Correct the running job flag for the test group. (apache#3865) * Release note for 4.15.4 (apache#3831) --- ### Motivation Release note for 4.15.4 * Add trigger entry location index rocksDB compact interface. (apache#3802) ### Motivation After the bookie instance running long time, the bookie entry location index rocksDB `.sst` file size maybe expand to 20-30GB as one ledger data dir's location index in some case, which will cause the rocksDB scan operator cost more time and cause the bookie client request timeout. Add trigger entry location index rocksDB compact REST API which can trigger entry location rocksDB compaction and get the compaction status. The full range entry location index rocksDB compact will cause the entry location index dir express higher IOUtils. So we'd better trigger the entry location rocksDB compact by the api in low data flow period. **Some case before rocksDB compact:** <img width="232" alt="image" src="https://user-images.githubusercontent.com/84127069/220893469-e6fbc1a3-c767-4ffe-8ae9-f05ad1833c50.png"> <img width="288" alt="image" src="https://user-images.githubusercontent.com/84127069/220891359-dc37e139-37b0-461b-8001-dcc48517366c.png"> **After rocksDB compact:** <img width="255" alt="image" src="https://user-images.githubusercontent.com/84127069/220891419-24267fa7-348c-4fbd-8b3e-70a99840bce5.png"> ### Changes 1. Add REST API to trigger entry location index rocksDB compact. * Pick the higher leak detection level between netty and bookkeeper. (apache#3794) ### Motivation 1. Pick the higher leak detection level between netty and bookkeeper. 2. Enhance the bookkeeper leak detection value match rule, now it's case insensitive. There are detailed information about it: https://lists.apache.org/thread/d3zw8bxhlg0wxfhocyjglq0nbxrww3sg * Disable code coverage and codecov report (apache#3863) ### Motivation There're two reasons that we want to disable the code coverage. 1. The current report result is not accurate. 2. We can't get the PR's unit test's code coverage because of the apache Codecov permission. * Add small files check in garbage collection (apache#3631) ### Motivation When we use `TransactionalEntryLogCompactor` to compact the entry log files, it will generate a lot of small entry log files, and for those files, the file usage is usually greater than 90%, which can not be compacted unless the file usage decreased. ![image](https://user-images.githubusercontent.com/5436568/201135615-4d6072f5-e353-483d-9afb-48fad8134044.png) ### Changes We introduce the entry log file size check during compaction, and the checker is controlled by `gcEntryLogSizeRatio`. If the total entry log file size is less than `gcEntryLogSizeRatio * logSizeLimit`, the entry log file will be compacted even though the file usage is greater than 90%. This feature is disabled by default and the `gcEntryLogSizeRatio` default value is `0.0` * [improvement] Delay all audit task when have a already delayed bookie check task (apache#3818) ### Motivation Fixes apache#3817 For details, see: apache#3817 ### Changes When there is an `auditTask` during the `lostBookieRecoveryDelay` delay, other detection tasks should be skipped. * Change order of doGcLedgers and extractMetaFromEntryLogs (apache#3869) * [Bugfix] make metadataDriver initialization more robust (apache#3873) Co-authored-by: zengqiang.xu <zengqiang.xu@shopee.com> * Enable CI for the streamstorage python client (apache#3875) * Fix compaction threshold default value precision problem. (apache#3871) * Fix compaction threshold precision problem. * Fix compaction threshold precision problem. * Single buffer for small add requests (apache#3783) * Single buffer for small add requests * Fixed checkstyle * Fixed treating of ComposityByteBuf * Fixed merge issues * Fixed merge issues * WIP * Fixed test and removed dead code * Removed unused import * Fixed BookieJournalTest * removed unused import * fix the checkstyle * fix failed test * fix failed test --------- Co-authored-by: chenhang <chenhang@apache.org> * Add log for entry log file delete. (apache#3872) * Add log for entry log file delete. * add log info. * Address the comment. * Address the comment. * revert the code. * Improve group and flush add-responses after journal sync (apache#3848) Descriptions of the changes in this PR: This is an improvement for apache#3837 ### Motivation 1. Now if the maxPendingResponsesSize is expanded large, it will not decrease. => We should make it flexible. 2. Now after prepareSendResponseV2 to the channel, then we trigger all channels to flush pendingSendResponses, maybe there is only a few channels that need to flush, but if we trigger all channels, it's a waste. => We only flush the channel which prepareSendResponseV2. --------- Co-authored-by: Penghui Li <penghui@apache.org> Co-authored-by: Yong Zhang <zhangyong1025.zy@gmail.com> Co-authored-by: Hang Chen <chenhang@apache.org> Co-authored-by: wenbingshen <oliver.shen999@gmail.com> Co-authored-by: ZhangJian He <shoothzj@gmail.com> Co-authored-by: lixinyang <84127069+Nicklee007@users.noreply.github.com> Co-authored-by: YANGLiiN <ielin@qq.com> Co-authored-by: Lishen Yao <yaalsn@gmail.com> Co-authored-by: Andrey Yegorov <8622884+dlg99@users.noreply.github.com> Co-authored-by: ZanderXu <zanderxu@apache.org> Co-authored-by: zengqiang.xu <zengqiang.xu@shopee.com> Co-authored-by: Matteo Merli <mmerli@apache.org>
Descriptions of the changes in this PR: This is an improvement for apache#3837 ### Motivation 1. Now if the maxPendingResponsesSize is expanded large, it will not decrease. => We should make it flexible. 2. Now after prepareSendResponseV2 to the channel, then we trigger all channels to flush pendingSendResponses, maybe there is only a few channels that need to flush, but if we trigger all channels, it's a waste. => We only flush the channel which prepareSendResponseV2.
Descriptions of the changes in this PR:
This is an improvement for #3837
Motivation