| commit | 1485f6cef0183564891a112454104c3780001301 | [log] [tgz] |
|---|---|---|
| author | Jacek Centkowski <geminica.programs@gmail.com> | Fri Dec 23 08:20:51 2022 +0100 |
| committer | Jacek Centkowski <geminica.programs@gmail.com> | Thu Dec 29 15:03:34 2022 +0100 |
| tree | b62eedd8830874078d42de72bdaaf43694c6f184 | |
| parent | 39f62a9da2aff9d712077d746997266247a94192 [diff] |
Don't attempt to cache refs containing '[:|@]' in resolve call The 'ref + ":" + filename' syntax is used to resolve the file's ObjectId under the specific ref. That attempt fails to succeed in 'exactRef' call therefore it falls back to the delegate 'resolve' call. The `ref + "@"` syntax (according to [1]) is used to reach out to a reflog entry or to a value of ref prior the point in time. The result (for both cases) is consistent with delegate 'resolve' call but cycles are wasted on 'exactRef' call that returns 'null'. This change adds refs that contain '[:|@]' as non cacheable. The check leverages on Guava's CharMatcher to check all chars in a single input String's loop. Note that the CachedRefRepositoryIT was rewritten to work on a FileRepository so that reflog could be enabled and `@`-related checks performed. [1] https://git-scm.com/docs/git-rev-parse#_specifying_revisions Bug: Issue 16553 Change-Id: Ia86e6bd2c0da4c7eceffc090ad1e7a63d63552fc
When Serialize AccountCache series was introduced it simplified the cache eviction by always reaching out to JGit for data. Unfortunately it comes with price which is especially high when All-Users repository is accessed through the NFS and core.trustFolderStat = false is configured in ${GERRIT_SITE}/etc/jgit.config (quite common setup for HA/Multi-Site ens).
This plugin was developed to introduce the in-memory cache (managed by Gerrit so that evictions could be coordinated to multiple nodes) that reduces the price for reaching to refs in JGit. It is a Gerrit native alternative (that can be applied to Gerrit 3.2) to work that is currently under progress for caching Refs in JGit.
Here is the short comparison of heavy-refs-related operations performance. The test scenario was to get random change details (over the same REST API that is used in Gerrit's details page) in 8 parallel threads over 5mins period of time. The core.trustFolderStat = false was set in ${GERRIT_SITE}/etc/jgit.config. It was called against:
stable-3.1 in the results)stable-3.2 in the results)stable-3.2-libCache in the results).Note that TRS is Reqs/Sec for each Thread.
| version | TRS Avg | TRS Std Dev | TRS Max | Total Reqs/sec | Transfer/sec(MB)| | --------------------------------- | ------- | ----------- | ------- | -------------- | --------------- | | stable-3.1 | 57,33 | 8,26 | 80 | 456,95 | 4,34 | | stable-3.2 | 13,87 | 4,92 | 20 | 110,18 | 1,07 | | stable-3.2-libCache | 105,27 | 14,55 | 150 | 834,88 | 8,41 | | stable-3.1 vs stable-3.2 | 313,34% | 67,89% | 300,00% | 314,73% | 305,61% | | stable-3.2-libCache vs stable-3.2 | 658,98% | 195,73% | 650,00% | 657,74% | 685,98% | | stable-3.2-libCache vs stable-3.1 | 83,62% | 76,15% | 87,50% | 82,71% | 93,78% |
One can clearly see that in this setup using this library module outperforms both Gerrit 3.2 and 3.1 by factor of 6 and 2 correspondingly. The test script, detailed description and more results are available here.
Clone or link this plugin to the plugins directory of Gerrit‘s source tree, and then run bazel build on the plugin’s directory.
Example:
git clone --recursive https://gerrit.googlesource.com/gerrit cd plugins git clone "https://gerrit.googlesource.com/modules/cached-refdb" cd .. && bazel build plugins/cached-refdb
The output plugin jar is created in:
bazel-bin/plugins/cached-refdb/cached-refdb.jar
Copy the cached-refdb.jar into the ${GERRIT_SITE}/lib/ so that it is being loaded when the Gerrit instance is started. Note that the following configuration options need to be added
git config --file ${GERRIT_SITE}/etc/gerrit.config --add gerrit.installDbModule\
com.googlesource.gerrit.plugins.cachedrefdb.LibDbModule
git config --file ${GERRIT_SITE}/etc/gerrit.config --add gerrit.installModule\
com.googlesource.gerrit.plugins.cachedrefdb.LibSysModule
NOTE: There are situations where the binding of the module to the Gerrit's GitRepositoryManager is not desired; e.g., when using this module together with others that are trying to override it at the same time.
It is possible to just load the module using the following two options:
git config --file ${GERRIT_SITE}/etc/gerrit.config --add gerrit.installDbModule\ com.googlesource.gerrit.plugins.cachedrefdb.LibModule git config --file ${GERRIT_SITE}/etc/gerrit.config --add gerrit.installModule\ com.googlesource.gerrit.plugins.cachedrefdb.LibSysModule
By default cache can hold up to 1024 refs which will not be sufficient for any production site therefore one can configure it through the standard Gerrit cache configuration means e.g.
git config --file ${GERRIT_SITE}/etc/gerrit.config cache.ref_by_name.memoryLimit 10240
Note that library module requires the Gerrit instance restart in order to pick up the configuration changes.