Cache an entire repository to increase cache hit rate #1098
Conversation
Motivation: The size of the repositories in Central Dogma are relatively small. Therefore, caching the entire repo rather than caching each file separately would result in a higher cache hit rate. Additionally, when checking repository access patterns internally, it was found that a client tends to send multiple queries to fully scan repository files instead of using the all path pattern `/**`. Modifications: - Changed `CachingRepository` to cache all files in a repository when finding files with a path pattern. Filtering is performed to the cached entries. - Doubled the default repository cache spec to approximately 256-megachars. - Strip `/` if a matching path starts with it to apply `PathPatternFilter` to `Entry.path()` Result: It increases cache hits through repo-level cache and allows the server to quickly fill the cache when starting up.
| cacheableOptions = newOptions.build(); | ||
| } | ||
|
|
||
| return cache.get(new CacheableFindCall(repo, normalizedRevision, ALL_PATH, cacheableOptions)) |
There was a problem hiding this comment.
I understood worst case, the cached size can be (number of revisions) * (size of directory).
There was a problem hiding this comment.
Possible. Based on the user pattern I observed, most clients were only interested in HEAD revision. Old revisions were accessed occasionally. The practical cached size would be (1 or 2 revisions) * (size of directory).
|
|
||
| // Use LinkedHashMap to 1) keep the order and 2) allow callers to mutate it. | ||
| return stream.collect(Collectors.toMap(Map.Entry::getKey, Map.Entry::getValue, | ||
| (oldV, newV) -> oldV, |
There was a problem hiding this comment.
Question) Is this just a precaution? or is it actually possible that multiple values (files) exist for a single key (path)?
There was a problem hiding this comment.
(oldV, newV) -> oldV was just added because of the signature of Collectors.toMap().
https://docs.oracle.com/javase/8/docs/api/java/util/stream/Collectors.html#toMap-java.util.function.Function-java.util.function.Function-java.util.function.BinaryOperator-java.util.function.Supplier-
To use LinkedHashMap, mergeFunction is necessary.
server/src/main/java/com/linecorp/centraldogma/server/storage/repository/Repository.java
Outdated
Show resolved
Hide resolved
…repository/Repository.java Co-authored-by: jrhee17 <guins_j@guins.org>
Motivation:
The size of the repositories in Central Dogma are relatively small. Therefore, caching the entire repo rather than caching each file separately would result in a higher cache hit rate. Additionally, when checking repository access patterns internally, it was found that a client tends to send multiple queries to fully scan repository files instead of using the all path pattern
/**.Modifications:
CachingRepositoryto cache all files in a repository when finding files with a path pattern. Filtering is performed to the cached entries./if a matching path starts with it to applyPathPatternFiltertoEntry.path()Result:
It increases cache hits through repo-level cache and allows the server to quickly fill the cache when starting up.