ben-manes commented on a change in pull request #3215:
URL: https://github.com/apache/hbase/pull/3215#discussion_r638239220
##########
File path:
hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/TinyLfuBlockCache.java
##########
@@ -158,7 +158,13 @@ public boolean containsBlock(BlockCacheKey cacheKey) {
@Override
public Cacheable getBlock(BlockCacheKey cacheKey,
boolean caching, boolean repeat, boolean updateCacheMetrics) {
- Cacheable value = cache.getIfPresent(cacheKey);
+ Cacheable value = cache.asMap().computeIfPresent(cacheKey, (blockCacheKey,
cacheable) -> {
+ // It will be referenced by RPC path, so increase here. NOTICE: Must do
the retain inside
+ // this block. because if retain outside the map#computeIfPresent, the
evictBlock may remove
+ // the block and release, then we're retaining a block with refCnt=0
which is disallowed.
+ cacheable.retain();
+ return cacheable;
+ });
Review comment:
Thanks @saintstack. This was from my analysis when contributing the
original patch.
Zipfian is wonderful for a perf benchmark by stressing locks, etc. to find
bottlenecks, but isn't realistic for actual production performance. I'm not
sure if there is a great approach other than network record/replay or
canarying.
If you have a workload trace we can try to simulate that, where the hit
rates should be better (e.g. [database
trace](https://github.com/ben-manes/caffeine/wiki/Efficiency#database). That
wouldn't show actual system behavior, just the cache's expected hit rates in
isolation. HBase's LRU is similar-ish to SLRU, so then ARC might be a good
upper bound of expectations.
Between zipfian benchmark and trace simulations, we can get a roughish idea
of if there is a benefit. Otherwise canarying is the best that I've seen so
far, which is a bit heavy handed but simple.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
