[
https://issues.apache.org/jira/browse/CASSANDRA-12453?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15771572#comment-15771572
]
Jay Zhuang commented on CASSANDRA-12453:
----------------------------------------
The issue can be consistently reproduced on machine with > 16 CPUs (which will
use [64 cache
segments|https://github.com/snazy/ohc/blob/develop/ohc-core/src/main/java/org/caffinitas/ohc/OHCacheBuilder.java#L159]),
or override the
[{{segmentCount}}|https://github.com/snazy/ohc/blob/develop/ohc-core/src/main/java/org/caffinitas/ohc/OHCacheBuilder.java#L162]
to [64 |
https://github.com/cooldoger/cassandra/commit/53facbb6e75a9adf69be8622ea0d951d3996cc8a]:
{code}
[junit] Testcase:
testRowCachePartialLoad(org.apache.cassandra.db.RowCacheTest): FAILED
[junit] expected:<50> but was:<48>
[junit] junit.framework.AssertionFailedError: expected:<50> but was:<48>
[junit] at
org.apache.cassandra.db.RowCacheTest.rowCacheLoad(RowCacheTest.java:413)
[junit] at
org.apache.cassandra.db.RowCacheTest.testRowCachePartialLoad(RowCacheTest.java:231)
{code}
When there're only 100 keys are [hashed to 64
segments|https://github.com/snazy/ohc/blob/develop/ohc-core/src/main/java/org/caffinitas/ohc/linked/OHCacheLinkedImpl.java#L275],
16 segments are actually empty. The hotKeyIterator try to get [one key per
segments|https://github.com/snazy/ohc/blob/develop/ohc-core/src/main/java/org/caffinitas/ohc/linked/OHCacheLinkedImpl.java#L1396],
which will only get 48.
The fix is to increase the data size to cover more segments, so the
hotKeyIterator is able to get 50 keys. I don't think it will impact Cassandra,
as in production, the cache segments should be filled with more data, even get
less cache keys should not be a problem.
The fix is attached, please review.
> AutoSavingCache does not store required keys making RowCacheTests Flaky
> -----------------------------------------------------------------------
>
> Key: CASSANDRA-12453
> URL: https://issues.apache.org/jira/browse/CASSANDRA-12453
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: sankalp kohli
> Assignee: Jay Zhuang
> Priority: Minor
> Fix For: 2.2.x, 3.0.x
>
> Attachments: 12453-2.2.txt
>
>
> RowCacheTests were flaky and while investigating, I found that it does not
> store all the keys to disk.
> The reason is that we use OHCache and call hotKeyIterator on it. This is not
> guaranteed to return the number of keys we want. Here is the documentation
> from OHCache
> /**
> * Builds an iterator over the N most recently used keys returning
> deserialized objects.
> * You must call {@code close()} on the returned iterator.
> * <p>
> * Note: During a rehash, the implementation might return keys twice
> or not at all.
> * </p>
> */
> CloseableIterator<K> hotKeyIterator(int n);
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
