[
https://issues.apache.org/jira/browse/CASSANDRA-13533?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16013284#comment-16013284
]
Eduard Tudenhoefner commented on CASSANDRA-13533:
-------------------------------------------------
[~jjirsa] we were using that flag in some of our integration tests where we
found out that, _very rarely_, simple things (usually taking < 1 second) would
take a long amount of time (> 30 seconds).
We boiled it then down to the point where we saw that this time was 99% spent
in the prepared stmt cache for adding/evicting elements. And in our case,
eviction was happening because a single prepared stmt would account *1mb - 6mb*
(prep stmt cache size was *8mb* for us). After looking closer why prep stmts
would be so large, we eventually saw that the binary representation of
{{ColumnIdentifier#text}} was be wrong and not what we would have expected (see
attached screenshot).
We also only detected all of this because on *cassandra-3.0* we don't use
*.omitSharedBufferOverhead()* with {{MemoryMeter}} object measurements (
[QueryProcessor.java#L67|https://github.com/apache/cassandra/blob/cassandra-3.0/src/java/org/apache/cassandra/cql3/QueryProcessor.java#L67]).
On *cassandra-3.11* things got refactored over time and the prep stmt cache was
using {{MemoryMeter}} stuff from {{ObjectSizes}}, where
*.omitSharedBufferOverhead()* is being applied and the issue wouldn't happen
([ObjectSizes.java#L35|https://github.com/apache/cassandra/blob/cassandra-3.11/src/java/org/apache/cassandra/utils/ObjectSizes.java#L35]).
> ColumnIdentifier object size wrong when tables are not flushed
> --------------------------------------------------------------
>
> Key: CASSANDRA-13533
> URL: https://issues.apache.org/jira/browse/CASSANDRA-13533
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: Eduard Tudenhoefner
> Assignee: Eduard Tudenhoefner
> Fix For: 3.0.14, 3.11.0, 4.0
>
> Attachments: columnidentifier.png
>
>
> It turns out that the object size of {{ColumnIdentifier}} is wrong when
> *cassandra.test.flush_local_schema_changes: false*. This looks like stuff is
> being wrongly reused when no flush is happening.
> We only noticed this because we were using the prepared stmt cache and
> noticed that prepared statements would account for *1-6mb* when
> *cassandra.test.flush_local_schema_changes: false*. With
> *cassandra.test.flush_local_schema_changes: true* (which is the default)
> those would be around *5000 bytes*.
> Attached is a test that reproduces the problem and also a fix.
> Also after talking to [~jkni] / [~blerer] we shouldn't probably take
> {{ColumnDefinition}} into account when measuring object sizes with
> {{MemoryMeter}}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
