[
https://issues.apache.org/jira/browse/CASSANDRA-21209?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Stefan Miklosovic updated CASSANDRA-21209:
------------------------------------------
Description: We should rework the current code to get rid of notions of
auto-training even further. We need to create a trainer on demand per each
training session rather than having one trainer per whole manager. That will
enable us to eventually do auto-training as well, albeit a little bit
differently (but simpler). (was: We can specify how much sampling data there
should be for training of a dictionary. The buffer for training is a direct
buffer. If we say that we will be training on 2GiB, then it will try to create
a direct buffer of size 2GiB.
This problem is visible e.g. when starting Cassandra via IDEA which uses 1GiB
heap. I think how it works is that max direct memory size for JVM is basically
same as xmx so it will fail.
The fix would consist of wrapping creation of that buffer in a try-catch and
propagate the error in some sanitized way up to a caller informing them it is
not possible to create such a big sampling buffer.)
> Rework ZSTD dictionary compression logic to create a trainer per training
> -------------------------------------------------------------------------
>
> Key: CASSANDRA-21209
> URL: https://issues.apache.org/jira/browse/CASSANDRA-21209
> Project: Apache Cassandra
> Issue Type: Bug
> Components: Feature/Compression
> Reporter: Stefan Miklosovic
> Priority: Normal
> Fix For: 5.x
>
> Time Spent: 0.5h
> Remaining Estimate: 0h
>
> We should rework the current code to get rid of notions of auto-training even
> further. We need to create a trainer on demand per each training session
> rather than having one trainer per whole manager. That will enable us to
> eventually do auto-training as well, albeit a little bit differently (but
> simpler).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
