[
https://issues.apache.org/jira/browse/CASSANDRA-15261?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16901055#comment-16901055
]
Benedict commented on CASSANDRA-15261:
--------------------------------------
Hi [~sdycreate],
This is in the category of optimisations the project has discussed on a number
of occasions. Am I right in understanding that your proposed solution caches
the _result_ of serialisation, not just its size? Since the latter should
already be cached, particularly in 4.0. This approach would also have some
negative consequences, by increasing the total amount of memory in flight at
any one time, potentially increasing the pressure on GC and exacerbating the
cost of a slow or dying recipient node.
CASSANDRA-9834 would be my starting point here, although it is non-trivial. As
part of this work, serialising to the same format would be valuable, (although
in 4.0 serialization to peers for some data is dependent on the current time in
nanos, so we would need to update this part when serialising to peers). This
has the advantage of _reducing_ in-flight memory usage, by eliminating the need
to maintain the {{Mutation}} object until we receive ACKs for writing hints,
and permits us to (if necessary) temporarily shed message from main-memory for
ordinary delivery, without actually shedding the message itself.
This is a fairly involved piece of work, which is why the project hasn't as yet
undertaken it.
> Improve performance by adding serialize cache in mutation
> ---------------------------------------------------------
>
> Key: CASSANDRA-15261
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15261
> Project: Cassandra
> Issue Type: Improvement
> Components: Consistency/Batch Log
> Reporter: Shen dayu
> Priority: Normal
>
> I am trying to add a serialize cache in mutation, for there is so many
> redundant serialization
> # calculate binary size before serialize to binary, this calculation double
> the serialization operation.
> # reuse binary in network message for commitlog, this can reduce a
> serialization operation
> In my test, I use cassandra version is 3.0.14, and it gains 20% performance
> improvement.
> I know cassandra 3.0.x only accept critical bug fixes now.
> So I was wondering any one would review my code if I patch this for 4.0
--
This message was sent by Atlassian JIRA
(v7.6.14#76016)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
