cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "graham sanderson (JIRA)" <>
Subject [jira] [Commented] (CASSANDRA-7546) AtomicSortedColumns.addAllWithSizeDelta has a spin lock that allocates memory
Date Tue, 15 Jul 2014 23:21:05 GMT


graham sanderson commented on CASSANDRA-7546:

A few random thoughts

1) for 2.0.x I'd be tempted (pending further review) to use something like my patch
We could make this a global cassandra.yaml option and/or make it a per CF setting, but I think
we should just make sure it looks good and go with it
2) It might be nice to track approximate (updated without locks) max contention count according
to the read of contentionCount, and expose it as a gauge for CF metrics - probably worth a
separate issue
3) for 2.1.x We should probably avoid the the duplicate code by better encapsulating the behavior
in an object (possibly an expanded variant of the ColumnUpdater), or move to the suggested
3.0.x approach

w.r.t. to 3)
a) It might be nice to do the updates very lazily at least under load, so as to reduce overhead
of per mutation partial tree copy-on-write by batching writes (you'd have to roll back and
start over in the case one atomic mutation failed mid-operation)
b) Have you given much thought to racing reads...

> AtomicSortedColumns.addAllWithSizeDelta has a spin lock that allocates memory
> -----------------------------------------------------------------------------
>                 Key: CASSANDRA-7546
>                 URL:
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Core
>            Reporter: graham sanderson
>            Assignee: graham sanderson
>         Attachments: suggestion1.txt, suggestion1_21.txt
> In order to preserve atomicity, this code attempts to read, clone/update, then CAS the
state of the partition.
> Under heavy contention for updating a single partition this can cause some fairly staggering
memory growth (the more cores on your machine the worst it gets).
> Whilst many usage patterns don't do highly concurrent updates to the same partition,
hinting today, does, and in this case wild (order(s) of magnitude more than expected) memory
allocation rates can be seen (especially when the updates being hinted are small updates to
different partitions which can happen very fast on their own) - see CASSANDRA-7545
> It would be best to eliminate/reduce/limit the spinning memory allocation whilst not
slowing down the very common un-contended case.

This message was sent by Atlassian JIRA

View raw message