[
https://issues.apache.org/jira/browse/CASSANDRA-7545?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
graham sanderson updated CASSANDRA-7545:
----------------------------------------
Description:
The worst side effect is potentially orders of magnitude larger than suspected
memory allocation due to a race condition updating the columns in the memtable
(see CASSANDRA-7546)
That said, having so many hints in a single partition has other negative side
effects (non linear growth of sorted tree data structure overhead in memtable,
just plain large number of tombstones in a row etc.)
Ideally, the hints would be partitioned by both the nodeUUID, and a fixed
number of (say 4) bits of a hash of the original rowmutation's partition key...
even this small spreading of load has a significant improvement.
was:
The worst side effect is potentially orders of magnitude larger than suspected
memory allocation due to a race condition updating the columns in the memtable.
That said, having so many hints in a single partition has other negative side
effects (non linear growth of sorted tree data structure overhead in memtable,
just plain large number of tombstones in a row etc.)
Ideally, the hints would be partitioned by both the nodeUUID, and a fixed
number of (say 4) bits of a hash of the original rowmutation's partition key...
even this small spreading of load has a significant improvement.
> Hints for a down node are written to a single partition in system.hints on
> the coordinator leading to contention
> ----------------------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-7545
> URL: https://issues.apache.org/jira/browse/CASSANDRA-7545
> Project: Cassandra
> Issue Type: Bug
> Reporter: graham sanderson
>
> The worst side effect is potentially orders of magnitude larger than
> suspected memory allocation due to a race condition updating the columns in
> the memtable (see CASSANDRA-7546)
> That said, having so many hints in a single partition has other negative side
> effects (non linear growth of sorted tree data structure overhead in
> memtable, just plain large number of tombstones in a row etc.)
> Ideally, the hints would be partitioned by both the nodeUUID, and a fixed
> number of (say 4) bits of a hash of the original rowmutation's partition
> key... even this small spreading of load has a significant improvement.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
