[
https://issues.apache.org/jira/browse/PHOENIX-6655?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Geoffrey Jacoby updated PHOENIX-6655:
-------------------------------------
Description:
High volume use cases of Phoenix sequences need to write to the same row of
SYSTEM.SEQUENCE very frequently, but each time we write, the block cache that
row is in is invalidated and has to be loaded again.
[~rushabh.shah] was looking at a case in our environment where a sequence was
being used heavily and found lots of time spent reading the HFiles for the same
Get of the sequence row within the SequenceRegionObserver increment coproc.
[~apurtell] pointed out offline that HColumnDescriptor has a property,
CACHE_DATA_ON_WRITE that keeps caches updated when a value is flushed. This
should help improve contention on sequence increments.
was:
High volume use cases of Phoenix sequences need to write to the same row of
SYSTEM.SEQUENCE very frequently, but each time we write, the block cache that
row is in is invalidated and has to be loaded again.
[~rushabh.shah] was looking at a case in our environment where a sequence was
being used heavily and found lots of time spent reading the HFiles for the same
Get of the sequence row within the SequenceRegionObserver increment coproc.
[~apurtell] pointed out offline that HColumnDescriptor has a property,
CACHE_DATA_ON_WRITE that keeps caches updated when a value is updated. This
should help improve contention on sequence increments.
> SYSTEM.SEQUENCE should have CACHE_DATA_ON_WRITE set to true
> -----------------------------------------------------------
>
> Key: PHOENIX-6655
> URL: https://issues.apache.org/jira/browse/PHOENIX-6655
> Project: Phoenix
> Issue Type: Improvement
> Reporter: Geoffrey Jacoby
> Priority: Major
>
> High volume use cases of Phoenix sequences need to write to the same row of
> SYSTEM.SEQUENCE very frequently, but each time we write, the block cache that
> row is in is invalidated and has to be loaded again.
> [~rushabh.shah] was looking at a case in our environment where a sequence was
> being used heavily and found lots of time spent reading the HFiles for the
> same Get of the sequence row within the SequenceRegionObserver increment
> coproc.
> [~apurtell] pointed out offline that HColumnDescriptor has a property,
> CACHE_DATA_ON_WRITE that keeps caches updated when a value is flushed. This
> should help improve contention on sequence increments.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
