[
https://issues.apache.org/jira/browse/HIVE-20656?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16779828#comment-16779828
]
Gopal V commented on HIVE-20656:
--------------------------------
HIVE-20364 changed the wrong config - the comment says
{code}
Update default for hive.map.aggr.hash.min.reduction
{code}
the actual config changed was
{code}
- HIVEMAPAGGRHASHMEMORY("hive.map.aggr.hash.percentmemory", (float) 0.5,
+ HIVEMAPAGGRHASHMEMORY("hive.map.aggr.hash.percentmemory", (float) 0.99,
{code}
The mem thresholds are good at 0.5 - the GC pauses start to really trouble us
at 80%, the tez buffers are approx ~30%, so 50% is a good enough high watermark.
> Map aggregation memory configs are too aggressive
> -------------------------------------------------
>
> Key: HIVE-20656
> URL: https://issues.apache.org/jira/browse/HIVE-20656
> Project: Hive
> Issue Type: Bug
> Affects Versions: 4.0.0, 3.2.0
> Reporter: Prasanth Jayachandran
> Priority: Major
>
> The defaults for the following configs seems to be too aggressive. In java
> this can easily lead to several full GC pauses whose memory cannot be
> reclaimed.
> {code:java}
> HIVEMAPAGGRHASHMEMORY("hive.map.aggr.hash.percentmemory", (float) 0.99,
> "Portion of total memory to be used by map-side group aggregation hash
> table"),
> HIVEMAPAGGRMEMORYTHRESHOLD("hive.map.aggr.hash.force.flush.memory.threshold",
> (float) 0.9,
> "The max memory to be used by map-side group aggregation hash table.\n" +
> "If the memory usage is higher than this number, force to flush
> data"),{code}
>
> We can be little bit conservative for these configs to avoid getting into GC
> pause.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
