[
https://issues.apache.org/jira/browse/HIVE-20153?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16558917#comment-16558917
]
Gopal V commented on HIVE-20153:
--------------------------------
LGTM - +1 tests pending.
This extra field is still taking up meaningful amounts of memory for the
objects in the heap.
>From JOL.
{code}
***** 64-bit VM: **********************************************************
org.apache.hadoop.hive.ql.udf.generic.GenericUDAFSum$GenericUDAFSumEvaluator$SumAgg
object internals:
OFFSET SIZE TYPE DESCRIPTION
VALUE
0 16 (object header) N/A
16 1 boolean SumAgg.empty N/A
17 7 (alignment/padding gap)
24 8 java.lang.Object SumAgg.sum N/A
32 8 java.util.HashSet SumAgg.uniqueObjects N/A
Instance size: 40 bytes
Space losses: 7 bytes internal + 0 bytes external = 7 bytes total
...
***** 64-bit VM, compressed references enabled: ***************************
org.apache.hadoop.hive.ql.udf.generic.GenericUDAFSum$GenericUDAFSumEvaluator$SumAgg
object internals:
OFFSET SIZE TYPE DESCRIPTION
VALUE
0 12 (object header) N/A
12 1 boolean SumAgg.empty N/A
13 3 (alignment/padding gap)
16 4 java.lang.Object SumAgg.sum N/A
20 4 java.util.HashSet SumAgg.uniqueObjects N/A
Instance size: 24 bytes
Space losses: 3 bytes internal + 0 bytes external = 3 bytes total
{code}
a PTF specific sub-class would remove that part & let me think of a way of
having a SumAggEmpty class (the "which class is it" goes into the 12 byte obj
header).
> Count and Sum UDF consume more memory in Hive 2+
> ------------------------------------------------
>
> Key: HIVE-20153
> URL: https://issues.apache.org/jira/browse/HIVE-20153
> Project: Hive
> Issue Type: Bug
> Components: UDF
> Affects Versions: 2.3.2
> Reporter: Szehon Ho
> Assignee: Aihua Xu
> Priority: Major
> Attachments: HIVE-20153.1.patch, Screen Shot 2018-07-12 at 6.41.28
> PM.png
>
>
> While playing with Hive2, we noticed that queries with a lot of count() and
> sum() aggregations run out of memory on Hadoop side where they worked before
> in Hive1.
> In many queries, we have to double the Mapper Memory settings (in our
> particular case mapreduce.map.java.opts from -Xmx2000M to -Xmx4000M), it
> makes it not so easy to upgrade to Hive 2.
> Taking heap dump, we see one of the main culprit is the field 'uniqueObjects'
> in GeneraicUDAFSum and GenericUDAFCount, which was added to support Window
> functions.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
