[ 
https://issues.apache.org/jira/browse/HIVE-19206?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Rajesh Balamohan updated HIVE-19206:
------------------------------------
    Comment: was deleted

(was: Problem:
 When there are 100s of record updaters open, the amount of memory required by 
orc writers keeps growing because of ORC's internal buffers. This can lead to 
potential high GC or OOM during streaming ingest.

Solution:
 The high level idea is for the streaming connection to remember all the open 
record updaters and flush the record updater periodically (at some interval). 
Records written to each record updater can be used as a metric to determine the 
candidate record updaters for flushing. 
 If stripe size of orc file is 64MB, the default memory management check 
happens only after every 5000 rows which may which may be too late when there 
are too many concurrent writers in a process. Example case would be 100 writers 
open and each of them have almost full stripe of 64MB buffered data, this would 
take 100*64MB ~=6GB of memory. When all of the record writers flush, the memory 
usage drops down to 100*~2MB which is just ~200MB memory usage.)

> Automatic memory management for open streaming writers
> ------------------------------------------------------
>
>                 Key: HIVE-19206
>                 URL: https://issues.apache.org/jira/browse/HIVE-19206
>             Project: Hive
>          Issue Type: Sub-task
>          Components: Streaming
>    Affects Versions: 3.0.0, 3.1.0
>            Reporter: Prasanth Jayachandran
>            Assignee: Prasanth Jayachandran
>            Priority: Major
>
> Problem:
>  When there are 100s of record updaters open, the amount of memory required 
> by orc writers keeps growing because of ORC's internal buffers. This can lead 
> to potential high GC or OOM during streaming ingest.
> Solution:
>  The high level idea is for the streaming connection to remember all the open 
> record updaters and flush the record updater periodically (at some interval). 
> Records written to each record updater can be used as a metric to determine 
> the candidate record updaters for flushing. 
>  If stripe size of orc file is 64MB, the default memory management check 
> happens only after every 5000 rows which may which may be too late when there 
> are too many concurrent writers in a process. Example case would be 100 
> writers open and each of them have almost full stripe of 64MB buffered data, 
> this would take 100*64MB ~=6GB of memory. When all of the record writers 
> flush, the memory usage drops down to 100*~2MB which is just ~200MB memory 
> usage.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to