[
https://issues.apache.org/jira/browse/HIVE-10940?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14615615#comment-14615615
]
Gopal V commented on HIVE-10940:
--------------------------------
[~hagleitn]: this fixes the leak, but reintroduces the performance issue. Added
log lines and it showed for query27
{code}
2015-07-06 13:08:31,521 INFO [InputInitializer [Map 5] #0] io.HiveInputFormat:
hasObj = false, hasExpr=true
2015-07-06 13:08:31,522 INFO [InputInitializer [Map 5] #0] io.HiveInputFormat:
hive.io.file.readcolumn.ids=0,6
2015-07-06 13:08:31,522 INFO [InputInitializer [Map 5] #0] io.HiveInputFormat:
hive.io.file.readcolumn.names=d_date_sk,d_year
{code}
so it hits the serialize codepath still
{code}
if (!hasObj) {
serializedFilterObj = Utilities.serializeObject(filterObject);
}
{code}
> HiveInputFormat::pushFilters serializes PPD objects for each getRecordReader
> call
> ---------------------------------------------------------------------------------
>
> Key: HIVE-10940
> URL: https://issues.apache.org/jira/browse/HIVE-10940
> Project: Hive
> Issue Type: Bug
> Components: File Formats
> Affects Versions: 1.2.0
> Reporter: Gopal V
> Assignee: Gunther Hagleitner
> Fix For: 2.0.0
>
> Attachments: HIVE-10940.01.patch, HIVE-10940.02.patch,
> HIVE-10940.03.patch, HIVE-10940.patch
>
>
> {code}
> String filterText = filterExpr.getExprString();
> String filterExprSerialized = Utilities.serializeExpression(filterExpr);
> {code}
> the serializeExpression initializes Kryo and produces a new packed object for
> every split.
> HiveInputFormat::getRecordReader -> pushProjectionAndFilters -> pushFilters.
> And Kryo is very slow to do this for a large filter clause.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
