[
https://issues.apache.org/jira/browse/HIVE-15339?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15714800#comment-15714800
]
Rajesh Balamohan edited comment on HIVE-15339 at 12/2/16 11:12 AM:
-------------------------------------------------------------------
Attaching .1 patch.
{noformat}
Without any patch: (compile time 14.2 seconds)
==============
2016-12-02T05:20:34,867 DEBUG [cf8155ce-cf85-41b5-b0a3-4d4a6c75da5e main]
log.PerfLogger: </PERFLOG method=compile start=1480656020666 end=1480656034867
duration=14201 from=org.apache.hadoop.hive.ql.Driver>
With Patch: (compile time 10.6 seconds)
=========
2016-12-02T05:34:53,820 DEBUG [bfe87e40-4260-4f67-9e84-cd89694be1ad main]
log.PerfLogger: </PERFLOG method=compile start=1480656883196 end=1480656893820
duration=10624 from=org.apache.hadoop.hive.ql.Driver>
{noformat}
metastore DB was hosted in postgres and flights table has around 7000
partitions. Prefetch is a wrong term mentioned in the jira heading. Patch tries
to send all the needed columns in same call, and in other side these columns
stats get cached in AggregateColStats. Any col stats call fired later fetches
the data from the cache itself making it faster.
\cc [~pxiong], [~ashutoshc], [~jcamachorodriguez]
was (Author: rajesh.balamohan):
Attaching .1 patch.
{noformat}
Without any patch: (compile time 14.2 seconds)
==============
2016-12-02T05:20:34,867 DEBUG [cf8155ce-cf85-41b5-b0a3-4d4a6c75da5e main]
log.PerfLogger: </PERFLOG method=compile start=1480656020666 end=1480656034867
duration=14201 from=org.apache.hadoop.hive.ql.Driver>
With Patch: (compile time 10.6 seconds)
=========
2016-12-02T05:34:53,820 DEBUG [bfe87e40-4260-4f67-9e84-cd89694be1ad main]
log.PerfLogger: </PERFLOG method=compile start=1480656883196 end=1480656893820
duration=10624 from=org.apache.hadoop.hive.ql.Driver>
{noformat}
metastore DB was hosted in postgres and flights table has around 7000
partitions. Prefetch is a wrong term in the jira. Patch tries to send all the
needed columns in same call, and in other side these columns stats get cached
in AggregateColStats. Any col stats call fired later fetches the data from the
cache itself making it faster.
\cc [~pxiong], [~ashutoshc], [~jcamachorodriguez]
> Prefetch column stats for fields needed in FilterSelectivityEstimator
> ---------------------------------------------------------------------
>
> Key: HIVE-15339
> URL: https://issues.apache.org/jira/browse/HIVE-15339
> Project: Hive
> Issue Type: Improvement
> Reporter: Rajesh Balamohan
> Priority: Minor
> Attachments: HIVE-15339.1.patch
>
>
> Based on query pattern, {{FilterSelectivityEstimator}} gets column statistics
> from metastore in multiple calls. For instance, in the following query, it
> ends up getting individual column statistics for for flights multiple number
> of times.
> When the table has large number of partitions, getting statistics for columns
> via multiple calls can be very expensive. This would adversely impact the
> overall compilation time. The following query took 14 seconds to compile.
> {noformat}
> SELECT COUNT(`flights`.`flightnum`) AS `cnt_flightnum_ok`,
> YEAR(`flights`.`dateofflight`) AS `yr_flightdate_ok`
> FROM `flights` as `flights`
> JOIN `airlines` ON (`flights`.`uniquecarrier` = `airlines`.`code`)
> JOIN `airports` as `source_airport` ON (`flights`.`origin` =
> `source_airport`.`iata`)
> JOIN `airports` as `dest_airport` ON (`flights`.`dest` =
> `dest_airport`.`iata`)
> GROUP BY YEAR(`flights`.`dateofflight`);
> {noformat}
> It may be helpful to club all columns that need statistics and fetch these
> details in single remote call.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
