spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Zhan Zhang <zzh...@hortonworks.com>
Subject Re: sql query orc slow
Date Fri, 09 Oct 2015 17:10:48 GMT
In your case, you manually set an AND pushdown, and the predicate is right based on your setting,
: leaf-0 = (EQUALS x 320)

The right way is to enable the predicate pushdown as follows.
sqlContext.setConf("spark.sql.orc.filterPushdown", "true”)

Thanks.

Zhan Zhang







On Oct 9, 2015, at 9:58 AM, patcharee <Patcharee.Thongtra@uni.no<mailto:Patcharee.Thongtra@uni.no>>
wrote:

Hi Zhan Zhang

Actually my query has WHERE clause "select date, month, year, hh, (u*0.9122461 - v*-0.40964267),
(v*0.9122461 + u*-0.40964267), z from 4D where x = 320 and y = 117 and zone == 2 and year=2009
and z >= 2 and z <= 8", column "x", "y" is not partition column, the others are partition
columns. I expected the system will use predicate pushdown. I turned on the debug and found
pushdown predicate was not generated ("DEBUG OrcInputFormat: No ORC pushdown predicate")

Then I tried to set the search argument explicitly (on the column "x" which is not partition
column)

   val xs = SearchArgumentFactory.newBuilder().startAnd().equals("x", 320).end().build()
   hiveContext.setConf("hive.io.file.readcolumn.names", "x")
   hiveContext.setConf("sarg.pushdown", xs.toKryo())

this time in the log pushdown predicate was generated but results was wrong (no results at
all)

15/10/09 18:36:06 INFO OrcInputFormat: ORC pushdown predicate: leaf-0 = (EQUALS x 320)
expr = leaf-0

Any ideas What wrong with this? Why the ORC pushdown predicate is not applied by the system?

BR,
Patcharee

On 09. okt. 2015 18:31, Zhan Zhang wrote:
Hi Patcharee,

>From the query, it looks like only the column pruning will be applied. Partition pruning
and predicate pushdown does not have effect. Do you see big IO difference between two methods?

The potential reason of the speed difference I can think of may be the different versions
of OrcInputFormat. The hive path may use NewOrcInputFormat, but the spark path use OrcInputFormat.

Thanks.

Zhan Zhang

On Oct 8, 2015, at 11:55 PM, patcharee <Patcharee.Thongtra@uni.no<mailto:Patcharee.Thongtra@uni.no>>
wrote:

Yes, the predicate pushdown is enabled, but still take longer time than the first method

BR,
Patcharee

On 08. okt. 2015 18:43, Zhan Zhang wrote:
Hi Patcharee,

Did you enable the predicate pushdown in the second method?

Thanks.

Zhan Zhang

On Oct 8, 2015, at 1:43 AM, patcharee <Patcharee.Thongtra@uni.no<mailto:Patcharee.Thongtra@uni.no>>
wrote:

Hi,

I am using spark sql 1.5 to query a hive table stored as partitioned orc file. We have the
total files is about 6000 files and each file size is about 245MB.

What is the difference between these two query methods below:

1. Using query on hive table directly

hiveContext.sql("select col1, col2 from table1")

2. Reading from orc file, register temp table and query from the temp table

val c = hiveContext.read.format("orc").load("/apps/hive/warehouse/table1")
c.registerTempTable("regTable")
hiveContext.sql("select col1, col2 from regTable")

When the number of files is large (query all from the total 6000 files) , the second case
is much slower then the first one. Any ideas why?

BR,




---------------------------------------------------------------------
To unsubscribe, e-mail: user-unsubscribe@spark.apache.org<mailto:user-unsubscribe@spark.apache.org>
For additional commands, e-mail: user-help@spark.apache.org<mailto:user-help@spark.apache.org>







Mime
View raw message