You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@hive.apache.org by Mich Talebzadeh <mi...@peridale.co.uk> on 2015/12/01 17:40:38 UTC
Using spark in tandem with Hive
What if we decide to use spark with Hive. I look to hear similar views
My test bed comprised
1. Spark version 1.5.2
2. Hive version 1.2.1
3. Hadoop version 2.6
I made Spark to use Hive metastore. So using spark-sql I can pretty do
whatever one can do with HiveQL
I created and populated an ORC table in spark-sql.. It took 90 seconds to
create and populate the table with 1.7 million rows
spark-sql> select count(1) from tt;
1767886
Time taken: 5.169 seconds, Fetched 1 row(s)
Now let me try to do the said operation on the same table with HCL and MR
hive> use asehadoop;
OK
Time taken: 0.639 seconds
hive> select count(1) from tt;
Query ID = hduser_20151201162717_e3102633-f501-413b-b9cb-384ac50880ac
Total jobs = 1
Launching Job 1 out of 1
Number of reduce tasks determined at compile time: 1
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
set mapreduce.job.reduces=<number>
Starting Job = job_1448969636093_0001, Tracking URL =
http://rhes564:8088/proxy/application_1448969636093_0001/
Kill Command = /home/hduser/hadoop-2.6.0/bin/hadoop job -kill
job_1448969636093_0001
Hadoop job information for Stage-1: number of mappers: 1; number of
reducers: 1
2015-12-01 16:27:27,154 Stage-1 map = 0%, reduce = 0%
2015-12-01 16:27:35,427 Stage-1 map = 100%, reduce = 0%, Cumulative CPU
2.65 sec
2015-12-01 16:27:41,611 Stage-1 map = 100%, reduce = 100%, Cumulative CPU
3.71 sec
MapReduce Total cumulative CPU time: 3 seconds 710 msec
Ended Job = job_1448969636093_0001
MapReduce Jobs Launched:
Stage-Stage-1: Map: 1 Reduce: 1 Cumulative CPU: 3.71 sec HDFS Read:
520151 HDFS Write: 8 SUCCESS
Total MapReduce CPU Time Spent: 3 seconds 710 msec
OK
1767886
Time taken: 25.635 seconds, Fetched: 1 row(s)
So 5 seconds in Spark versus 25 seconds in Hive
On a point query Hive does not seem to return the correct timing?
hive> select * from tt where data_object_id = 10;
Time taken: 0.063 seconds, Fetched: 72 row(s)
Whereas in Spark I get
spark-sql> select * from tt where data_object_id = 10;
Time taken: 9.002 seconds, Fetched 72 row(s)
9 seconds looks far more plausible to me than 0.063 seonds. Or in an
unlikely event Spark returns elapsed time, whereas Hive returns execution
time?
Thanks
Mich Talebzadeh
Sybase ASE 15 Gold Medal Award 2008
A Winning Strategy: Running the most Critical Financial Data on ASE 15
<http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908
.pdf>
http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.
pdf
Author of the books "A Practitioner's Guide to Upgrading to Sybase ASE 15",
ISBN 978-0-9563693-0-7.
co-author "Sybase Transact SQL Guidelines Best Practices", ISBN
978-0-9759693-0-4
Publications due shortly:
Complex Event Processing in Heterogeneous Environments, ISBN:
978-0-9563693-3-8
Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume
one out shortly
<http://talebzadehmich.wordpress.com/> http://talebzadehmich.wordpress.com
NOTE: The information in this email is proprietary and confidential. This
message is for the designated recipient only, if you are not the intended
recipient, you should destroy it immediately. Any information in this
message shall not be understood as given or endorsed by Peridale Technology
Ltd, its subsidiaries or their employees, unless expressly so stated. It is
the responsibility of the recipient to ensure that this email is virus free,
therefore neither Peridale Ltd, its subsidiaries nor their employees accept
any responsibility.
RE: Using spark in tandem with Hive
Posted by Mich Talebzadeh <mi...@peridale.co.uk>.
Thanks.
My test bed has the following components.
1. Spark version 1.5.2
2. Hive version 1.2.1
3. Hadoop version 2.6
I will try your suggestions, however, we have to consider that the underlying table is based on a Hive table, to keep the systematics the same so to speak for comparison.
I increased the rows in table tt to 56 million and this is what a simple query compares 10 seconds in spark compared to 143 seconds with HIVE/MR.
In spark-sql
spark-sql> select count(1) from tt where object_id > 1000 and object_type = 'TABLE';
2251008
Time taken: 10.295 seconds, Fetched 1 row(s)
In hive with standard MR (no TEZ)
hive> select count(1) from tt where object_id > 1000 and object_type = 'TABLE';
Query ID = hduser_20151201202549_b698c002-6de0-4353-9a4a-3ba06e7c0428
Total jobs = 1
Launching Job 1 out of 1
Number of reduce tasks determined at compile time: 1
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
set mapreduce.job.reduces=<number>
Starting Job = job_1448969636093_0004, Tracking URL = http://rhes564:8088/proxy/application_1448969636093_0004/
Kill Command = /home/hduser/hadoop-2.6.0/bin/hadoop job -kill job_1448969636093_0004
Hadoop job information for Stage-1: number of mappers: 11; number of reducers: 1
2015-12-01 20:25:58,765 Stage-1 map = 0%, reduce = 0%
2015-12-01 20:26:08,140 Stage-1 map = 5%, reduce = 0%, Cumulative CPU 6.38 sec
2015-12-01 20:26:10,201 Stage-1 map = 9%, reduce = 0%, Cumulative CPU 7.5 sec
2015-12-01 20:26:19,539 Stage-1 map = 14%, reduce = 0%, Cumulative CPU 13.92 sec
2015-12-01 20:26:20,570 Stage-1 map = 18%, reduce = 0%, Cumulative CPU 15.07 sec
2015-12-01 20:26:30,870 Stage-1 map = 23%, reduce = 0%, Cumulative CPU 21.55 sec
2015-12-01 20:26:31,897 Stage-1 map = 27%, reduce = 0%, Cumulative CPU 22.7 sec
2015-12-01 20:26:42,191 Stage-1 map = 32%, reduce = 0%, Cumulative CPU 29.45 sec
2015-12-01 20:26:43,217 Stage-1 map = 36%, reduce = 0%, Cumulative CPU 30.41 sec
2015-12-01 20:26:52,465 Stage-1 map = 41%, reduce = 0%, Cumulative CPU 36.83 sec
2015-12-01 20:26:53,493 Stage-1 map = 45%, reduce = 0%, Cumulative CPU 37.78 sec
2015-12-01 20:27:04,778 Stage-1 map = 50%, reduce = 0%, Cumulative CPU 44.12 sec
2015-12-01 20:27:05,806 Stage-1 map = 55%, reduce = 0%, Cumulative CPU 45.1 sec
2015-12-01 20:27:17,126 Stage-1 map = 59%, reduce = 0%, Cumulative CPU 51.67 sec
2015-12-01 20:27:18,150 Stage-1 map = 64%, reduce = 0%, Cumulative CPU 52.75 sec
2015-12-01 20:27:28,424 Stage-1 map = 69%, reduce = 0%, Cumulative CPU 59.24 sec
2015-12-01 20:27:29,453 Stage-1 map = 73%, reduce = 0%, Cumulative CPU 60.2 sec
2015-12-01 20:27:40,805 Stage-1 map = 77%, reduce = 0%, Cumulative CPU 66.63 sec
2015-12-01 20:27:42,855 Stage-1 map = 82%, reduce = 0%, Cumulative CPU 68.53 sec
2015-12-01 20:27:54,156 Stage-1 map = 87%, reduce = 0%, Cumulative CPU 74.99 sec
2015-12-01 20:27:55,180 Stage-1 map = 91%, reduce = 0%, Cumulative CPU 76.15 sec
2015-12-01 20:28:05,483 Stage-1 map = 96%, reduce = 0%, Cumulative CPU 82.64 sec
2015-12-01 20:28:06,509 Stage-1 map = 100%, reduce = 0%, Cumulative CPU 83.4 sec
2015-12-01 20:28:10,622 Stage-1 map = 100%, reduce = 100%, Cumulative CPU 84.72 sec
MapReduce Total cumulative CPU time: 1 minutes 24 seconds 720 msec
Ended Job = job_1448969636093_0004
MapReduce Jobs Launched:
Stage-Stage-1: Map: 11 Reduce: 1 Cumulative CPU: 84.72 sec HDFS Read: 56718493 HDFS Write: 8 SUCCESS
Total MapReduce CPU Time Spent: 1 minutes 24 seconds 720 msec
OK
2251008
Time taken: 143.452 seconds, Fetched: 1 row(s)
Mich Talebzadeh
Sybase ASE 15 Gold Medal Award 2008
A Winning Strategy: Running the most Critical Financial Data on ASE 15
<http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf> http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf
Author of the books "A Practitioner’s Guide to Upgrading to Sybase ASE 15", ISBN 978-0-9563693-0-7.
co-author "Sybase Transact SQL Guidelines Best Practices", ISBN 978-0-9759693-0-4
Publications due shortly:
Complex Event Processing in Heterogeneous Environments, ISBN: 978-0-9563693-3-8
Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume one out shortly
<http://talebzadehmich.wordpress.com/> http://talebzadehmich.wordpress.com
NOTE: The information in this email is proprietary and confidential. This message is for the designated recipient only, if you are not the intended recipient, you should destroy it immediately. Any information in this message shall not be understood as given or endorsed by Peridale Technology Ltd, its subsidiaries or their employees, unless expressly so stated. It is the responsibility of the recipient to ensure that this email is virus free, therefore neither Peridale Ltd, its subsidiaries nor their employees accept any responsibility.
From: Jörn Franke [mailto:jornfranke@gmail.com]
Sent: 01 December 2015 20:38
To: user@hive.apache.org
Subject: Re: Using spark in tandem with Hive
You Should use TEZ (preferably >0.8 and a release of hive supporting it, because it has tez service which allows more lower latency queries) instead of mr to get the first query faster. The second query is probably faster in hive because you use statistics, which to my knowledge are not leveraged by Spark (only for broadcast joins).
However you can still create statistics for each column by executing the same command you use for the table, but adding for columns to it (probably also not supported by Spark). You may get faster access if you do not mark the table as transactional (I am not sure if spark can handle them properly). I do not know your data, but you should check if you want to sort the data on certain columns. Probably the bloom filter on object_id is not necessary because this is covered by the storage index. Bloom filters are only available in Hive >= 1.2. There might be further optimizations (eg partitioning, increasing replication etc), but this would require more knowledge of the data.
On 01 Dec 2015, at 19:20, Mich Talebzadeh <mich@peridale.co.uk <ma...@peridale.co.uk> > wrote:
The table was created in spark-sql as ORC table
use asehadoop;
drop table if exists tt;
create table tt (
owner varchar(30)
,object_name varchar(30)
,subobject_name varchar(30)
,object_id bigint
,data_object_id bigint
,object_type varchar(19)
,created timestamp
,last_ddl_time timestamp
,timestamp varchar(19)
,status varchar(7)
,temporary2 varchar(1)
,generated varchar(1)
,secondary varchar(1)
,namespace bigint
,edition_name varchar(30)
,padding1 varchar(4000)
,padding2 varchar(3500)
,attribute varchar(32)
,op_type int
,op_time timestamp
)
CLUSTERED BY (object_id) INTO 256 BUCKETS
STORED AS ORC
TBLPROPERTIES ( "orc.compress"="SNAPPY",
"transactional"="true",
"orc.create.index"="true",
"orc.bloom.filter.columns"="object_id",
"orc.bloom.filter.fpp"="0.05",
"orc.stripe.size"="268435456",
"orc.row.index.stride"="10000" )
;
show create table tt;
INSERT INTO TABLE tt
SELECT
owner
, object_name
, subobject_name
, object_id
, data_object_id
, object_type
, cast(created AS timestamp)
, cast(last_ddl_time AS timestamp)
, timestamp
, status
, temporary2
, generated
, secondary
, namespace
, edition_name
, padding1
, padding2
, attribute
, 1
, cast(from_unixtime(unix_timestamp()) AS timestamp)
FROM t_staging
;
And it was analysed ass below
hive> analyze table tt compute statistics;
Table asehadoop.tt stats: [numFiles=30, numRows=1767886, totalSize=88388380, rawDataSize=5984968162]
OK
Time taken: 0.241 seconds
HTH
Mich Talebzadeh
Sybase ASE 15 Gold Medal Award 2008
A Winning Strategy: Running the most Critical Financial Data on ASE 15
<http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf> http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf
Author of the books "A Practitioner’s Guide to Upgrading to Sybase ASE 15", ISBN 978-0-9563693-0-7.
co-author "Sybase Transact SQL Guidelines Best Practices", ISBN 978-0-9759693-0-4
Publications due shortly:
Complex Event Processing in Heterogeneous Environments, ISBN: 978-0-9563693-3-8
Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume one out shortly
<http://talebzadehmich.wordpress.com/> http://talebzadehmich.wordpress.com
NOTE: The information in this email is proprietary and confidential. This message is for the designated recipient only, if you are not the intended recipient, you should destroy it immediately. Any information in this message shall not be understood as given or endorsed by Peridale Technology Ltd, its subsidiaries or their employees, unless expressly so stated. It is the responsibility of the recipient to ensure that this email is virus free, therefore neither Peridale Ltd, its subsidiaries nor their employees accept any responsibility.
From: Jörn Franke [mailto:jornfranke@gmail.com]
Sent: 01 December 2015 16:58
To: user@hive.apache.org <ma...@hive.apache.org>
Subject: Re: Using spark in tandem with Hive
How did you create the tables? Do you have automated statistics activated in Hive?
Btw mr is outdated as a Hive execution engine. Use TEZ (maybe wait for 0.8 for sub second queries ) or use Spark as an execution engine in Hive.
On 01 Dec 2015, at 17:40, Mich Talebzadeh <mich@peridale.co.uk <ma...@peridale.co.uk> > wrote:
What if we decide to use spark with Hive. I look to hear similar views
My test bed comprised
1. Spark version 1.5.2
2. Hive version 1.2.1
3. Hadoop version 2.6
I made Spark to use Hive metastore. So using spark-sql I can pretty do whatever one can do with HiveQL
I created and populated an ORC table in spark-sql.. It took 90 seconds to create and populate the table with 1.7 million rows
spark-sql> select count(1) from tt;
1767886
Time taken: 5.169 seconds, Fetched 1 row(s)
Now let me try to do the said operation on the same table with HCL and MR
hive> use asehadoop;
OK
Time taken: 0.639 seconds
hive> select count(1) from tt;
Query ID = hduser_20151201162717_e3102633-f501-413b-b9cb-384ac50880ac
Total jobs = 1
Launching Job 1 out of 1
Number of reduce tasks determined at compile time: 1
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
set mapreduce.job.reduces=<number>
Starting Job = job_1448969636093_0001, Tracking URL = http://rhes564:8088/proxy/application_1448969636093_0001/
Kill Command = /home/hduser/hadoop-2.6.0/bin/hadoop job -kill job_1448969636093_0001
Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 1
2015-12-01 16:27:27,154 Stage-1 map = 0%, reduce = 0%
2015-12-01 16:27:35,427 Stage-1 map = 100%, reduce = 0%, Cumulative CPU 2.65 sec
2015-12-01 16:27:41,611 Stage-1 map = 100%, reduce = 100%, Cumulative CPU 3.71 sec
MapReduce Total cumulative CPU time: 3 seconds 710 msec
Ended Job = job_1448969636093_0001
MapReduce Jobs Launched:
Stage-Stage-1: Map: 1 Reduce: 1 Cumulative CPU: 3.71 sec HDFS Read: 520151 HDFS Write: 8 SUCCESS
Total MapReduce CPU Time Spent: 3 seconds 710 msec
OK
1767886
Time taken: 25.635 seconds, Fetched: 1 row(s)
So 5 seconds in Spark versus 25 seconds in Hive
On a point query Hive does not seem to return the correct timing?
hive> select * from tt where data_object_id = 10;
Time taken: 0.063 seconds, Fetched: 72 row(s)
Whereas in Spark I get
spark-sql> select * from tt where data_object_id = 10;
Time taken: 9.002 seconds, Fetched 72 row(s)
9 seconds looks far more plausible to me than 0.063 seonds. Or in an unlikely event Spark returns elapsed time, whereas Hive returns execution time?
Thanks
Mich Talebzadeh
Sybase ASE 15 Gold Medal Award 2008
A Winning Strategy: Running the most Critical Financial Data on ASE 15
<http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf> http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf
Author of the books "A Practitioner’s Guide to Upgrading to Sybase ASE 15", ISBN 978-0-9563693-0-7.
co-author "Sybase Transact SQL Guidelines Best Practices", ISBN 978-0-9759693-0-4
Publications due shortly:
Complex Event Processing in Heterogeneous Environments, ISBN: 978-0-9563693-3-8
Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume one out shortly
<http://talebzadehmich.wordpress.com/> http://talebzadehmich.wordpress.com
NOTE: The information in this email is proprietary and confidential. This message is for the designated recipient only, if you are not the intended recipient, you should destroy it immediately. Any information in this message shall not be understood as given or endorsed by Peridale Technology Ltd, its subsidiaries or their employees, unless expressly so stated. It is the responsibility of the recipient to ensure that this email is virus free, therefore neither Peridale Ltd, its subsidiaries nor their employees accept any responsibility.
Re: Using spark in tandem with Hive
Posted by Jörn Franke <jo...@gmail.com>.
You Should use TEZ (preferably >0.8 and a release of hive supporting it, because it has tez service which allows more lower latency queries) instead of mr to get the first query faster. The second query is probably faster in hive because you use statistics, which to my knowledge are not leveraged by Spark (only for broadcast joins).
However you can still create statistics for each column by executing the same command you use for the table, but adding for columns to it (probably also not supported by Spark). You may get faster access if you do not mark the table as transactional (I am not sure if spark can handle them properly). I do not know your data, but you should check if you want to sort the data on certain columns. Probably the bloom filter on object_id is not necessary because this is covered by the storage index. Bloom filters are only available in Hive >= 1.2. There might be further optimizations (eg partitioning, increasing replication etc), but this would require more knowledge of the data.
> On 01 Dec 2015, at 19:20, Mich Talebzadeh <mi...@peridale.co.uk> wrote:
>
> The table was created in spark-sql as ORC table
>
> use asehadoop;
> drop table if exists tt;
> create table tt (
> owner varchar(30)
> ,object_name varchar(30)
> ,subobject_name varchar(30)
> ,object_id bigint
> ,data_object_id bigint
> ,object_type varchar(19)
> ,created timestamp
> ,last_ddl_time timestamp
> ,timestamp varchar(19)
> ,status varchar(7)
> ,temporary2 varchar(1)
> ,generated varchar(1)
> ,secondary varchar(1)
> ,namespace bigint
> ,edition_name varchar(30)
> ,padding1 varchar(4000)
> ,padding2 varchar(3500)
> ,attribute varchar(32)
> ,op_type int
> ,op_time timestamp
> )
> CLUSTERED BY (object_id) INTO 256 BUCKETS
> STORED AS ORC
> TBLPROPERTIES ( "orc.compress"="SNAPPY",
> "transactional"="true",
> "orc.create.index"="true",
> "orc.bloom.filter.columns"="object_id",
> "orc.bloom.filter.fpp"="0.05",
> "orc.stripe.size"="268435456",
> "orc.row.index.stride"="10000" )
> ;
> show create table tt;
> INSERT INTO TABLE tt
> SELECT
> owner
> , object_name
> , subobject_name
> , object_id
> , data_object_id
> , object_type
> , cast(created AS timestamp)
> , cast(last_ddl_time AS timestamp)
> , timestamp
> , status
> , temporary2
> , generated
> , secondary
> , namespace
> , edition_name
> , padding1
> , padding2
> , attribute
> , 1
> , cast(from_unixtime(unix_timestamp()) AS timestamp)
> FROM t_staging
> ;
>
> And it was analysed ass below
>
> hive> analyze table tt compute statistics;
> Table asehadoop.tt stats: [numFiles=30, numRows=1767886, totalSize=88388380, rawDataSize=5984968162]
> OK
> Time taken: 0.241 seconds
>
> HTH
>
> Mich Talebzadeh
>
> Sybase ASE 15 Gold Medal Award 2008
> A Winning Strategy: Running the most Critical Financial Data on ASE 15
> http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf
> Author of the books "A Practitioner’s Guide to Upgrading to Sybase ASE 15", ISBN 978-0-9563693-0-7.
> co-author "Sybase Transact SQL Guidelines Best Practices", ISBN 978-0-9759693-0-4
> Publications due shortly:
> Complex Event Processing in Heterogeneous Environments, ISBN: 978-0-9563693-3-8
> Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume one out shortly
>
> http://talebzadehmich.wordpress.com
>
> NOTE: The information in this email is proprietary and confidential. This message is for the designated recipient only, if you are not the intended recipient, you should destroy it immediately. Any information in this message shall not be understood as given or endorsed by Peridale Technology Ltd, its subsidiaries or their employees, unless expressly so stated. It is the responsibility of the recipient to ensure that this email is virus free, therefore neither Peridale Ltd, its subsidiaries nor their employees accept any responsibility.
>
> From: Jörn Franke [mailto:jornfranke@gmail.com]
> Sent: 01 December 2015 16:58
> To: user@hive.apache.org
> Subject: Re: Using spark in tandem with Hive
>
> How did you create the tables? Do you have automated statistics activated in Hive?
>
> Btw mr is outdated as a Hive execution engine. Use TEZ (maybe wait for 0.8 for sub second queries ) or use Spark as an execution engine in Hive.
>
> On 01 Dec 2015, at 17:40, Mich Talebzadeh <mi...@peridale.co.uk> wrote:
>
> What if we decide to use spark with Hive. I look to hear similar views
>
> My test bed comprised
>
> 1. Spark version 1.5.2
> 2. Hive version 1.2.1
> 3. Hadoop version 2.6
>
>
> I made Spark to use Hive metastore. So using spark-sql I can pretty do whatever one can do with HiveQL
>
> I created and populated an ORC table in spark-sql.. It took 90 seconds to create and populate the table with 1.7 million rows
>
> spark-sql> select count(1) from tt;
> 1767886
> Time taken: 5.169 seconds, Fetched 1 row(s)
>
> Now let me try to do the said operation on the same table with HCL and MR
>
> hive> use asehadoop;
> OK
> Time taken: 0.639 seconds
> hive> select count(1) from tt;
> Query ID = hduser_20151201162717_e3102633-f501-413b-b9cb-384ac50880ac
> Total jobs = 1
> Launching Job 1 out of 1
> Number of reduce tasks determined at compile time: 1
> In order to change the average load for a reducer (in bytes):
> set hive.exec.reducers.bytes.per.reducer=<number>
> In order to limit the maximum number of reducers:
> set hive.exec.reducers.max=<number>
> In order to set a constant number of reducers:
> set mapreduce.job.reduces=<number>
> Starting Job = job_1448969636093_0001, Tracking URL = http://rhes564:8088/proxy/application_1448969636093_0001/
> Kill Command = /home/hduser/hadoop-2.6.0/bin/hadoop job -kill job_1448969636093_0001
> Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 1
> 2015-12-01 16:27:27,154 Stage-1 map = 0%, reduce = 0%
> 2015-12-01 16:27:35,427 Stage-1 map = 100%, reduce = 0%, Cumulative CPU 2.65 sec
> 2015-12-01 16:27:41,611 Stage-1 map = 100%, reduce = 100%, Cumulative CPU 3.71 sec
> MapReduce Total cumulative CPU time: 3 seconds 710 msec
> Ended Job = job_1448969636093_0001
> MapReduce Jobs Launched:
> Stage-Stage-1: Map: 1 Reduce: 1 Cumulative CPU: 3.71 sec HDFS Read: 520151 HDFS Write: 8 SUCCESS
> Total MapReduce CPU Time Spent: 3 seconds 710 msec
> OK
> 1767886
> Time taken: 25.635 seconds, Fetched: 1 row(s)
>
> So 5 seconds in Spark versus 25 seconds in Hive
>
> On a point query Hive does not seem to return the correct timing?
>
> hive> select * from tt where data_object_id = 10;
> Time taken: 0.063 seconds, Fetched: 72 row(s)
>
> Whereas in Spark I get
>
> spark-sql> select * from tt where data_object_id = 10;
> Time taken: 9.002 seconds, Fetched 72 row(s)
>
> 9 seconds looks far more plausible to me than 0.063 seonds. Or in an unlikely event Spark returns elapsed time, whereas Hive returns execution time?
>
> Thanks
>
> Mich Talebzadeh
>
> Sybase ASE 15 Gold Medal Award 2008
> A Winning Strategy: Running the most Critical Financial Data on ASE 15
> http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf
> Author of the books "A Practitioner’s Guide to Upgrading to Sybase ASE 15", ISBN 978-0-9563693-0-7.
> co-author "Sybase Transact SQL Guidelines Best Practices", ISBN 978-0-9759693-0-4
> Publications due shortly:
> Complex Event Processing in Heterogeneous Environments, ISBN: 978-0-9563693-3-8
> Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume one out shortly
>
> http://talebzadehmich.wordpress.com
>
> NOTE: The information in this email is proprietary and confidential. This message is for the designated recipient only, if you are not the intended recipient, you should destroy it immediately. Any information in this message shall not be understood as given or endorsed by Peridale Technology Ltd, its subsidiaries or their employees, unless expressly so stated. It is the responsibility of the recipient to ensure that this email is virus free, therefore neither Peridale Ltd, its subsidiaries nor their employees accept any responsibility.
>
RE: Using spark in tandem with Hive
Posted by Mich Talebzadeh <mi...@peridale.co.uk>.
The table was created in spark-sql as ORC table
use asehadoop;
drop table if exists tt;
create table tt (
owner varchar(30)
,object_name varchar(30)
,subobject_name varchar(30)
,object_id bigint
,data_object_id bigint
,object_type varchar(19)
,created timestamp
,last_ddl_time timestamp
,timestamp varchar(19)
,status varchar(7)
,temporary2 varchar(1)
,generated varchar(1)
,secondary varchar(1)
,namespace bigint
,edition_name varchar(30)
,padding1 varchar(4000)
,padding2 varchar(3500)
,attribute varchar(32)
,op_type int
,op_time timestamp
)
CLUSTERED BY (object_id) INTO 256 BUCKETS
STORED AS ORC
TBLPROPERTIES ( "orc.compress"="SNAPPY",
"transactional"="true",
"orc.create.index"="true",
"orc.bloom.filter.columns"="object_id",
"orc.bloom.filter.fpp"="0.05",
"orc.stripe.size"="268435456",
"orc.row.index.stride"="10000" )
;
show create table tt;
INSERT INTO TABLE tt
SELECT
owner
, object_name
, subobject_name
, object_id
, data_object_id
, object_type
, cast(created AS timestamp)
, cast(last_ddl_time AS timestamp)
, timestamp
, status
, temporary2
, generated
, secondary
, namespace
, edition_name
, padding1
, padding2
, attribute
, 1
, cast(from_unixtime(unix_timestamp()) AS timestamp)
FROM t_staging
;
And it was analysed ass below
hive> analyze table tt compute statistics;
Table asehadoop.tt stats: [numFiles=30, numRows=1767886, totalSize=88388380, rawDataSize=5984968162]
OK
Time taken: 0.241 seconds
HTH
Mich Talebzadeh
Sybase ASE 15 Gold Medal Award 2008
A Winning Strategy: Running the most Critical Financial Data on ASE 15
<http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf> http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf
Author of the books "A Practitioner’s Guide to Upgrading to Sybase ASE 15", ISBN 978-0-9563693-0-7.
co-author "Sybase Transact SQL Guidelines Best Practices", ISBN 978-0-9759693-0-4
Publications due shortly:
Complex Event Processing in Heterogeneous Environments, ISBN: 978-0-9563693-3-8
Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume one out shortly
<http://talebzadehmich.wordpress.com/> http://talebzadehmich.wordpress.com
NOTE: The information in this email is proprietary and confidential. This message is for the designated recipient only, if you are not the intended recipient, you should destroy it immediately. Any information in this message shall not be understood as given or endorsed by Peridale Technology Ltd, its subsidiaries or their employees, unless expressly so stated. It is the responsibility of the recipient to ensure that this email is virus free, therefore neither Peridale Ltd, its subsidiaries nor their employees accept any responsibility.
From: Jörn Franke [mailto:jornfranke@gmail.com]
Sent: 01 December 2015 16:58
To: user@hive.apache.org
Subject: Re: Using spark in tandem with Hive
How did you create the tables? Do you have automated statistics activated in Hive?
Btw mr is outdated as a Hive execution engine. Use TEZ (maybe wait for 0.8 for sub second queries ) or use Spark as an execution engine in Hive.
On 01 Dec 2015, at 17:40, Mich Talebzadeh <mich@peridale.co.uk <ma...@peridale.co.uk> > wrote:
What if we decide to use spark with Hive. I look to hear similar views
My test bed comprised
1. Spark version 1.5.2
2. Hive version 1.2.1
3. Hadoop version 2.6
I made Spark to use Hive metastore. So using spark-sql I can pretty do whatever one can do with HiveQL
I created and populated an ORC table in spark-sql.. It took 90 seconds to create and populate the table with 1.7 million rows
spark-sql> select count(1) from tt;
1767886
Time taken: 5.169 seconds, Fetched 1 row(s)
Now let me try to do the said operation on the same table with HCL and MR
hive> use asehadoop;
OK
Time taken: 0.639 seconds
hive> select count(1) from tt;
Query ID = hduser_20151201162717_e3102633-f501-413b-b9cb-384ac50880ac
Total jobs = 1
Launching Job 1 out of 1
Number of reduce tasks determined at compile time: 1
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
set mapreduce.job.reduces=<number>
Starting Job = job_1448969636093_0001, Tracking URL = http://rhes564:8088/proxy/application_1448969636093_0001/
Kill Command = /home/hduser/hadoop-2.6.0/bin/hadoop job -kill job_1448969636093_0001
Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 1
2015-12-01 16:27:27,154 Stage-1 map = 0%, reduce = 0%
2015-12-01 16:27:35,427 Stage-1 map = 100%, reduce = 0%, Cumulative CPU 2.65 sec
2015-12-01 16:27:41,611 Stage-1 map = 100%, reduce = 100%, Cumulative CPU 3.71 sec
MapReduce Total cumulative CPU time: 3 seconds 710 msec
Ended Job = job_1448969636093_0001
MapReduce Jobs Launched:
Stage-Stage-1: Map: 1 Reduce: 1 Cumulative CPU: 3.71 sec HDFS Read: 520151 HDFS Write: 8 SUCCESS
Total MapReduce CPU Time Spent: 3 seconds 710 msec
OK
1767886
Time taken: 25.635 seconds, Fetched: 1 row(s)
So 5 seconds in Spark versus 25 seconds in Hive
On a point query Hive does not seem to return the correct timing?
hive> select * from tt where data_object_id = 10;
Time taken: 0.063 seconds, Fetched: 72 row(s)
Whereas in Spark I get
spark-sql> select * from tt where data_object_id = 10;
Time taken: 9.002 seconds, Fetched 72 row(s)
9 seconds looks far more plausible to me than 0.063 seonds. Or in an unlikely event Spark returns elapsed time, whereas Hive returns execution time?
Thanks
Mich Talebzadeh
Sybase ASE 15 Gold Medal Award 2008
A Winning Strategy: Running the most Critical Financial Data on ASE 15
<http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf> http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf
Author of the books "A Practitioner’s Guide to Upgrading to Sybase ASE 15", ISBN 978-0-9563693-0-7.
co-author "Sybase Transact SQL Guidelines Best Practices", ISBN 978-0-9759693-0-4
Publications due shortly:
Complex Event Processing in Heterogeneous Environments, ISBN: 978-0-9563693-3-8
Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume one out shortly
<http://talebzadehmich.wordpress.com/> http://talebzadehmich.wordpress.com
NOTE: The information in this email is proprietary and confidential. This message is for the designated recipient only, if you are not the intended recipient, you should destroy it immediately. Any information in this message shall not be understood as given or endorsed by Peridale Technology Ltd, its subsidiaries or their employees, unless expressly so stated. It is the responsibility of the recipient to ensure that this email is virus free, therefore neither Peridale Ltd, its subsidiaries nor their employees accept any responsibility.
Re: Using spark in tandem with Hive
Posted by Jörn Franke <jo...@gmail.com>.
How did you create the tables? Do you have automated statistics activated in Hive?
Btw mr is outdated as a Hive execution engine. Use TEZ (maybe wait for 0.8 for sub second queries ) or use Spark as an execution engine in Hive.
> On 01 Dec 2015, at 17:40, Mich Talebzadeh <mi...@peridale.co.uk> wrote:
>
> What if we decide to use spark with Hive. I look to hear similar views
>
> My test bed comprised
>
> 1. Spark version 1.5.2
> 2. Hive version 1.2.1
> 3. Hadoop version 2.6
>
>
> I made Spark to use Hive metastore. So using spark-sql I can pretty do whatever one can do with HiveQL
>
> I created and populated an ORC table in spark-sql.. It took 90 seconds to create and populate the table with 1.7 million rows
>
> spark-sql> select count(1) from tt;
> 1767886
> Time taken: 5.169 seconds, Fetched 1 row(s)
>
> Now let me try to do the said operation on the same table with HCL and MR
>
> hive> use asehadoop;
> OK
> Time taken: 0.639 seconds
> hive> select count(1) from tt;
> Query ID = hduser_20151201162717_e3102633-f501-413b-b9cb-384ac50880ac
> Total jobs = 1
> Launching Job 1 out of 1
> Number of reduce tasks determined at compile time: 1
> In order to change the average load for a reducer (in bytes):
> set hive.exec.reducers.bytes.per.reducer=<number>
> In order to limit the maximum number of reducers:
> set hive.exec.reducers.max=<number>
> In order to set a constant number of reducers:
> set mapreduce.job.reduces=<number>
> Starting Job = job_1448969636093_0001, Tracking URL = http://rhes564:8088/proxy/application_1448969636093_0001/
> Kill Command = /home/hduser/hadoop-2.6.0/bin/hadoop job -kill job_1448969636093_0001
> Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 1
> 2015-12-01 16:27:27,154 Stage-1 map = 0%, reduce = 0%
> 2015-12-01 16:27:35,427 Stage-1 map = 100%, reduce = 0%, Cumulative CPU 2.65 sec
> 2015-12-01 16:27:41,611 Stage-1 map = 100%, reduce = 100%, Cumulative CPU 3.71 sec
> MapReduce Total cumulative CPU time: 3 seconds 710 msec
> Ended Job = job_1448969636093_0001
> MapReduce Jobs Launched:
> Stage-Stage-1: Map: 1 Reduce: 1 Cumulative CPU: 3.71 sec HDFS Read: 520151 HDFS Write: 8 SUCCESS
> Total MapReduce CPU Time Spent: 3 seconds 710 msec
> OK
> 1767886
> Time taken: 25.635 seconds, Fetched: 1 row(s)
>
> So 5 seconds in Spark versus 25 seconds in Hive
>
> On a point query Hive does not seem to return the correct timing?
>
> hive> select * from tt where data_object_id = 10;
> Time taken: 0.063 seconds, Fetched: 72 row(s)
>
> Whereas in Spark I get
>
> spark-sql> select * from tt where data_object_id = 10;
> Time taken: 9.002 seconds, Fetched 72 row(s)
>
> 9 seconds looks far more plausible to me than 0.063 seonds. Or in an unlikely event Spark returns elapsed time, whereas Hive returns execution time?
>
> Thanks
>
> Mich Talebzadeh
>
> Sybase ASE 15 Gold Medal Award 2008
> A Winning Strategy: Running the most Critical Financial Data on ASE 15
> http://login.sybase.com/files/Product_Overviews/ASE-Winning-Strategy-091908.pdf
> Author of the books "A Practitioner’s Guide to Upgrading to Sybase ASE 15", ISBN 978-0-9563693-0-7.
> co-author "Sybase Transact SQL Guidelines Best Practices", ISBN 978-0-9759693-0-4
> Publications due shortly:
> Complex Event Processing in Heterogeneous Environments, ISBN: 978-0-9563693-3-8
> Oracle and Sybase, Concepts and Contrasts, ISBN: 978-0-9563693-1-4, volume one out shortly
>
> http://talebzadehmich.wordpress.com
>
> NOTE: The information in this email is proprietary and confidential. This message is for the designated recipient only, if you are not the intended recipient, you should destroy it immediately. Any information in this message shall not be understood as given or endorsed by Peridale Technology Ltd, its subsidiaries or their employees, unless expressly so stated. It is the responsibility of the recipient to ensure that this email is virus free, therefore neither Peridale Ltd, its subsidiaries nor their employees accept any responsibility.
>