hive-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mich Talebzadeh <mich.talebza...@cloudtechnologypartners.co.uk>
Subject Re: Hive 2 performance
Date Thu, 25 Feb 2016 10:15:08 GMT
 

hanks Gopal I made the following observation so far: 

Using the old MR you get this message now which is fine 

Hive-on-MR is deprecated in Hive 2 and may not be available in the
future versions. Consider using a different execution engine (i.e. tez,
spark) or using Hive 1.X releases. 

use oraclehadoop;
--set hive.execution.engine=spark;
set hive.execution.engine=mr;
--
-- Get the total amount sold for each calendar month
-- 

select from_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss') AS
StartTime; 

CREATE TEMPORARY TABLE tmp AS
SELECT t.calendar_month_desc, c.channel_desc, SUM(s.amount_sold) AS
TotalSales
--FROM smallsales s, times t, channels c
FROM smallsales s, times t, channels c
WHERE s.time_id = t.time_id
AND s.channel_id = c.channel_id
GROUP BY t.calendar_month_desc, c.channel_desc
; 

select from_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss') AS
FirstQuery;
SELECT calendar_month_desc AS MONTH, channel_desc AS CHANNEL, TotalSales
from tmp
ORDER BY MONTH, CHANNEL LIMIT 5
;
select from_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss') AS
SecondQuery;
SELECT channel_desc AS CHANNEL, MAX(TotalSales) AS SALES
FROM tmp
GROUP BY channel_desc
order by SALES DESC LIMIT 5
;
select from_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss') AS
EndTime; 

This batch returns results on MR in 2 min, 3 seconds 

If I change my engine to Hive 2 on Spark 1.3.1. I get it back in 1 min,
9 sec 

If I run that job on Spark 1.5.2 shell against the same tables using
Functional programming and Hive Context for tables 

val HiveContext = new org.apache.spark.sql.hive.HiveContext(sc)
println ("nStarted at"); HiveContext.sql("SELECT
FROM_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss')
").collect.foreach(println)
HiveContext.sql("use oraclehadoop")
var s =
HiveContext.table("sales").select("AMOUNT_SOLD","TIME_ID","CHANNEL_ID")
val c =
HiveContext.table("channels").select("CHANNEL_ID","CHANNEL_DESC")
val t =
HiveContext.table("times").select("TIME_ID","CALENDAR_MONTH_DESC")
println ("ncreating data set at"); HiveContext.sql("SELECT
FROM_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss')
").collect.foreach(println)
val rs =
s.join(t,"time_id").join(c,"channel_id").groupBy("calendar_month_desc","channel_desc").agg(sum("amount_sold").as("TotalSales"))
println ("nfirst query at"); HiveContext.sql("SELECT
FROM_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss')
").collect.foreach(println)
val rs1 =
rs.orderBy("calendar_month_desc","channel_desc").take(5).foreach(println)
println ("nsecond query at"); HiveContext.sql("SELECT
FROM_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss')
").collect.foreach(println)
val rs2
=rs.groupBy("channel_desc").agg(max("TotalSales").as("SALES")).orderBy("SALES").sort(desc("SALES")).take(5).foreach(println)
println ("nFinished at"); HiveContext.sql("SELECT
FROM_unixtime(unix_timestamp(), 'dd/MM/yyyy HH:mm:ss.ss')
").collect.foreach(println) 

I get the job done in under 8 min. Ok this is not a benchmark for Spark
but shows that Hive 2 has improved significantly IMO. I also had Hive on
Spark 1.3.1 crashing on certain large tables(had to revert to MR) but no
issues now. 

HTH 

On 25/02/2016 09:13, Gopal Vijayaraghavan wrote: 

>> Correct hence the question as I have done some preliminary tests on Hive 2. I want
to share insights with other people who have performed the same
> 
> If you have feedback on Hive-2.0, I'm all ears.
> 
> I'm building up 2.1 features & fixes, so now would be a good time to bring
> stuff up.
> 
> Speed mostly depends on whether you're using Hive-2.0 with LLAP or not -
> if you're using the old engines, the plans still get much better (even for
> MR).
> 
> Tez does get some stuff out of it, like the new shuffle join vertex
> manager (hive.optimize.dynamic.partition.hashjoin).
> 
> LLAP will still win that out for <10s queries, because it takes approx ~10
> mins for all the auto-generated vectorized classes to get JIT'd into tight
> SIMD loops.
> 
> For something like TPC-H Q1, you can slowly see it turning all the null
> checks into UncommonTrapBlob as the JIT slowly learns about the data &
> finds .noNulls is always true.
> 
> Cheers,
> Gopal

-- 

Dr Mich Talebzadeh

LinkedIn
https://www.linkedin.com/profile/view?id=AAEAAAAWh2gBxianrbJd6zP6AcPCCdOABUrV8Pw

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 Cloud
Technology Partners 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 Cloud Technology
partners Ltd, its subsidiaries nor their employees accept any
responsibility.

 
Mime
View raw message