[
https://issues.apache.org/jira/browse/HIVE-16972?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Chaozhong Yang updated HIVE-16972:
----------------------------------
Description:
* Background
We can describe the basic work flow of common HQL query as follows:
1. compile and execute
2. fetch results
In many cases, we don't need to worry about the issues fetching results from
HDFS(iff there are mapreduce jobs generated in planning step). However, the
number of results files on HDFS and data distribution will affect the final
status of HQL query, especially for HiveServer2. We have some map-only queries,
e.g:
{code:sql}
select * from myTable where date > '20170101' and date <= '20170301' and id =
88 and type=99;
{code}
This query will generate more than 20,000 files(look at screenshot image
uploaded) on HDFS and most of those files are empty. Of course, they are very
sparse. If we send TFetchResultsRequest from HiveServer2 client with some
parameters(timeout: 90s, maxRows: 1024) , FetchOperator can not fetch 1024 rows
in 90 seconds and our HiveServer2 client will mark this TFetchResultsRequest as
timed out failure. Why? In fact, It's expensive to fetch results from empty
file. In our HDFS cluster( 5000+ DataNodes) , reading data from an empty file
will cost almost 100 ms (100ms * 1000 ==> 100s > 90s timeout). Obviously, we
can filter out those empty files or splits to speed up the process of
FetchResults.
was:
* Background
We can describe the basic work flow of common HQL query as follows:
1. compile and execute
2. fetch results
In many cases, we don't need to worry about the issues fetching results from
HDFS(iff there are mapreduce jobs generated in planning step). However, the
number of results files on HDFS and data distribution will affect the final
status of HQL query, especially for HiveServer2. We have some map-only queries,
e.g:
{code:sql}
select * from myTable where date > '20170201' and date <= '20170301' and id =
88;
{code}
This query will generate more than 10,000 files on HDFS and most of those
files are empty. Of course, they are very sparse. If we send
TFetchResultsRequest from HiveServer2 client with some parameters(timeout:
90s, maxRows: 1024) , FetchOperator can not fetch 1024 rows in 90 seconds and
our HiveServer2 client will mark this TFetchResultsRequest as timed out
failure. Why? In fact, It's expensive to fetch results from empty file. In our
HDFS cluster( 5000+ DataNodes) , reading data from an empty file will cost
almost 100 ms (100ms * 1000 ==> 100s > 90s timeout). Obviously, we can filter
out those empty files or splits to speed up the process of FetchResults.
> FetchOperator: filter out inputSplits which length is zero
> ----------------------------------------------------------
>
> Key: HIVE-16972
> URL: https://issues.apache.org/jira/browse/HIVE-16972
> Project: Hive
> Issue Type: Improvement
> Components: HiveServer2, Physical Optimizer, Query Planning
> Affects Versions: 2.1.0, 2.1.1
> Reporter: Chaozhong Yang
> Assignee: Chaozhong Yang
> Fix For: 2.1.2
>
> Attachments: HIVE-16972.patch, screenshot-1.png
>
>
> * Background
> We can describe the basic work flow of common HQL query as follows:
> 1. compile and execute
> 2. fetch results
> In many cases, we don't need to worry about the issues fetching results
> from HDFS(iff there are mapreduce jobs generated in planning step). However,
> the number of results files on HDFS and data distribution will affect the
> final status of HQL query, especially for HiveServer2. We have some map-only
> queries, e.g:
> {code:sql}
> select * from myTable where date > '20170101' and date <= '20170301' and id =
> 88 and type=99;
> {code}
> This query will generate more than 20,000 files(look at screenshot image
> uploaded) on HDFS and most of those files are empty. Of course, they are very
> sparse. If we send TFetchResultsRequest from HiveServer2 client with some
> parameters(timeout: 90s, maxRows: 1024) , FetchOperator can not fetch 1024
> rows in 90 seconds and our HiveServer2 client will mark this
> TFetchResultsRequest as timed out failure. Why? In fact, It's expensive to
> fetch results from empty file. In our HDFS cluster( 5000+ DataNodes) ,
> reading data from an empty file will cost almost 100 ms (100ms * 1000 ==>
> 100s > 90s timeout). Obviously, we can filter out those empty files or splits
> to speed up the process of FetchResults.
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
