[
https://issues.apache.org/jira/browse/DRILL-4589?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15229298#comment-15229298
]
Jinfeng Ni commented on DRILL-4589:
-----------------------------------
I have a patch for this JIRA. Using the same dataset used in the comparison
done in DRILL-2517(With 115k parquet files in total, it's organized in 25
directories (1990, 1991, ... ), and each directory has four subdirectories (Q1,
Q2, Q3, Q4).), here is the query planning time measured on a mac laptop.
{code}
explain plan for select * from dfs.`/drill/testdata/tpch-sf10/lineitem115k`
where dir0 = '1990' and dir1 = 'Q1';
{code}
Without the patch (on today's master branch:
{code}
1 row selected (8.084 seconds)
{code}
With the patch
{code}
1 row selected (4.306 seconds)
{code}
If the partition filter contains complex expression, then the improvement
percentage is even higher. For this query, the improvement is 24.951 seconds
vs. 4.393 seconds
{code}
explain plan for select * from dfs.`/drill/testdata/tpch-sf10/lineitem115k`
where concat(substr(dir0, 1, 4), substr(dir1, 1, 2)) = '1990Q1';
{code}
> Reduce planning time for file system partition pruning by reducing filter
> evaluation overhead
> ---------------------------------------------------------------------------------------------
>
> Key: DRILL-4589
> URL: https://issues.apache.org/jira/browse/DRILL-4589
> Project: Apache Drill
> Issue Type: Bug
> Components: Query Planning & Optimization
> Reporter: Jinfeng Ni
> Assignee: Jinfeng Ni
>
> When Drill is used to query hundreds of thousands, or even millions of files
> organized into multi-level directories, user typically will provide a
> partition filter like : dir0 = something and dir1 = something2 and .. .
> For such queries, we saw the query planning time could be unacceptable long,
> due to three main overheads: 1) to expand and get the list of files, 2) to
> evaluate the partition filter, 3) to get the metadata, in the case of parquet
> files for which metadata cache file is not available.
> DRILL-2517 targets at the 3rd part of overhead. As a follow-up work after
> DRILL-2517, we plan to reduce the filter evaluation overhead. For now, the
> partition filter evaluation is applied to file level. In many cases, we saw
> that the number of leaf subdirectories is significantly lower than that of
> files. Since all the files under the same leaf subdirecctory share the same
> directory metadata, we should apply the filter evaluation at the leaf
> subdirectory. By doing that, we could reduce the cpu overhead to evaluate the
> filter, and the memory overhead as well.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
