[
https://issues.apache.org/jira/browse/DRILL-4363?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15139864#comment-15139864
]
Jinfeng Ni commented on DRILL-4363:
-----------------------------------
Did some performance comparison with two datasets:
1. Dataset containing 300 parquet files, with total 46M rows. Each parquet
file has about 2000 columns. That's for wide table use case.
Without the patch, the query will crash drillbits in the cluster, since the
query is executed in multi minor fragments for Scan operator, and each minor
scan fragment will use around 500M ~ 1GB memory.
With the patch, the query completed in under 30 seconds, with warm cache.
2. Dataset containing 115 small parquet files. The file was created from TPCH
lineitem table.
Without patch,
{code}
select * from dfs.`/Users/jni/work/data/tpch-sf10/lineitem115k` limit 1;
1 row selected (34.165 seconds)
{code}
With patch
{code}
select * from dfs.`/Users/jni/work/data/tpch-sf10/lineitem115k` limit 1;
1 row selected (14.021 seconds)
{code}
Basically, it reduce from 34 seconds to 14 seconds with warm cache.
> Apply row count based pruning for parquet table in LIMIT n query
> ----------------------------------------------------------------
>
> Key: DRILL-4363
> URL: https://issues.apache.org/jira/browse/DRILL-4363
> Project: Apache Drill
> Issue Type: Improvement
> Reporter: Jinfeng Ni
> Assignee: Jinfeng Ni
> Fix For: 1.6.0
>
>
> In interactive data exploration use case, one common and probably first query
> that users would use is " SELECT * from table LIMIT n", where n is a small
> number. Such query will give user idea about the columns in the table.
> Normally, user would expect such query should be completed in very short
> time, since it's just asking for small amount of rows, without any
> sort/aggregation.
> When table is small, there is no big problem for Drill. However, when the
> table is extremely large, Drill's response time is not as fast as what user
> would expect.
> In case of parquet table, it seems that query planner could do a bit better
> job : by applying row count based pruning for such LIMIT n query. The
> pruning is kind of similar to what partition pruning will do, except that it
> uses row count, in stead of partition column values. Since row count is
> available in parquet table, it's possible to do such pruning.
> The benefit of doing such pruning is clear: 1) for small "n", such pruning
> would end up with a few parquet files, in stead of thousands, or millions of
> files to scan. 2) execution probably does not have to put scan into multiple
> minor fragments and start reading the files concurrently, which will cause
> big IO overhead. 3) the physical plan itself is much smaller, since it does
> not include the long list of parquet files, reduce rpc cost of sending the
> fragment plans to multiple drillbits, and the overhead to
> serialize/deserialize the fragment plans.
>
>
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
