[
https://issues.apache.org/jira/browse/IMPALA-9744?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Tim Armstrong updated IMPALA-9744:
----------------------------------
Summary: Treat corrupt table stats as missing to avoid bad plans (was:
Treat corrupt table stats as )
> Treat corrupt table stats as missing to avoid bad plans
> -------------------------------------------------------
>
> Key: IMPALA-9744
> URL: https://issues.apache.org/jira/browse/IMPALA-9744
> Project: IMPALA
> Issue Type: Bug
> Components: Frontend
> Reporter: Tim Armstrong
> Priority: Major
> Labels: ramp-up
>
> We currently detect corrupt stats (0 rows but data in partition) but only
> flag it. The 0 row count is used for planning. I ran into a scenario where
> this lead to an extremely pathological plan - the 0 row count lead to
> flipping a nested loop join to put the big table on the build side and
> running out of memory.
> I propose doing something very conservative to avoid this scenario: if we see
> corrupt stats in any partition, and the row count is computed to be zero,
> ignore the row count and treat it the same as missing stats in the planner.
> Here's an example where we end up with corrupt stats. Warning: this can
> remove the data file from your alltypes type, I recommend copying the file to
> a different location before running this.
> {noformat}
> # In beeline against HS2
> !connect jdbc:hive2://localhost:11050 hive org.apache.hive.jdbc.HiveDrive
> set hive.stats.autogather=true;
> CREATE TABLE `alltypes_insert_only`(
> `id` int COMMENT 'Add a comment',
> `bool_col` boolean,
> `tinyint_col` tinyint,
> `smallint_col` smallint,
> `int_col` int,
> `bigint_col` bigint,
> `float_col` float,
> `double_col` double,
> `date_string_col` string,
> `string_col` string,
> `timestamp_col` timestamp)
> PARTITIONED BY (
> `year` int,
> `month` int)
> STORED AS PARQUET
> TBLPROPERTIES ("transactional"="true",
> "transactional_properties"="insert_only");
> load data inpath
> 'hdfs://172.19.0.1:20500/test-warehouse/alltypes_parquet/year=2009/month=1/154473eafa08ea0e-f9d70e7100000004_1040780996_data.0.parq'
> into table alltypes_insert_only partition (year=2009,month=9);
> # In Impala
> show table stats alltypes_insert_only;
> +-------+-------+-------+--------+--------+--------------+-------------------+---------+-------------------+----------------------------------------------------------------------------------------+
> | year | month | #Rows | #Files | Size | Bytes Cached | Cache Replication
> | Format | Incremental stats | Location
> |
> +-------+-------+-------+--------+--------+--------------+-------------------+---------+-------------------+----------------------------------------------------------------------------------------+
> | 2009 | 10 | 0 | 1 | 7.75KB | NOT CACHED | NOT CACHED
> | PARQUET | false |
> hdfs://172.19.0.1:20500/test-warehouse/managed/alltypes_insert_only/year=2009/month=10
> |
> | Total | | -1 | 1 | 7.75KB | 0B |
> | | |
> |
> +-------+-------+-------+--------+--------+--------------+-------------------+---------+-------------------+----------------------------------------------------------------------------------------+
> {noformat}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
