[HIVE-23716] Support Anti Join in Hive

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.0.0-alpha-1
• Labels: pull-request-available
• Assignee: mahesh kumar behera
• Reporter: mahesh kumar behera
• Created: 17/Jun/20 18:08
• Updated: 27/Feb/24 22:23
• Resolved: 08/Aug/20 01:23

Description

Currently hive does not support Anti join. The query for anti join is converted to left outer join and null filter on right side join key is added to get the desired result. This is causing

1. Extra computation — The left outer join projects the redundant columns from right side. Along with that, filtering is done to remove the redundant rows. This is can be avoided in case of anti join as anti join will project only the required columns and rows from the left side table.
2. Extra shuffle — In case of anti join the duplicate records moved to join node can be avoided from the child node. This can reduce significant amount of data movement if the number of distinct rows( join keys) is significant.
3. Extra Memory Usage - In case of map based anti join , hash set is sufficient as just the key is required to check  if the records matches the join condition. In case of left join, we need the key and the non key columns also and thus a hash table will be required.

For a query like

 select wr_order_number FROM web_returns LEFT JOIN web_sales  ON wr_order_number = ws_order_number WHERE ws_order_number IS NULL;

The number of distinct ws_order_number in web_sales table in a typical 10TB TPCDS set up is just 10% of total records. So when we convert this query to anti join, instead of 7 billion rows, only 600 million rows are moved to join node.

In the current patch, just one conversion is done. The pattern of project->filter->left-join is converted to project->anti-join. This will take care of sub queries with “not exists” clause. The queries with “not exists” are converted first to filter + left-join and then its converted to anti join. The queries with “not in” are not handled in the current patch.

From execution side, both merge join and map join with vectorized execution  is supported for anti join.

Issue Links

• causes: HIVE-26659 TPC-DS query 16, 69, 94 return wrong results.
• causes: HIVE-26135 Invalid Anti join conversion may cause missing results
• is related to: HIVE-25953 Drop HiveRelMdPredicates::getPredicates(Join...) to use that of RelMdPredicates
• relates to: HIVE-23903 Support "not-in" for bloom filter
• relates to: HIVE-23904 Update TestMapJoinOperator for adding anti join test cases.
• relates to: HIVE-23905 Remove duplicate code in vector map join execution for Anti join and Semi Join.
• relates to: HIVE-23906 Analyze and implement PK-FK based optimization for Anti join
• relates to: HIVE-23907 Hash table type should be considered for calculating the Map join table size
• relates to: HIVE-23919 Merge all kind of Join operator variants (Semi, Anti, Normal) into one.
• relates to: HIVE-23920 Need to handle HiveJoinConstraintsRule for Anti Join
• relates to: HIVE-23921 Support HiveJoinProjectTransposeRule for Anti Join
• relates to: HIVE-23928 Support conversion of not-exists to Anti join directly
• relates to: HIVE-23991 Support isAlwaysTrue for Anti Join
• relates to: HIVE-23992 Support null filter within or clause for Anti Join
• relates to: HIVE-24013 Move anti join conversion after join reordering rule
• links to: GitHub Pull Request #1147