[ 
https://issues.apache.org/jira/browse/HIVE-27924?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Krisztian Kasa reassigned HIVE-27924:
-------------------------------------

    Assignee: Krisztian Kasa

> Incremental rebuild goes wrong when inserts and deletes overlap between the 
> source tables
> -----------------------------------------------------------------------------------------
>
>                 Key: HIVE-27924
>                 URL: https://issues.apache.org/jira/browse/HIVE-27924
>             Project: Hive
>          Issue Type: Bug
>          Components: Materialized views
>    Affects Versions: 4.0.0-beta-1
>         Environment: * Docker version : 19.03.6
>  * Hive version : 4.0.0-beta-1
>  * Driver version : Hive JDBC (4.0.0-beta-1)
>  * Beeline version : 4.0.0-beta-1
>            Reporter: Wenhao Li
>            Assignee: Krisztian Kasa
>            Priority: Critical
>              Labels: bug, hive, materializedviews
>         Attachments: 截图.PNG, 截图1.PNG, 截图2.PNG, 截图3.PNG, 截图4.PNG, 截图5.PNG, 
> 截图6.PNG, 截图7.PNG, 截图8.PNG, 截图9.PNG
>
>
> h1. Summary
> The incremental rebuild plan and execution output are incorrect when one side 
> of the table join has inserted/deleted join keys that the other side has 
> deleted/inserted (note the order).
> The argument is that tuples that have never been present simultaneously 
> should not interact with one another, i.e., one's inserts should not join the 
> other's deletes.
> h1. Related Test Case
> The bug was discovered during replication of the test case:
> ??hive/ql/src/test/queries/clientpositive/materialized_view_create_rewrite_5.q??
> h1. Steps to Reproduce the Issue
>  # Configurations:
> {code:sql}
> SET hive.vectorized.execution.enabled=false;
> set hive.support.concurrency=true;
> set hive.txn.manager=org.apache.hadoop.hive.ql.lockmgr.DbTxnManager;
> set hive.strict.checks.cartesian.product=false;
> set hive.materializedview.rewriting=true;{code}
>  # 
> {code:sql}
> create table cmv_basetable_n6 (a int, b varchar(256), c decimal(10,2), d int) 
> stored as orc TBLPROPERTIES ('transactional'='true'); {code}
>  # 
> {code:sql}
> insert into cmv_basetable_n6 values
> (1, 'alfred', 10.30, 2),
> (1, 'charlie', 20.30, 2); {code}
>  # 
> {code:sql}
> create table cmv_basetable_2_n3 (a int, b varchar(256), c decimal(10,2), d 
> int) stored as orc TBLPROPERTIES ('transactional'='true'); {code}
>  # 
> {code:sql}
> insert into cmv_basetable_2_n3 values
> (1, 'bob', 30.30, 2),
> (1, 'bonnie', 40.30, 2);{code}
>  # 
> {code:sql}
> CREATE MATERIALIZED VIEW cmv_mat_view_n6 TBLPROPERTIES 
> ('transactional'='true') AS
> SELECT cmv_basetable_n6.a, cmv_basetable_2_n3.c
> FROM cmv_basetable_n6 JOIN cmv_basetable_2_n3 ON (cmv_basetable_n6.a = 
> cmv_basetable_2_n3.a)
> WHERE cmv_basetable_2_n3.c > 10.0;{code}
>  # 
> {code:sql}
> show tables; {code}
> !截图.PNG!
>  # Select tuples, including deletion and with VirtualColumn's, from the MV 
> and source tables. We see that the MV is correctly built upon creation:
> {code:sql}
> SELECT ROW__IS__DELETED, ROW__ID, * FROM 
> cmv_mat_view_n6('acid.fetch.deleted.rows'='true');{code}
> !截图1.PNG!
>  # 
> {code:sql}
> SELECT ROW__IS__DELETED, ROW__ID, * FROM 
> cmv_basetable_n6('acid.fetch.deleted.rows'='true'); {code}
> !截图2.PNG!
>  # 
> {code:sql}
> SELECT ROW__IS__DELETED, ROW__ID, * FROM 
> cmv_basetable_2_n3('acid.fetch.deleted.rows'='true'); {code}
> !截图3.PNG!
>  # Now make an insert to the LHS and a delete to the RHS source table:
> {code:sql}
> insert into cmv_basetable_n6 values
> (1, 'kevin', 50.30, 2);
> DELETE FROM cmv_basetable_2_n3 WHERE b = 'bonnie';{code}
>  # Select again to see what happened:
> {code:sql}
> SELECT ROW__IS__DELETED, ROW__ID, * FROM 
> cmv_basetable_n6('acid.fetch.deleted.rows'='true'); {code}
> !截图4.PNG!
>  # 
> {code:sql}
> SELECT ROW__IS__DELETED, ROW__ID, * FROM 
> cmv_basetable_2_n3('acid.fetch.deleted.rows'='true'); {code}
> !截图5.PNG!
>  # Use {{EXPLAIN CBO}} to produce the incremental rebuild plan for the MV, 
> which is incorrect already:
> {code:sql}
> EXPLAIN CBO
> ALTER MATERIALIZED VIEW cmv_mat_view_n6 REBUILD; {code}
> !截图6.PNG!
>  # Rebuild MV and see (incorrect) results:
> {code:sql}
> ALTER MATERIALIZED VIEW cmv_mat_view_n6 REBUILD;
> SELECT ROW__IS__DELETED, ROW__ID, * FROM 
> cmv_mat_view_n6('acid.fetch.deleted.rows'='true');{code}
> !截图7.PNG!
>  # Run MV definition directly, which outputs incorrect results because the MV 
> is enabled for MV-based query rewrite, i.e., the following query will output 
> what's in the MV for the time being:
> {code:sql}
> SELECT cmv_basetable_n6.a, cmv_basetable_2_n3.c
> FROM cmv_basetable_n6 JOIN cmv_basetable_2_n3 ON (cmv_basetable_n6.a = 
> cmv_basetable_2_n3.a)
> WHERE cmv_basetable_2_n3.c > 10.0; {code}
> !截图8.PNG!
>  # Disable MV-based query rewrite for the MV and re-run the definition, which 
> should give the correct results:
> {code:sql}
> ALTER MATERIALIZED VIEW cmv_mat_view_n6 DISABLE REWRITE;
> SELECT cmv_basetable_n6.a, cmv_basetable_2_n3.c
> FROM cmv_basetable_n6 JOIN cmv_basetable_2_n3 ON (cmv_basetable_n6.a = 
> cmv_basetable_2_n3.a)
> WHERE cmv_basetable_2_n3.c > 10.0;{code}
> !截图9.PNG!
> h1. Note
> This issue is also seen in update-incurred inserts/deletes.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to