[ATLAS-2891] Incorrect column lineage: each output column has input from *all columns* of the input table

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 0.8.2
• Fix Version/s: 0.8.3, 1.2.0, 2.0.0
• Component/s: atlas-intg
• Assignee: Madhan Neethiraj
• Reporter: Madhan Neethiraj
• Created: 25/Sep/18 21:47
• Updated: 26/Sep/19 00:38
• Resolved: 26/Oct/18 14:03

Description

Column lineage generated by Atlas Hive hook is incorrect for certain queries - like the following INSERT:

CREATE TABLE source_tbl(col_001 INT, col_002 INT, col_003 INT);

CREATE TABLE target_tbl(col_001 INT, col_002 INT, col_003 INT);

INSERT INTO target_tbl SELECT v1.col_001, v1.col_002, v1.col_003 FROM (SELECT col_001, col_002, col_003, ROW_NUMBER() OVER() AS r_num FROM source_tbl) v1;

In this case, lineage for each column in target_tbl shows input from all columns in source_tbl. In this case, the lineage information provided to post hooks (like Atlas hook) contains 3 entries, one for each column in target_tbl. Note the dependency for each column has all columns of the source_tbl.

DependencyKey=default.target_tbl:FieldSchema(name:col_001, type:int, comment:null)
Dependency=[SCRIPT]
           [default.source_tbl(src):FieldSchema(name:col_001, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:col_002, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:col_003, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:BLOCK__OFFSET__INSIDE__FILE, type:bigint, comment:),
            default.source_tbl(src):FieldSchema(name:INPUT__FILE__NAME, type:string, comment:),
            default.source_tbl(src):FieldSchema(name:ROW__ID, type:struct<transactionId:bigint,bucketId:int,rowId:bigint>, comment:)
           ];
 
DependencyKey=default.target_tbl:FieldSchema(name:col_002, type:int, comment:null)
Dependency=[SCRIPT]
           [default.source_tbl(src):FieldSchema(name:col_001, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:col_002, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:col_003, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:BLOCK__OFFSET__INSIDE__FILE, type:bigint, comment:),
            default.source_tbl(src):FieldSchema(name:INPUT__FILE__NAME, type:string, comment:),
            default.source_tbl(src):FieldSchema(name:ROW__ID, type:struct<transactionId:bigint,bucketId:int,rowId:bigint>, comment:)
           ];
 
DependencyKey=default.target_tbl:FieldSchema(name:col_003, type:int, comment:null)
Dependency=[SCRIPT]
           [default.source_tbl(src):FieldSchema(name:col_001, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:col_002, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:col_003, type:int, comment:null),
            default.source_tbl(src):FieldSchema(name:BLOCK__OFFSET__INSIDE__FILE, type:bigint, comment:),
            default.source_tbl(src):FieldSchema(name:INPUT__FILE__NAME, type:string, comment:),
            default.source_tbl(src):FieldSchema(name:ROW__ID, type:struct<transactionId:bigint,bucketId:int,rowId:bigint>, comment:)
           ];

When INSERT statement doesn't include "ROW_NUMBER() OVER() AS r_num", the lineage details look correct.

This issue is seen in Hive version 1; but not in Hive2 or Hive3.

Issue Links

• is caused by: HIVE-20633 Incorrect column lineage: each output column has input from *all columns* of the input table