Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
2.0-incubating
-
None
Description
Selva found the cause of this problem and suggested the fix.
The problem was discovered as a failure of regression test executor/test015, with some missing index rows.
Selva found that this is because we insert a "tombstone" into each alternate index for every row, even if it is a new row. Those tombstones for the non-existing old rows contain junk or NULL values, and they can potentially wipe out existing index rows.
Selva's suggested fix is to use the "pre-condition" for the index delete to check whether an old row actually existed, and only do a delete if it did.
Here is a simple test case, also from Selva:
drop table t015t8 cascade ; create table t015t8 (i int not null, j int, k int, primary key(i)); create unique index t015t8i1 on t015t8(j); insert into t015t8 values (4,null,4); select * from t015t8; -- enable this CQD to try the new, efficient tree -- cqd TRAF_UPSERT_TO_EFF_TREE 'on'; prepare s1 from upsert into t015t8 values (1,2,3),(11,12,13) ; execute s1 ; set parserflags 1; select * from t015t8; select * from table (index_table t015t8i1); -- wrong answer, 2 rows instead of 3
If this does not reproduce the error or if you want to see the extra tombstones, do an HBase raw scan to see them:
hbase shell
scan 'TRAFODION.SCH.T015T8I1' , {RAW=>true}
