[DERBY-6788] Wrong value inserted by INSERT INTO with multiple subselects

• Type: Bug
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Affects Version/s: 10.1.1.0, 10.2.2.0, 10.3.3.0, 10.4.2.0, 10.5.3.0, 10.6.2.1, 10.7.1.1, 10.8.2.2, 10.9.1.0, 10.10.2.0, 10.11.1.1
• Component/s: SQL
• Environment: not relevant, tested under Windows 7 32bit
• Urgency: Normal
• Issue & fix info: Repro attached
• Bug behavior facts: Data corruption
• Assignee: Abhinav Gupta
• Reporter: Manfred Steinsiek
• Created: 09/Jan/15 09:44
• Updated: 30/Aug/15 14:50

Description

Certain INSERT INTO statements with multiple subselects insert wrong values, i.e. may lead to data corruption.

Here is a very simple example how to reproduce this bug: Start with a new (empty) Derby database, create 3 tiny tables, and insert two records:

create table m1 (k varchar(64), s decimal);
create table m2 (k varchar(64), s decimal);
create table v (s decimal);

insert into m1 values ('Bug', 2015);
insert into m2 values ('Bug', 1957);

Now, the following (likewise simple) select

select res.* from (select d2.s from m1
left join (select k,s from m2) as d2 on m1.k=d2.k) as res

yields 1957, of course. - Of course? Not entirely: If I add an INSERT INTO to that select, i.e.

insert into v (select res.* from (select d2.s from m1
left join (select k,s from m2) as d2 on m1.k=d2.k) as res)

then table v contains 1 row (as it should), but in this row s=4,355,431. This value is interesting, because 4355431 = Hex 427567 and 'B' = X42, 'u'=X75, 'g'=X67.

Finally, if I slightly modify the INSERT INTO above as

insert into v (select res.* from (select d2.s*1 from m1
left join (select k,s from m2) as d2 on m1.k=d2.k) as res)

then it works correct.

This phenomenon arose with every Derby version 10.x I tried (see list above).

Possibly this bug is related to DERBY-6786, where similar INSERT INTOs with subselects appear.

Addendum: There is indeed a close relationship between DERBY-6786 and this one: Let's denote by SQL1 the first INSERT INTO above, by SQL2 the second one, i.e.

SQL1 = insert into v (select res.* from (select d2.s from m1 left join (select k,s from m2) as d2 on m1.k=d2.k) as res)
SQL2 = insert into v (select res.* from (select d2.s*1 from m1 left join (select k,s from m2) as d2 on m1.k=d2.k) as res)

We further assume that there are exactly 1 record R1 in table M1 and exactly 1 record R2 in table M2, both without NULL-values. Then:

If R1.k is equal to R2.k, then SQL1 -> data corruption, SQL2 -> correct.
If R1.k is not equal to R2.k, then SQL1 -> correct, SQL2 -> NullPointerException.