Re: [Bug] Duplicate results for inheritance and FOR UPDATE.

Hello, we found (and maybe fixed) two wrong behavior related to
inheritance and FOR UPDATE. This report is about one of them.

This behavior could be observed on master and back to 8.4 but 8.4
behaves a bit more funny. I haven't checked on 8.3.

This issue is that some query returns duplicate rows after FOR
UPDATE was blocked, in other words, after getting
HeapTupleUpdated in ExecLockRows.

- Reproducing the symptom

This behavior is easy to reproduce as following.

A=# DROP TABLE IF EXISTS p CASCADE;
A=# CREATE TABLE p (a int, b int, c int);
A=# CREATE TABLE c1 (LIKE p INCLUDING INDEXES) INHERITS (p);
A=# CREATE TABLE c2 (LIKE p INCLUDING INDEXES) INHERITS (p);
A=# CREATE TABLE c3 (LIKE p INCLUDING INDEXES) INHERITS (p);
A=# INSERT INTO c1 (SELECT 0, a / 3, a % 3 FROM generate_series(0, 9) a);
A=# INSERT INTO c2 (SELECT 1, a / 3, a % 3 FROM generate_series(0, 9) a);
A=# INSERT INTO c3 (SELECT 2, a / 3, a % 3 FROM generate_series(0, 9) a);
A=# ANALYZE;
A=# BEGIN;
A=# SELECT tableoid, ctid, * FROM p WHERE b IN (0, 1) AND c = 0 FOR UPDATE;
A=# UPDATE p SET b = -1 WHERE a = 1 AND b = 1 AND c = 0;
B=# SELECT tableoid, ctid, * FROM p WHERE b IN (0, 1) AND c = 0 FOR UPDATE;
-- the above is blocked by session A until it commits.
A=# COMMIT;

Where A and B are individual sessions.

You will get the follwing result on sesison B just after the
COMMIT on session A.

| tableoid | ctid | a | b | c
| ----------+-------+---+---+---
* 33834 | (0,1) | 0 | 0 | 0
| 33834 | (0,4) | 0 | 1 | 0
| 33838 | (0,1) | 1 | 0 | 0
* 33834 | (0,1) | 0 | 0 | 0
| 33842 | (0,1) | 2 | 0 | 0
| 33842 | (0,4) | 2 | 1 | 0
| (6 rows)

The lines prefixed with '*' appear twice in the result.

The plan for the query is as following,

LockRows
-> Result
-> Append
-> Seq Scan on p
Filter: ((b = ANY ('{0,1}'::integer[])) AND (c = 0))
-> Seq Scan on c1
Filter: ((b = ANY ('{0,1}'::integer[])) AND (c = 0))
-> Seq Scan on c2
Filter: ((b = ANY ('{0,1}'::integer[])) AND (c = 0))
-> Seq Scan on c3
Filter: ((b = ANY ('{0,1}'::integer[])) AND (c = 0))

- Analysys and solution

This seems to be caused in ExecScanFetch under EvalPlanQualNext,
when es_epqTuleSet[scanrelid - 1] is false (p and c1 in the above
example). In that case, execution confluents onto non-EPQ
route. As the result, the accessMtd (= SeqScan, IndexScan will
behaves as the same) retunes null slot at the first iteration on
p. Then the second iteration on c1, it reuturns the c1's first
tuple (33834:(0,1)) and EvalPlanQualNext is satisfied with the
wrong tuple.

In the EPQ block in ExecScanFetch, it seems should return NULL if
the relation does not has test tuple. The attached patch does so
on the current master and the problem has disappeared.

regards,

--
Kyotaro Horiguchi
NTT Open Source Software Center

Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> writes:
> This issue is that some query returns duplicate rows after FOR
> UPDATE was blocked, in other words, after getting
> HeapTupleUpdated in ExecLockRows.

Good catch!

> In the EPQ block in ExecScanFetch, it seems should return NULL if
> the relation does not has test tuple. The attached patch does so
> on the current master and the problem has disappeared.

... but bad fix. This would break join cases, wherein it's important
that other scan nodes still return all the required tuples. (It's
unfortunate that the eval-plan-qual isolation test fails to cover
joins :-(.)

After digging around a bit, I realized that the problem is actually in
nodeLockRows.c. It is supposed to do EvalPlanQualSetTuple(..., NULL)
for each child table that's not supposed to return any row during the
current EPQ test cycle. Unfortunately, it only does that reliably once
the EPQ environment is already set up. If we discover we need an EPQ
test while looking at a non-first child table, tables already passed
over in the loop over node->lr_arowMarks don't get the word.

So the correct fix (or a correct fix, anyway; this could also have been
done with more-invasive loop logic changes) is as attached. I'm working
on back-patching this.

regards, tom lane

Hello,

> > This issue is that some query returns duplicate rows after FOR
> > UPDATE was blocked, in other words, after getting
> > HeapTupleUpdated in ExecLockRows.
>
> Good catch!

Thank you.

> > In the EPQ block in ExecScanFetch, it seems should return NULL if
> > the relation does not has test tuple. The attached patch does so
> > on the current master and the problem has disappeared.
>
> ... but bad fix. This would break join cases, wherein it's important
> that other scan nodes still return all the required tuples. (It's
> unfortunate that the eval-plan-qual isolation test fails to cover
> joins :-(.)

Hmm. I regretfully forgot to do isolation check. The confluent
didn't seem a bug but I couldn't see its intention.

> After digging around a bit, I realized that the problem is actually in
> nodeLockRows.c. It is supposed to do EvalPlanQualSetTuple(..., NULL)
> for each child table that's not supposed to return any row during the
> current EPQ test cycle. Unfortunately, it only does that reliably once
> the EPQ environment is already set up. If we discover we need an EPQ
> test while looking at a non-first child table, tables already passed
> over in the loop over node->lr_arowMarks don't get the word.

Thank you for the detailed explanation. I was wondering during
investigation about EvalPlanQualSetTuple(, NULL) was called only
for children after the first one that had EPQ test tuple. Now I
understand that it was the core of this bug.

> So the correct fix (or a correct fix, anyway; this could also have been
> done with more-invasive loop logic changes) is as attached. I'm working
> on back-patching this.

That really helps. Thank you.

regards,

--
Kyotaro Horiguchi
NTT Open Source Software Center