Konstantin Bereznyakov created HIVE-29598:
---------------------------------------------
Summary: Vectorization: Scratch column reuse sometimes produces
wrong results
Key: HIVE-29598
URL: https://issues.apache.org/jira/browse/HIVE-29598
Project: Hive
Issue Type: Bug
Reporter: Konstantin Bereznyakov
The vectorizer can assign the same physical `LongColumnVector` slot to two
different logical columns: an intermediate scratch result (e.g. `cast(STRING AS
INT)`) and a small-side broadcast value column carried by a
`VectorMapJoinOuter*Operator`. The intermediate writes a typed value (an
`int`); the broadcast then writes a typed value of a different logical type
(e.g. a `boolean`) to the same slot for matched rows.
For *unmatched* rows the broadcast write is skipped — only `isNull[i]` is
flipped — and the slot retains the stale intermediate value. Any downstream
operator that reads `vector[i]` without first checking `isNull[i]` then
propagates a value of the wrong logical type into the query result. The bug is
silent.
Reproduced on Apache Hive master, commit `6f2355e22c` (`pom.xml` declares
version `4.3.0-SNAPSHOT`). Released 4.x versions sharing this `Vectorizer` /
`VectorMapJoinOuter*Operator` / `ColOrCol` code are likely also exposed but
were not separately verified by this filing.
The repro below uses the projection-mode `ColOrCol` path; analogous
projection-mode operators that read `vector[i]` directly (`ColAndCol`,
long-column arithmetic, equality/comparison expressions,
`VectorUDFAdaptor`-routed UDFs) are exposed by the same root cause but are not
separately verified by this filing.
1. When does the bug fire?
All of the following co-occurring conditions are required:
- Vectorized execution is on (default).
- A vectorized OUTER-join MapJoin runs (`VectorMapJoinOuter*Operator`). INNER
/ LEFT SEMI / ANTI variants do not propagate stale slot data through unmatched
rows.
- The vectorizer aliases a scratch column with a `smallTableValueMapping`
target — typical when a `LongColumnVector`-backed type (e.g. `int`) is computed
for the join-key expression and a `LongColumnVector`-backed broadcast value
(e.g. `boolean`) is also produced from the small side.
- Some output row is unmatched on the join keys, going through
`generateOuterNulls`.
- A downstream operator reads `vector[i]` of the affected slot without
consulting `isNull[i]`.
- The reader's contract is violated by the dirty value (e.g. a strict `== 1`
check fails on a non-` \{0,1}` long).
Removing any one condition masks the bug.
2. Reproduction
{code:java}
SET hive.auto.convert.join=true; – default true
SET hive.auto.convert.join.noconditionaltask=true; – default true
CREATE TABLE t (k STRING, v STRING) STORED AS ORC;
INSERT INTO t VALUES
('A','1'),('A','2'),('A','3'),
('B','2'),('B','3'),
('C','3'),
('D','1'),('D','3');
WITH
probe AS (
SELECT k, v, (CAST(v AS INT) > 0) AS p_bool
FROM t WHERE CAST(v AS INT) >= 3
),
small_side AS (
SELECT k, v, (CAST(v AS INT) > 9999) AS s_bool
FROM t
),
classified AS (
SELECT p.k, p.v, CAST((s.s_bool OR p.p_bool) AS INT) AS observed_value
FROM probe p
LEFT JOIN small_side s
ON p.k = s.k
AND CAST(p.v AS INT) - 1 = CAST(s.v AS INT)
),
diagnosed AS (
– Aggregate barrier: prevents Calcite from inlining the inner expression
– back into the outer WHERE and simplifying the bug surface away.
SELECT k, v, MAX(observed_value) AS observed_value
FROM classified
GROUP BY k, v
)
SELECT k, v,
observed_value AS observed_value_returned_by_select,
1 AS required_value_per_sql_semantics
FROM diagnosed
WHERE observed_value = 0
ORDER BY k, v;
{code}
{*}Expected{*}: zero rows. Every probe row's `(s_bool OR p_bool)` is `TRUE` per
SQL semantics (matched: `FALSE OR TRUE`; unmatched: `NULL OR TRUE`), so
`cast(... AS INT) = 1` for every row, and no row can satisfy `WHERE
observed_value = 0`.
{*}Actual{*}:
{code:java}
C 3 0 1
D 3 0 1{code}
Two rows are returned. The filter `WHERE observed_value = 0` accepts them
because their stored `observed_value` is 0. The `1 AS
required_value_per_sql_semantics` projected alongside makes the contradiction
visible per row: SQL semantics demand value 1; the database returned 0.
`C` and `D` are exactly the unmatched probe rows.
3 Workaround
{code:java}
SET hive.vectorized.execution.enabled=false;{code}
Same query then returns zero rows.
4. Mechanism (pointer to the slot reuse)
`EXPLAIN VECTORIZATION DETAIL` of the repro shows physical column *{*}`4`{*}*
allocated for two purposes:
{code:java}
bigTableKeyExpressions:
LongColSubtractLongScalar(col 4:int, val 1)(children:
CastStringToLong(col 1:string) -> 4:int) -> 6:int
className: VectorMapJoinOuterMultiKeyOperator
smallTableValueMapping: 4:boolean{code}
For matched rows the broadcast writes a `boolean` (0/1) into `vector[col
4][i]`, overwriting the cast intermediate. For unmatched rows
`generateOuterNulls`
(`ql/.../mapjoin/VectorMapJoinOuterGenerateResultOperator.java`) flips
`isNull[col 4][i] = true` but leaves `vector[col 4][i]` holding the stale `int`
from `CastStringToLong`. A subsequent projection of `cast(s_bool OR p_bool AS
INT)` reads the slot through `ColOrCol` (`ql/.../expressions/ColOrCol.java`)
followed by `IfExprLongScalarLongScalar`, and the dirty value (`stale_int | 1`)
fails the latter's strict `== 1` check, producing 0.
Additional note: disabling CBO (which steers the plan away from this vectorized
path) is also observed to produce correct results, consistent with the analysis
above.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
