David Mollitor created SPARK-59934:
--------------------------------------
Summary: pmod silently wraps around for INT/BIGINT with
large-magnitude negative divisors
Key: SPARK-59934
URL: https://issues.apache.org/jira/browse/SPARK-59934
Project: Spark
Issue Type: Improvement
Components: SQL
Affects Versions: 4.0.0, 4.1.0
Reporter: David Mollitor
{code:sql}
-- INT: expected -1, actual 2147483647
SELECT pmod(-1, CAST(-2147483648 AS INT));
-- BIGINT: expected -1, actual 9223372036854775807
SELECT pmod(CAST(-1 AS BIGINT), CAST(-9223372036854775808 AS BIGINT));
-- Correct result for the same magnitude via DECIMAL: -1
SELECT pmod(CAST(-1 AS DECIMAL(38,0)), CAST(-2147483648 AS DECIMAL(38,0)));
{code}
h3. Root cause
For {{{}n < 0{}}}, {{MathUtils.pmod}} computes {{r = a % n}} and, when {{{}r <
0{}}}, returns
{{(r + n) % n}} to give a result that shares the sign of the dividend. For
sufficiently
large-magnitude negative divisors (roughly {{n < -2^30}} for {{{}INT{}}}, {{n <
-2^62}} for
{{{}BIGINT{}}}), the intermediate {{r + n}} overflows the integral type and
wraps, so the
final value is positive and no longer congruent to {{a}} modulo {{{}n{}}}. The
{{DECIMAL}}
path uses {{BigDecimal}} arithmetic and does not overflow, hence the cross-type
disagreement.
h3. Expected behavior
{{pmod(a, n)}} should return a result consistent across all integral types
(matching the
{{DECIMAL}}/{{BYTE}}/{{SHORT}} result), or raise {{ARITHMETIC_OVERFLOW}} under
ANSI mode
rather than silently returning a wrapped value.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
