Luo Zhengjie created FLINK-41327:
------------------------------------
Summary: Binary IN/NOT IN predicates compare byte arrays by
identity in SEARCH point sets
Key: FLINK-41327
URL: https://issues.apache.org/jira/browse/FLINK-41327
Project: Flink
Issue Type: Bug
Components: Table SQL / Runtime
Reporter: Luo Zhengjie
The runtime ObjectHashSet used for SEARCH point sets delegates to HashSet. SQL
BINARY/VARBINARY values are represented as byte[], whose Java equality is
identity-based. A separately allocated value containing the same bytes as an
IN-list literal can therefore be reported as absent; NOT IN can produce the
opposite incorrect result.
h3. Minimal runtime reproduction
{code:java}
ObjectHashSet<byte[]> set = new ObjectHashSet<>();
set.add(new byte[] {1});
boolean found = set.contains(new byte[] {1});
// Expected: true. Actual: false.
{code}
Reproduced on community master ce373fe8b9e4131a0d756b739449e410318a74b5
(2.5-SNAPSHOT), using Java 17 and a source reactor build. New runtime
regression tests run: 3; failures: 2; errors: 0. The failures cover
equal-content membership and isolation from mutation of the inserted byte array.
CodeGeneratorContext.addReusableHashSet selects ObjectHashSet for binary
values. SQL coverage must use non-folded input and enough point literals to
exercise the SEARCH hash-set path rather than a short OR chain. SQL integration
validation is still in progress; no full-suite success is claimed.
h3. Proposed fix
Use content-based equality and hashing for byte-array keys, retain normal
equality for other key types, and copy inserted byte-array keys so later caller
mutation cannot invalidate membership. No dependency or storage-format changes.
Prepared with assistance from OpenAI Codex (GPT-5).
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
