leerho opened a new pull request, #785:
URL: https://github.com/apache/datasketches-java/pull/785
## Summary
Java tuple sketches always hashed with the default seed. They wrote a fixed
seed hash and never checked the seed hash, either when reading an image or when
combining sketches. As a result, Java accepted images created with any seed,
and set operations could silently mix sketches built with different seeds. The
C++, Go and Rust tuple sketches support custom seeds and reject mismatches.
This PR brings Java in line with them, and with Java theta.
### API: all changes are additive overloads
No existing signature changes. The new methods and constructors are:
- `UpdatableTupleSketchBuilder.setSeed(long)`, and a seed constructor on
`UpdatableTupleSketch`.
- Seed constructors on `DoubleTupleSketch`, `IntegerTupleSketch`,
`ArrayOfStringsTupleSketch` and `FdtSketch`.
- `TupleUnion(nomEntries, summarySetOps, seed)`,
`TupleIntersection(summarySetOps, seed)` and `TupleAnotB(seed)`.
- Seed overloads of the two static `TupleAnotB.aNotB` methods and of the 8
tuple `JaccardSimilarity` methods. The Jaccard overloads match what #782 added
for theta.
- `TupleSketch.heapifySketch(seg, deserializer, seed)` and a public
`TupleSketch.getSeedHash()`, as on `ThetaSketch` since #782.
### Behavior
- **Writing:** compact images record the sketch's seed hash. Empty images
record zero, as in theta. Images written with the default seed are
byte-for-byte unchanged.
- **Reading:** reading a non-empty compact image checks its seed hash
against the expected seed. That is the default seed unless the new overload is
used. Legacy images (serial versions 1 and 2) were always written with the
default seed.
- **Set operations:** union, intersection, A-not-B and Jaccard check the
seed hash of every non-empty tuple or theta input. Their results carry the
operation's seed hash.
- **Updatable form:** the deprecated updatable serialized form has no seed
hash, so it is always read as the default seed. Serializing an updatable sketch
built with a custom seed throws `SketchesStateException`. Call `compact()`
first. C++ doesn't serialize updatable tuple sketches at all.
For default-seed users nothing changes. The only new exceptions are seed
mismatches, which used to produce wrong results silently.
## Verification
- `mvn verify` passes, including javadoc and style checks.
- The cross-language checks (`check_java_files`, `check_cpp_files` and
`check_go_files`) pass for `TupleCrossLanguageTest`. Java still reads the C++
and Go `tuple_int_*` images, now with their seed hash checked.
- The new test file is `TupleCustomSeedTest`. It covers:
- seed hashes through `copy()`, `compact()` and serialization;
- that the default-seed image still carries 37836, as C++ does;
- the empty-image handling;
- reading compact images with a matching and with a mismatched seed;
- the updatable-form behavior;
- the stateful and stateless union, intersection and A-not-B, with both
tuple and theta inputs;
- Jaccard and `Filter`;
- the derived sketches, including `FdtSketch`.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
