OnPair trainer should draw Rust's sample (StdRng + partial_shuffle) for byte-identical dictionaries

Open
  • Is something wrong?
#534 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
22/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
java, rust
Domain
data

Research direction

Start in OnPairEncodingEncoder.Encoder.train, which uses SplittableRandom(42) and a full Fisher-Yates shuffle, and compare it with the Rust training_order in encoding/trainer.rs. Port the partial_shuffle tail selection and doubling retry loop on top of SampleRng, and check that SampleRng.nextIntInclusive matches random_range from rand 0.9.5. Done when the 100k-row URL column gives the same dict_bytes_len and encoded size per chunk as the Rust 0.88.0 oracle.

Written by the indexing model from the issue text.

Description

enhancement PARKED

Problem

OnPairEncodingEncoder does not draw the same training sample as the Rust onpair crate, so it trains different dictionaries on identical data. Prime directive: behavioral parity with Rust.

Found by comparing a 100k-row URL column (https://example.com/products/<n>/reviews?page=<m>, 7 chunks of 16,384 rows) written by vortex-java and by Rust 0.88.0 (the release vortex-jni uses), both at cascade 3 with global dict off. Both pick vortex.onpair per chunk.

dict_bytes_len per chunk:

chunk Rust Java
0 8191 8429
1 8525 8368
2 8447 8387
3 8363 8452
4 8379 8473
5 8757 5670
6 1872 1877

Chunks 0-4 end up within 0.04% in total, but Java's chunk 5 is 18% smaller (131,904 B vs 161,507 B): its trainer built a 761-token dictionary and the cascade then dict-encoded the 467 distinct codes. That one chunk is 29.6 KB of the 31.6 KB (3.2%) by which Java's file is smaller. This is training luck, not a systematic advantage, and it could go the other way on other data.

Cause

  • Rust (onpair 0.2.1, training_order in encoding/trainer.rs): StdRng::seed_from_u64(42) (ChaCha12, rand 0.9.5) and order.partial_shuffle(&mut rng, k) with k = min(n, 2 * sample_fraction * n + 1024). Training consumes only the selected tail order[n-k..]; if that tail is shorter than the byte budget, k doubles and the shuffle restarts.
  • Java (OnPairEncodingEncoder.Encoder.train): SplittableRandom(42), a full Fisher-Yates shuffle, and training consumes order from the start.

Proposed fix

  1. Port rand 0.9.5 partial_shuffle on top of SampleRng (already a bit-exact ChaCha12 StdRng and random_range for u32). It uses IncreasingUniform / calculate_bound_u32, which batches several swaps into one random chunk. Check that SampleRng.nextIntInclusive (written against rand 0.10) gives the same values as 0.9.5's random_range.
  2. Reproduce the k computation, the tail consumption and the doubling retry loop in train.
  3. Parity test: the URL column above, Rust as the oracle, comparing dict_bytes_len and encoded size per chunk (expect identical).

Repro

Scratch crate with vortex = "=0.88.0"; data dumped from the Java CascadeDepthGainIntegrationTest patterns (utf8-urls), seed 42. Not committed.

Dominant language
Java
Stars
3
Forks
2
Avg merge
9h 19m
Merged PRs (30d)
154

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from dfa1/vortex-java

All issues in dfa1/vortex-java

Similar issues

More Java issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.