Skip to main content
GoldenMatch 3.0.0 completes the public half of the Polars-eviction program: results are Arrow-native, and the Arrow frame backend is the default engine lane. Inputs are unchanged.

​
The one breaking change: result frames are pyarrow.Table

DedupeResult.golden, .dupes, .unique and MatchResult.matched, .unmatched now return pyarrow.Table (they were polars.DataFrame). If you consumed those frames with Polars, migration is one line:
Common direct translations if you’d rather stay on Arrow: Everything else on the result objects is unchanged: result.clusters, result.stats, result.scored_pairs, to_csv, the notebook display, and __repr__ all behave as before.

​
Inputs are NOT affected

dedupe_df / match_df accept Polars, pandas, and Arrow inputs exactly as in v2 (Arrow C stream polymorphism). Your ingestion code does not change.

​
The Arrow backend is now the default

GOLDENMATCH_FRAME defaults to arrow — the measured-faster lane (~36% faster end-to-end on the 100K zero-config A/B benchmark). Behavior is output-equivalent to the Polars lane, enforced by a differential harness with frozen fixtures. If you need the old lane while validating the upgrade:
v3.1.0 resolution: the Arrow descent completed. polars moved out of the required dependencies — pip install goldenmatch runs the engine end to end on Arrow (a zero-polars CI gate proves a full dedupe with polars imports blocked). The polars-free install is also the FAST configuration: measured head-to-head, the Arrow lane without polars beats the polars-present run (the Rust fused kernels own the hot paths). The [polars] extra is a COMPATIBILITY surface — the classic GOLDENMATCH_FRAME=polars lane, the golden fast-columnar replay when the native kernel is absent, and goldencheck cell-quality weighting — byte-identical to 3.0.x behavior, not an accelerator.

​
Notes for specific surfaces

  • MCP / A2A: tool responses were already JSON — no change.
  • dbt (dbt-goldensuite): unaffected — the internal pipeline result frames are pyarrow.Table on both lanes since the engine descent completed.
  • One behavioral nuance: sampling-based auto-config decisions are deterministic per backend but may legitimately differ between the arrow and polars lanes (documented statistical contract of Frame.sample).