[GraysonWu]

Cached Tabular predictions build a full envelope, including encoding every context row, and then discard the context before sending the request. Skip that encoding for session scoring while retaining schema checks, pinned dtype conversion, and query serialization. One-shot predictions and session creation still encode the context.

Measured benefits

Mean one-row prediction latency fell from 467.4 ms to 52.4 ms: approximately 415 ms saved per call, 89% lower latency, or 8.9× faster. This improvement comes from removing repeated encoding of the already-cached context.

Benchmark configuration

• GPU: NVIDIA A10G; client and server on the same machine.
• Model: Kumo Tabular Large, BF16, one estimator.
• Task: fraud classification using a fitted session with a cached context.
• Input: 10,000 context rows, 30 numerical features, one query row.
• Timed operation: FittedModel.predict(...).to_pandas(); context fit and Request construction are outside the timer.

Validation

• Same server and fitted session, two warmups followed by 20 alternating measurements per version. This is a sequential latency comparison, not a throughput or P99 claim.
• Payload contents matched excluding request IDs. Predictions were bitwise identical for query batches of 1, 8 and 1,000 rows.
• Four regression cases covering classification/regression with and without a pinned schema fail before the change and pass afterward. Public API and all nine wire-payload snapshots are unchanged.
• Ruff, formatting, mypy, repository hygiene and attribution checks passed for this patch.
• CI currently needs the compatibility and built-wheel test fixes in Support DuckDB 2 and pandas 3.1 on main, and test the wheels CI built #51. Verified this patch applied to Support DuckDB 2 and pandas 3.1 on main, and test the wheels CI built #51 at f54b0938af0fcb12706b3a740304d26ae4837f0c with pandas 3.1.0 and DuckDB 2.0.0: main suite 2,679 passed, separate engine suite 2,580 passed, and 1,963 client tests passed against the exact newly built wheel installed by filename. Live-service tests were excluded; engine verification reused the unchanged native extension. These are local combined-tree results, not a green CI claim for this PR alone.

[GraysonWu]

CI follow-up for 513c860: the six failing jobs match the dependency and wheel-installation issues addressed by #51.

• Python 3.11–3.14: each fails the existing large-integer description test with pandas 3.1. Example log.
• Engine: DuckDB 2.0 emits an Arrow-conversion warning that the suite treats as an error. Log.
• Installed wheel: the log shows pip downloading the released client 1.0.0 from PyPI instead of installing this PR's built wheel. The four new context-encoding regression tests therefore run against the old implementation and fail; this job also hits the pandas failure. Log.

#51 addresses the dependency compatibility problems and installs the built wheels explicitly, with a check of their origin. It is still open. The recommended next step is to update this branch after #51 merges and rerun CI, rather than duplicate those fixes here.

The envelope/session test file was rerun locally on this PR: 29 passed, including the new regression tests. Hosted CI remains failing until the shared fixes are incorporated and verified.