Release review and verification

Reviewed 2026-10-01 against current primary documentation, PyPI release metadata and released source. This skill documents Vaex 4.19.0/core 4.19.0, HDF5 0.15.0, viz 0.6.0 and ML 0.19.0. The core wheel requires Python >=3.9,<3.13; meta-package version and upload date alone do not establish interpreter or NumPy support.

Native evidence

Python 3.12.10 on macOS ARM; NumPy 2.5.3, pandas 2.3.3, PyArrow 25.0.1, Matplotlib 3.11.2, scikit-learn 1.9.1, Dask 2024.8.2, Numba 0.68.0, XGBoost 3.4.1, LightGBM 4.7.0, CatBoost 1.2.10, cachetools 7.2.0. Dependencies were installed in a separate environment, never the project environment.

The repository suite executes 23 runnable Python documentation blocks and additional behavior checks (18 tests total): numerical expressions, filters/nulls/NaNs, population variance, approximate percentiles, explicit grouped aggregators, grid orientation and upper-edge exclusion, extracted joins with duplicate keys, async reductions, bounded pandas conversion, cache controls, five local file formats, HDF5 group append, scalers, encoders, PCA/projections, KMeans, sklearn full/incremental fits, three boosting adapters, probability class axis, trusted model-state transfer without importing the training filter, vectorized apply, and numerical Dask-array conversion. Histogram/scatter/grid PNGs were rendered and visually checked for axes, labels, transposition and cropping.

The heatmap failure with current Matplotlib was reproduced: Vaex calls removed matplotlib.cm.get_cmap. The documented fallback uses Vaex's bounded grid aggregation and native Matplotlib pcolormesh; no monkeypatch or legacy plotting claim is made. CatBoost's default Probability output was also observed on an RMSE-trained model; regression checks explicitly use RawFormulaVal.

These tiny checks establish API mechanics and numerical invariants, not billion-row performance, memory peaks, statistical calibration, model convergence or scientific validity. Approximate percentiles need not match NumPy's interpolation convention. No data were obtained from external scientific datasets and no cloud/server write occurred.

Source-only boundaries

The installed contents of 23 core, HDF5, viz and ML modules were compared byte-for-byte with the official v4.19.0 tag and matched. Additional current server 0.10.0 and Jupyter 0.9.0 wheel sources were inspected for commands/accessors. Their optional integrations remain illustrative:

• S3/GCS filesystem handling and authentication-option routing were inspected; no authenticated cloud reads/writes, Azure integration or provider endpoint was tested.
• Vaex server uses vaex server and a named WebSocket DataFrame URL. No running server, TLS/auth deployment or Python-client round trip was validated. The skill does not invent REST endpoints or pagination semantics.
• Widget accessor names were checked in Jupyter source; no notebook frontend was tested.
• TensorFlow KerasModel.fit is intentionally unimplemented; inference-wrapper source was checked but no TensorFlow installation/training/inference/serialization was run.
• FITS/vaex-astro, arbitrary external HDF5 schemas, large remote data and other platforms remain unexecuted. Use small representative round trips before production use.

Primary sources

• Core 4.19.0 metadata and wheels
• Vaex 4.19.0 meta-package
• ML 0.19.0
• HDF5 0.15.0
• Viz 0.6.0
• Official tagged source
• API reference
• I/O guide
• Missing values
• Async execution
• Caching
• Dask arrays
• ML tutorial
• Server guide
• Notebook integration
• Matplotlib 3.9 API removals

Earlier references mixed pandas/NumPy conventions with unsupported Vaex calls. This refresh preserves the six existing capability areas, consolidates repeated examples and removes unsupported APIs rather than claiming them as untested recipes.