tea-tasting 2.0.0

Summary

tea-tasting 2.0 is here with:

• Fewer default dependencies.
• A more consistent and cleaner API.
• Optional performance improvements for parallel simulations with batched execution.
• Adoption of Scientific Python SPEC standards.

tea-tasting 2.0 shouldn't break most users' workflows. Read the breaking changes below if you:

• Use deprecated RNG aliases such as seed or random_state.
• Pass custom data generators or progress callbacks to Experiment.simulate.
• Use Mean(value=...) by keyword.
• Rely on Quantile's previous default NaN handling.
• Pass Ibis tables without installing Ibis explicitly.
• Use tea-tasting with Python <3.12, NumPy <2.1, SciPy <1.15, or PyArrow <17.

Breaking changes

• Deprecated RNG aliases were removed. Use rng instead of seed in make_users_data, make_sessions_data, and Experiment.simulate. Data generator callables passed to Experiment.simulate must also accept rng, not seed. Use rng instead of random_state in Bootstrap and Quantile; the random_state attribute was removed. More details in SPEC 7.
• Mean(value=...) was renamed to Mean(column=...). Positional calls such as Mean("orders") still work, but keyword usage must change from value= to column=, and the public attribute is now column instead of value.
• nan_policy is now configurable globally with set_config() and config_context(). MannWhitneyU, Bootstrap, and Quantile now inherit nan_policy from global config when it is not passed explicitly. The default global policy is "propagate". This changes Quantile’s default behavior from "omit" to "propagate".
• Experiment.simulate now runs simulations in batches. Use the new batch_size parameter to control how many simulations run per batch, including when using map_ or progress reporting. Custom progress callbacks must follow the new context-manager style: progress(total=n_simulations) must return a context manager whose __enter__() value has an update(n) method.
• Ibis is now an optional dependency. Installing tea-tasting no longer installs ibis-framework; users who pass Ibis tables must install Ibis separately. For example, install the backend extra you need: uv add 'ibis-framework[duckdb]' or pip install 'ibis-framework[duckdb]'.
• Minimum Narwhals dependency was raised to 2.8 to support direct DuckDB relation inputs without Ibis.
• Python 3.12 or later is now required. Minimum dependencies were raised to NumPy 2.1, PyArrow 17, and SciPy 1.15. More details in SPEC 0.

All changes

• docs: recommend uv add for installation commands by @e10v in #274
• refactor: narrow variant key annotations to hashable by @e10v in #275
• chore: migrate type checking from pyright to ty by @e10v in #276
• refactor: support mapping metric results by @e10v in #277
• build!: require Python 3.12 and raise dependency minimums by @e10v in #278
• feat!: remove deprecated rng aliases by @e10v in #279
• refactor: pass rng to scipy bootstrap by @e10v in #280
• feat!: make Ibis an optional dependency by @e10v in #281
• feat!: add global nan_policy config by @e10v in #282
• feat!: batch experiment simulations by @e10v in #283
• docs: clarify Ibis installation examples by @e10v in #284
• feat!: rename Mean value parameter to column by @e10v in #285
• build!: require Narwhals 2.8 for direct DuckDB support by @e10v in #286
• docs: add DuckDB relation data backend example by @e10v in #287
• chore: pin workflow action versions by @e10v in #288
• chore: ignore Marimo local files by @e10v in #289
• docs: clarify multiprocessing limits in simulation guide by @e10v in #290

Full Changelog: v1.14.0...v2.0.0