New CTable.group_by() implementation

[FrancescAlted]

This PR adds CTable.group_by() with grouped reductions, optimized Cython kernels, and a public blosc2.group_reduce() array API.

Highlights:

• Added CTable.group_by() with:
  • size(), count(), agg()
  • convenience methods: sum(), mean(), min(), max()
  • sort= and dropna= support
  • persistent output via urlpath=
• Added optimized Cython group-by paths:
  • fused integer-key kernels for all integer widths
  • size, count, sum, mean, min, max
  • arbitrary float-key hash path
  • conservative two-key integer/dictionary hash path
• Added public:

    blosc2.group_reduce(keys, values=None, op="sum")

• Extended benchmarks and documentation.
• Added tests for group-by semantics, optimized paths, persistent output, and group_reduce().

[Copilot]

Pull request overview

This PR introduces group-by functionality for blosc2.CTable and a new array-oriented grouped reduction API, backed by optimized Cython kernels and covered by new tests/benchmarks/docs.

Changes:

• Added CTable.group_by() returning a deferred CTableGroupBy with grouped reductions (size, count, agg, plus sum/mean/min/max) and optional persistent output via urlpath=.
• Added public blosc2.group_reduce(keys, values=None, op=..., sort=..., dropna=...) plus NumPy fallback and optimized Cython execution paths.
• Integrated a new groupby_ext Cython extension into the build, and added tests/benchmarks/documentation for the new APIs.

Reviewed changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/test_group_reduce.py	New unit tests for blosc2.group_reduce() semantics, NaN handling, sorting, and input validation.
tests/ctable/test_object_spec.py	New tests for schema-less object columns (heterogeneous values, persistence, Arrow export rejection).
tests/ctable/test_nested_append.py	New tests for appending/extending nested dict rows into dotted (nested) column names.
tests/ctable/test_groupby.py	New comprehensive tests for CTable.group_by() behavior, fast paths, dropna/sort, multi-key, persistence.
src/blosc2/indexing_ext.pyx	Minor cleanup / section delimiter after removing group-by related kernels.
src/blosc2/groupby.py	New Phase-1 group-by implementation, multiple optimized execution paths, and group_reduce() public API.
src/blosc2/groupby_ext.pyx	New Cython kernels for dense integer keys, integral-float dense keys, arbitrary-float hash, and two-key hash.
src/blosc2/ctable.py	Added CTable.group_by() API and adjusted dictionary-column initialization sizing.
src/blosc2/init.py	Exported CTableGroupBy and group_reduce at the top level.
plans/ctable-groupby.md	Added/updated implementation plan and current status write-up.
doc/reference/reduction_functions.rst	Documented group_reduce in the reduction functions reference.
doc/reference/ctable.rst	Documented CTable.group_by() and CTableGroupBy API and examples.
CMakeLists.txt	Added build/install integration for the new groupby_ext extension module.
bench/ctable/groupby.py	New benchmark harness for group-by performance comparisons and options.
bench/ctable/bench_nested_filter_index.py	New benchmark for nested (dotted) vs flat column filter/index/aggregate overhead.

Comments suppressed due to low confidence (2)

src/blosc2/groupby_ext.pyx:623

• groupby_dense_int_f64_max_checked is missing the basic input/state shape validation that the other "*_checked" kernels perform. Because the function runs with boundscheck=False, length mismatches can cause memory corruption. Please add explicit checks for keys/values/valid lengths and for maxs/has_value/keys_present lengths (and ideally dtype checks for public robustness).

def groupby_dense_int_f64_max_checked(
    dense_int_key_t[:] keys,
    double[:] values,
    np.npy_bool[:] valid,
    double[:] maxs,

src/blosc2/groupby_ext.pyx:694

• groupby_dense_int_i64_max_checked lacks the standard length checks for inputs and state arrays. Since this runs in nogil with bounds checking disabled, a mismatch can crash Python. Add shape validation (and keep it consistent with sum/mean kernels) before calling _dense_int_key_scan.

def groupby_dense_int_i64_max_checked(
    dense_int_key_t[:] keys,
    int64_t[:] values,
    np.npy_bool[:] valid,
    int64_t[:] maxs,

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.