array.layout.form touches shape

[lgray]

Version of Awkward Array

2.6.1

Description and code to reproduce

import awkward as ak

a = ak.Array([1, 2, 3, 4, 5])

layout, report = ak.typetracer.typetracer_with_report(a.layout.form_with_key())

b = ak.Array(layout)

print(report.data_touched, report.shape_touched)

b.layout.form

print(report.data_touched, report.shape_touched)

results in:

[] []
[] ['node0']

Asking for the form shouldn't touch shape.

[lgray]

This is what is needed in addition to #3017 to avoid problems with dask's deterministic keys project dask/dask#10905.

Also related: dask/dask#10918 (partial fix that doesn't quite hold)

[agoose77]

Technically this does require shape because it's a NumpyArray, and NumpyArray's have shapes as part of their type signature. In this case, it's asking for shape[1:], because the outermost length can be unknown. We don't have a way of saying touch_inner_shape, so we have to touch the whole thing.

[agoose77]

Well, saying that, we could introduce a "fast path" that checks self.data.ndim. This is defined as non-touching because it's "valueless metadata". I think it's the only example of such an operation that doesn't even touch shape; I remember us discussing it recently (and we could git blame that line in _nplikes/typetracer.py to find more discussion if needs be).

e.g.

if self._data.ndim == 1:
    ...
else:
    do_something_with(
        self._data.shape[1:]
    )