fix(python): Raise properly for slices not supported by LazyFrame

[petrosbar]

Found a couple more slice cases that are mishandled by LazyFrame. Specifically, both cases below are not supported by LazyFrame but they currently raise incorrect exceptions/errors:

1. slice(None, 0, k), where k < 0

import polars as pl

ldf = pl.LazyFrame({"a": [1, 2, 3]})
ldf[:0:-1].collect()
# TypeError: '>=' not supported between instances of 'int' and 'NoneType'

2. slice(i, None, k), where i, k < 0

import polars as pl

ldf = pl.LazyFrame({"a": [1, 2, 3]})
ldf[-1::-1].collect()
# OverflowError: can't convert negative int to unsigned

Both these cases should raise a ValueError: the given slice is not supported by lazy computation

[alexander-beedie]

Interesting! I need to check why the parametric tests didn't catch some of these; I think the fixes looks good, but I want to update the parametric tests such that they find these, to make sure nothing else slips through 🤔

[petrosbar]

@alexander-beedie Actually, I did find these cases (and also #15297) using Hypothesis. I just had to actually modify the strategy a bit.

I think the problem is here:
https://github.com/pola-rs/polars/blob/main/py-polars/polars/testing/parametric/primitives.py#L426

Probably because of the way Hypothesis works, this isn't really very "random", and it is one of the reasons why upgrading Hypothesis gives warnings and #14798 cannot be merged.

[alexander-beedie]

it is one of the reasons why upgrading Hypothesis gives warnings and #14798 cannot be merged.

Adjusting the weights of the test (null probability) made it work better - it found both errors. I have a TODO to fix use of random though; this is more incentive to get that done 😅

[petrosbar]

it is one of the reasons why upgrading Hypothesis gives warnings and #14798 cannot be merged.

Adjusting the weights of the test (null probability) made it work better - it found both errors. I have a TODO to fix use of random though; this is more incentive to get that done 😅

Yes. But as you said that wouldn't solve the actual problem either.
Can you see this branch? main...petrosbar:polars:parametric
This would expose the faulty cases.

[petrosbar]

Note that in this branch, the changed probabilities in test_strategy_null_probability are a bit irrelevant. I had to adjust them, otherwise the test case was flaky. That also shows that the check
if random.random() < null_probability: does not work as expected.

[alexander-beedie]

Note that in this branch, the changed probabilities in test_strategy_null_probability are a bit irrelevant. I had to adjust them, otherwise the test case was flaky. That also shows that the check if random.random() < null_probability: does not work as expected.

Indeed; I need to run around the parametric code addressing usage of the standard random.
Anyway, the fixes here look good 😎