Very slow unique sampling of np.int8 about hypothesis

[jobh]

(fyi: just reporting, not working on)

[Zac-HD]

Unclear what we can do about this, beyond complicating our fast path with support for a broader range of strategies. "Seeing through" a MappedStrategy is only valid in cases where the transform callable is fairly pure, which means we'd also need an allowlist for that, and I'm just not convinced it's worth it. At some point we have to tell power users to write st.sets(st.sampled_from([your stuff here])) rather than rewriting it for them.

[jobh]

Thanks, and I appreciate your point. I guess the test name (test_efficiently_generates_all_unique_array) got me wound up.

Still: it would be nice for that power user, or the np extra, to be able to derive custom types that are as powerful as the built-in types. A few ideas below, for posterity if this is revisited.

1. We don't really need to check purity of the mapping, just repeat the uniqueness check afterwards. A mapping can only narrow the underlying strategy.
2. We could even add purity hints for MappedStrategy. If we know the mapping is bijective, we don't need to repeat the uniqueness check. While possibly making users feel clever and empowered, I agree this is probably not worth the effort — unless we see other cases. OTOH, it would be simpler than (1) if we limited see-through to bijective mappings.
3. Simpler would be to add a (possibly private/undocumented) special form of .map(). Perhaps an astype= argument to the strategy. This would be applied immediately to the result of .do_draw() without deferring to another strategy, and also anywhere these draws are replaced of course.

Anyway, feel free to close.

[jobh]

Actually, don't close quite yet. I'll fix it quickly (I think) in the numpy extra.