Is PLR6201 technically incorrect?

[trag1c]

From the docs:

When testing for membership in a static sequence, prefer a set literal over a list or tuple, as Python optimizes set membership tests.

This optimization isn't significant when using literals—while sets do have faster membership checks, doing 1 in (1, 2, 3) is faster than 1 in {1, 2, 3} as it takes 4–5x as much time to just construct a set compared to a tuple (see benchmarks below).

Benchmarks

Creating an object with elements 1, 2, 3

Version	tuple	set	list	tuple vs set	list vs set
3.12	10.28ns	48.71ns	42.37ns	+374%	+15%
3.11	8.24ns	39.88ns	25.55ns	+384%	+56%
3.10	7.98ns	40.73ns	26.86ns	+410%	+52%
3.9	8.73ns	40.00ns	26.03ns	+358%	+54%
3.8	9.43ns	39.84ns	27.38ns	+322%	+46%

Checking whether 1 is a member of the object (w/ a literal)

Version	tuple	set	list	tuple vs set	list vs set
3.12	15.19ns	16.54ns	15.20ns	+9%	+9%
3.11	13.47ns	14.74ns	14.14ns	+9%	+4%
3.10	13.83ns	15.20ns	13.83ns	+10%	+10%
3.9	15.66ns	16.94ns	15.61ns	+8%	+9%
3.8	17.83ns	19.00ns	17.81ns	+7%	+7%

Therefore (even though the difference isn't that substantial) I think it would make more sense for the rule to suggest

• using a tuple when checking against a literal, e.g. 1 in (1, 2, 3), and
• using a set when checking against a constant, e.g. S = {1, 2, 3}; 1 in S (although I suspect literals may be faster than constants for small objects (because of variable lookup?), in which case tuples would be the way).

Hopefully I haven't made any mistakes with my reasoning 😄

[charliermarsh]

I think this was discussed a bit in #8758. Do you mind reading that issue first?

[trag1c]

I somehow missed this specific one, my bad—seems like I missed a lot of details here, so I guess it's fine to keep it the way it is in most cases 🙂