On: Python dicts are now ordered

[onei]

I was under the impression python dict ordering was deterministic, just not in an order recognisable by humans, i.e. ordered by hashes. Was this not the case?

[_gok2 → onei]

From the docs: "CPython implementation detail: Keys and values are listed in an arbitrary order which is non-random, varies across Python implementations, and depends on the dictionary’s history of insertions and deletions."

So I would claim that equates to non-deterministic

[onei → _gok2]

I read that as: if two dicts are constructed in the same way and use the same python implementation then they'd be the same. If you change the way the dicts were generated or the underlying implementation then they're not the same.

To me, nondeterministic is when the input does not change but the output does. The docs essentially say to not rely on the order, not that the result is not reproducible.

[int_19h → onei]

CPython hashes are intentionally non-deterministic, so even the same exact Python install will produce different hash orderings on the same code, if you run it twice (although you can make it deterministic by using PYTHONHASHSEED).

[baq → _gok2]

I’d call undependable but maybe it’s the same thing in practice?

[tshaddox → baq]

That's generally what people mean when they say "nondeterministic" in the context of computing. Yeah, in philosophy it generally means something like "the future is not completely determined by the past," but in computing it means something closer to "the programmer cannot reasonably determine the behavior and thus should not depend on specific behavior."

[mypalmike → tshaddox]

In computing it means a given set of inputs lead to a given set of outputs. It has nothing to do with how difficult it is for a programmer to reason about. Deterministic builds, deterministic tests, etc.

[Dylan16807 → mypalmike]

But what counts as "input" will vary based on who you ask and under what context.

[masklinn → onei]

> Was this not the case?

From the origin to Python 3.2, iteration order was arbitrary but deterministic (in CPython, it was not guaranteed to be determinstic in other implementations)

From Python 3.3 to Python 3.5, iteration order was non-determinstic (across runs, it was deterministic per-process instance)

In Python 3.6 it's unspecified but deterministic and follows insertion order (in CPython)

In Python 3.7, Python 3.6's behaviour was specified and documented