[Python-Dev] Guarantee ordered dict literals in v3.7?

[Nick Coghlan]

On 6 November 2017 at 18:46, Victor Stinner <victor.stinner at gmail.com> wrote:
> 2017-11-05 18:50 GMT+01:00 Guido van Rossum <guido at python.org>:
>> I don't see this as a reason to not put this into the language spec at 3.7.
>
> It can prevent some kinds of optimizations. Dictionaries are used
> "everywhere" in Python, so they are very important for performance.
>
> I would prefer to only keep the ordering warranty for function keyword
> arguments and class members, and use explicitly an ordered dictionary
> when needed.
>
> Sorry, I don't have any example right now of a concrete optimization
> which would not be possible with ordered dictionary. But Serhiy
> mentioned the performance impact of ordering in Python 3.6 dictionary
> on deletion.

Note that I *don't* think we should mandate that regular dictionaries
be synonymous with collections.OrderedDict - I think it's fine to say
that regular dicts are insertion ordered *until* a key is deleted, and
after that their ordering is arbitrary.
Insertion-ordered-until-the-first-delete is sufficient to guarantee
serialisation round trips, dict display ordering, and keyword order
preservation in the dict constructor (it's not sufficient to guarantee
ordering for class namespaces, as those may contain "del" statements,
but class bodies are also permitted to use a non-default mapping
type).

However, if we decide *not* to require that dictionaries be insertion
ordered, then I think we should do the same thing Go did, and have
dict iterators start at a random offset in the item sequence (that's
not actually my idea - Greg Smith suggested it at the core dev
sprints).

Otherwise we'll end up standardising the 3.6 behaviour by default, as
folks come to rely on it, and decline to support Python
implementations that don't provide insertion ordering semantics.

Cheers,
Nick.

-- 
Nick Coghlan   |   ncoghlan at gmail.com   |   Brisbane, Australia