frozendict: an experiment

[Barry Scott]

> On 13 Jul 2020, at 03:20, Marco Sulla <elbarbun at gmail.com> wrote:
> 
> TL;DR: I tried to implement in CPython a frozendict here:
> https://github.com/Marco-Sulla/cpython
> 
> Long explaining:
> 
> What is a frozendict? It's an immutable dict. The type was proposed in
> the past but rejected: https://www.python.org/dev/peps/pep-0416/
> 
> So why did I try to implement it? IMO, apart the considerations in PEP
> 416, a frozendict can be useful:
> 
> - as a faster base for types.MutableMappingProxy
> - as a substitute of namedtuple
> - as set values
> - as a faster dict. Indeed a frozendict does not need views. Keys,
> values and items can be cached and could be a subclass of frozendict
> that implements also the set API (values apart).


Why is it faster? Have you benchmarked that claim?

Why do you think I do not need views to use the frozendict?
I thought that is what make d.key(), d.items() etc work?

> 
> A possible problem is that frozendict requires more memory, since the
> hash is cached.

The hash is tiny 64 bits?

> 
> My implementation is very naif. I tried to do the simplest thing,
> without any speed improvement. I haven't added frozendict in marshal,
> and the pickle implementation is very raw... Furthermore, I didn't
> touch the AST parser at all, and I did not add any test for Python or
> the C API in the battery, but I've done my tests apart, in pytest.
> 
> PS: I used for tests and benchmarks a little modified version of the
> tests for my Python implementation of frozendict:
> https://github.com/Marco-Sulla/python-frozendict/tree/master/test .
> The results are interesting. A CPython builtin frozendict is faster in
> every bench that immutables.Map, that uses the HAMT algorithm. and was
> proposed as alternative in PEP 603:
> https://www.python.org/dev/peps/pep-0603/ . Maybe because I
> implemented it in CPython and immutables.Map is a C extension.
> -- 
> https://mail.python.org/mailman/listinfo/python-list
>