Dict Unpacking in Python | Hacker News Reader

Dict Unpacking in Python

github.com

146 points_ZeD_65 comments

Lame: Submit a PEP, campaign for community support, write a patch, go back and forth with the maintainers, endure weeks and months of bikeshedding, then maybe, eventually have your feature included in the next Python release.

Game: Use the codec hack, immediately publish your feature for all Python versions, then write "please do not use" to be safe.

While not nearly as fun as the OP, I’d note that this sort of unpacking is very pleasant in the newish PEP636 match case statements:

https://peps.python.org/pep-0636/#matching-builtin-classes

Looks really cool!

Will this allow combinations of bound and unbound variables?

E.g.:

  def is_on_horizontal_line(point, line_y):
    match point:
      case (x, line_y):
        return f"Yes, with x={x}"
      case _:
        return "No"

Seems both useful and potentially confusing.

It allows you to use bound variables/constants as long as the expression includes a dot so you can distinguish it from a capture variable.

Scala allows matching against bound variables but requires it either start with an uppercase letter or be surrounded in backtics in the pattern. I don't know that that would make sense for python, but there could potentially be some special syntax to spicify you want to compare against an existing variable instead of capturing a new variable.

Ah, that makes sense. Maybe the "exceptions" (dots, uppercase letters, etc) are needed to permit bound variables that we usually don't think of as variables at all, like class or package identifiers?

Did not know that such things could be accomplished by registering a new file coding format. Reminds me of https://pypi.org/project/goto-statement/

This one is arguably even more of a hack; it's working at the source code level rather than the AST level.

The "coding" here is a bytes-to-text encoding. The Python lexer expects to see character data; you get to insert arbitrary code to convert the bytes to characters (or just use existing schemes the implement standards like UTF-8).

> it's working at the source code level rather than the AST level.

this (lexing) is the only use of the codec hack - if you want to manipulate the AST you do not need this and can just to `ast.parse` and then recompile the function.

Indeed; the goto hack works that way (and uses a decorator to make it easier to invoke the AST-manipulation logic).

I think there's a package to treat Jupyter notebooks as source code (so you can import them as modules).

While the OP package is obviously a joke, the one with notebooks is kind of useful. And, of course, obligatory quote about how languages that don't have meta-programming at the design level will reinvent it, but poorly.

You talking about this?

https://jupyter-notebook.readthedocs.io/en/stable/examples/N...

I'd argue "import from notebooks" is still only helpful in the "space bar heating" sense.

I think Notebooks are great for quick, "explorative" sketches of code. They are absolutely terrible for organizing "production" code.

I know it often happens that something starts in a notebook and then sort of morphs into a generic script or full-on application. But I think, this is usually the signal you should refactor, pull out the "grown" parts from the notebooks and organize them into proper Python modules.

If you have parts that are still experimental or explorative, consider importing your new modules into the notebook instead of the other way around.

Source: personal experience

So I see asottile has gone from backporting released features [1] to backporting unreleased ones!

[1]: https://pypi.org/p/future-fstrings, mentioned in https://github.com/asottile/dict-unpacking-at-home#please-do...

I found dictionary unpacking to be quite useful, when you don't want to mutate things. Code like:

    new_dict = {**old_dict, **update_keys_and_values_dict}

Or even complexer:

    new_dict = {
        **old_dict,
        **{
            key: val
            for key, val in update_keys_and_values_dict
            if key not in some_other_dict
        }
    }

It is quite flexible.

I love the union syntax in 3.9+:

  new_dict = old_dict | update_keys_and_values_dict

Don’t forget the in place variant!

  the_dict |= update_keys_and_values_dict

No no, do forget about it: like += for lists, |= mutates “the dict”, which often makes for awkward bugs.

And like += over list.extend, |= over dict.update is very little gain, and restricts legal locations (augmented assignments are statements, method calls are expressions even if they return "nothing")

The |= does exactly what it says on the tin. How could it not mutate the left side of the assignment?

In typed languages, I’m all about using nice safe immutable variables/values.

But in python, everything is mutable so there’s only so much safety you can wring out of adhering the an immutable style. Any other function can hop in and start mutating things (even your “private” attributes). Plan for mutations occurring everywhere.

I find the fewer mutations the easier code is to understand, at the level of an individual function.

Of course you do not have the safety you would have in a language that enforces immutability, but there is still a cost (in terms of maintenance and the likelihood of bugs) to mutation.

> The |= does exactly what it says on the tin. How could it not mutate the left side of the assignment?

The normal way? If the LHS is an integer. |= updates the binding but does not mutate the object.

Nothing requires that |= mutate the LHS let alone do so unconditionally (e.g. it could update the LHS in place as an optimisation iff the refcount indicated that was the only reference, which would optimise the case where you create a local then update it in multiple steps, but would avoid unwittingly updating a parameter in-place).

edit: you might not be understanding what dict.__ior__ is doing:

  >>> a = b = {}
  >>> c = {1: 2}
  >>> b |= c
  >>> a
  {1: 2}

That is, `a |= b` does not merely desugar to `a = a | b`, dict.__ior__ does a `self.update(other)` internally before updating the binding to its existing value. Which also leads to this fun bit of trivial (most known from list.__iadd__ but "working" just as well here):

  >>> t = ({},)
  >>> t[0] |= {1: 2}
  Traceback (most recent call last):
    File "<stdin>", line 1, in <module>
  TypeError: 'tuple' object does not support item assignment
  >>> t
  ({1: 2},)

That’s not “the normal way”, it’s just the case when the LHS is immutable.

This behavior is congruent to C++ and it’s documented and idiomatic.

It’d be weird if the in-place operator deeply copied the LHS, and problematic for a garbage-collected language in high throughput applications.

> That’s not “the normal way”, it’s just the case when the LHS is immutable.

It’s the default behaviour of |= if you override |, you need to specifically override |= to do otherwise.

> This behavior is congruent to C++

An excellent reason to do something else in and of itself.

> it’s documented and idiomatic.

Documenting a trap does not make it less of a trap. Innocently replacing | by |= induces a massive behavioural change which the user has little reason to expect, and which can easily lead to bugs due to mutating shared data.

> It’d be weird if the in-place operator deeply copied the LHS

I see you understand neither Python nor the behaviour of the | operator.

> problematic for a garbage-collected language in high throughput applications.

dict.update exists, if you specifically need to update the LHS in place it works just fine.

I use the Python package 'sorcery' [0] in all my production services.

It gives dict unpacking but also a shorthand dict creation like this:

    from sorcery import dict_of, unpack_keys
    a, b = unpack_keys({'a': 1, 'b': 42})
    assert a == 1
    assert b == 42
    assert dict_of(a, b) == {'a': 1, 'b': 42}

[0] https://github.com/alexmojaki/sorcery

That seems a bit crazy and like it would lead to unpredictable and hard-to-mantain code. (pardon my candor).

im curios why you think so ?

After using JS, Python dicts and objects feel so cumbersome. I don't see why they need to be separate things, and why you can't access a dict like `dict.key`. Destructuring is the icing on the cake. In JS, it even handles the named args use case like

   const foo = ({name, age, email}) => { }

I'm guessing all of this has been proposed in Python before, and rejected in part because at this point it'd create way more confusion than it's worth.

I don’t mind the distinction of it as a map container keeping dot properties/methods separate from the keyed values. But yeah the endless string quoting is painful coming back from JS, bare key literals in constructors like JS would be a welcome addition for sure, as would named key unpacking like this whole post is about.

In short, it runs a text preprocessor as the source text decoder (like you would decode from Latin-1 or Shift-JIS to Unicode).

yeah that's the funny part here, would never have thought of this

Yeah I had totally forgotten about this. I remember seeing it around a bit in the python 2 days when UTF-8 wasn’t always assumed. The fact a ~macro system can be bolted on using this is impressive, hilarious, and shockingly terrible.

The author also has an accompanying video: https://youtu.be/eqiM0xRmFJg

Coming from lisp/haskell I always wanted destructuring but after using it quite a lot in ES6/Typescript, I found it's not always as ergonomic and readable as I thought.

This confuses me a bit

  dct = {'a': [1, 2, 3]}
  {'a': [1, *rest]} = dct
  print(rest)  # [2, 3]

Does this mean that i can use?

  dct = {'a': [1, 2, 3]}
  {'b': [4, *rest]} = dct
  print(rest)  # [2, 3]

and more explicit

  dct = {'a': [1, 2, 3]}
  {'_': [_, *rest]} = dct
  print(rest)  # [2, 3]

None of the last two LHSes will match `dct`, so you'll get a runtime error.

> Does this mean that i can use?

They'll both trigger a runtime error, since the key you're using in the pattern (LHS) does not match any key in the dict.

Note that `'_'` is an actual string, and thus key, it's not any sort of wildcard. Using a bare `_` as key yields a syntax error, I assume because it's too ambiguous for the author to want to support it.

Anthony is also the maintainer of the deadsnake ppa, if you were searching for reasons to love him more.

Believe he’s the same person who won’t allow pyflakes to support # noqa, because it’s “opinionated.”

As if dropping that word is some sort of justification. I don’t know what the opinion is! Worse is better?

I would donate $500 to the PSF tomorrow if they added this, the lack of it is daily pain

you can't do this consistently across all cases without compiler assistance (see https://doc.rust-lang.org/book/ch19-03-pattern-syntax.html or https://peps.python.org/pep-0636/#matching-builtin-classes linked below).

perfect is enemy of good imo, dict destructuring is so valuable that I'm willing to bend some rules / add some rules to make it possible. can't we just copy whatever JS does?

If it's that valuable to you personally you can use that project to remove your "daily pain". No need to inflict the pain caused by such a thing being present in official Python. Some of us like for the language to remain highly readable.

> you can use that project

It’s not meant for production use. Quite clearly so: https://github.com/asottile/dict-unpacking-at-home#please-do...

> perfect is enemy of good imo

You can't land a language feature that only sometimes works - that's absolutely horrid UX.

> can't we just copy whatever JS does?

I wasn't aware that js does this and I don't know it's implemented. So maybe I should retract my claim about compiler assistance.

It's been one of the headline features keeping me happy in JS for 9+ years.

You shouldn't be using dicts for data that you know the name of anyway - use dataclasses or named tuples. Dicts are best for things with keys that are not known at compile time.

Since when can you use data classes for kwargs? There are plenty of times when you should use a dict even if you know the keys.

You shouldn't be using kwargs! That is also well known to be bad practice (or it should be anyway).

https://medium.com/codex/stop-using-kwargs-as-method-argumen...

http://ivory.idyll.org/blog/on-kwargs.html

Give me another one.

lol i think you didn't read/understand this - the article is about **kwargs (which is sometimes sloppy) while the person you're responding to is talking about "exploding" a dict when calling a function (this does not require **kwargs at all).

So when they said "kwargs" they didn't actually mean **kwargs they meant something else which doesn't even require kwargs? How did you magically infer that?

I’d agree with this, unless the kwargs is typed with the new-ish PEP-692: https://peps.python.org/pep-0692/

Sure, but destructuring makes sense for both. Just assign the variables without corresponding keys a value of None.

Now come on... for code golf? Why on Earth would anyone want extra syntax in a language with already tons of bloat in the syntax that contribute nothing to language's capabilities? It's, in Bill Gates words, like paying to make airplanes heavier...

This package is a funny gimmick, to illustrate, probably, unintended consequences of some of the aspects of Python's parser. Using this for anything other than another joke is harmful...

Destructuring/unpacking and packing add more to the language than whatever crap they stuff into every new python release these days. Let's never forget they _removed_ in its entirety a standard library package, that is the level we are dealing with here...

Not saying it needs to be done with the decoder hack, it should have proper compiler support, but it's basically necessary. It's the best feature in JS hands down.

  >> a, b = 2, 3 # say

  >> d = dict_of(a, b)
  >> d
  {"a": 2, "b": 3}

  >> # somewhere else...
  >> { a, b } = d
  >> a + b
  5

The dict_of() function is already possible at runtime with python `ast` magic, see [1]. That package also has unpack_keys() again made possible with `ast`, but we'd of course want proper language support.

[1] https://github.com/alexmojaki/sorcery

The confusing bit to me is that the LHS of this

{greeting, thing} = dct

is a set, which is not ordered, so why would greeting and thing be assigned in the order in which they appear?

I don't think they are. They are matched by variable names, so this:

  {thing, greeting} = dct

Should have the exact same result.

  def u(**kwargs):
    return tuple(kwargs.values())

Am I missing something, is this effectively the same?

*I realize the tuple can be omitted here

Yours relies on ordering, OP's presumably does not.

You have to pull them out by key name, and not just get everything. Here's a working version, though with a totally different syntax (to avoid having to list the keys twice, once as keys and once as resulting variable names):

  >>> def u(locals, dct, keys):
  ...     for k in keys:
  ...         locals[k] = dct[k]
  ... 
  >>> dct = {'greeting': 'hello', 'thing': 'world', 'farewell': 'bye'}
  >>> u(locals(), dct, ['greeting', 'thing'])
  >>> greeting
  'hello'
  >>> thing
  'world'
  >>> farewell
  Traceback (most recent call last):
    File "<stdin>", line 1, in <module>
  NameError: name 'farewell' is not defined

Modifying locals() is generally frowned upon, as there's no guarantee it'll work. But it does for this example.

Or use itemgetter:

  >>> from operator import itemgetter
  >>> dct = {'greeting': 'hello', 'thing': 'world', 'farewell': 'bye'}
  >>> thing, greeting = itemgetter("thing", "greeting")(dct)
  >>> thing
  'world'
  >>> greeting
  'hello'

Ohhh, nice. Also, attrgetter (which also supports dotted notation to get nested attrs! Sadly, no dotted notation for itemgetter.)

https://docs.python.org/3/library/operator.html#operator.att...

Dotted notation would not work because the keys in a dict can also contain dots. I am not terrible familiar with them but there is something called `lenses` that comes from functional programming that should allow you to access nested structures. And I am pretty sure there must be at least one python library that implements that.

TFA looks things up by key, and allows pulling a subset of the dict.

Python needs a better dictionary. Also, Python needs better names for things than dict.

In the meantime I find myself resorting to these:

- https://pypi.org/project/python-benedict/ - https://docs.pydantic.dev/ - https://github.com/alexmojaki/sorcery

Reply on news.ycombinator.com