This is a very common need when implementing network protocols. You want to match keys case-insensitively but also preserve the original casing (e.g. for presentation).

When searching on the Web, you see many people reimplementing their own variant (often incomplete, or buggy). For example, Twisted has its own, the email package has something resembling it, WebOb also.

Having an implementation in the stdlib would spare many people the effort, ensure the implementation is complete and well-tested, and perhaps also add some optimizations to mitigate the overhead compared to a plain dict.

Note this is an instance of a more general pattern, where they key used for matching is derived from the lookup key using a constant derivation function. So maybe we want to implement the more general pattern and let users specify str.lower as the key derivation function.

Surely a case-insensitive dict should use str.casefold, not str.lower?

> Surely a case-insensitive dict should use str.casefold, not
> str.lower?

Perhaps. Network protocols will usually only allow ASCII in parts
where case is insensitive (e.g. header names), so it shouldn't make a
difference.

Implementing the generic pattern means this is left at the user's
discretion, though.

For the record, I have my own implementation here:
https://bitbucket.org/optiflowsrd/obelus/src/tip/obelus/casedict.py?at=default
https://bitbucket.org/optiflowsrd/obelus/src/tip/obelus/test/test_casedict.py?at=default

For the record, email is not a good argument for this, since email could not use this data structure (its data structure is *not* a dict, but a list with dict-like features grafted on).

I do think this would be useful, and the generic version (analogous to defaultdict) would seem to make sense.

Ok, let's bikeshed this a bit. What should be the name?
- projectdict?
- normalizedict?
- normdict?
- derivedict?
- transformdict?
- any ideas?

I would say

- transformkeydict


Too bad we can't just add an extra 'transform_key' keyword to defaultdict.

It would be nice to combine the behaviors that defaultdict and the case-insensitive comparisons.

> It would be nice to combine the behaviors that defaultdict and the case-insensitive comparisons.

Any use case? In my experience they are used in completely different
situations. defaultdict mostly to use the writing of some (internal)
algorithms, a case-insensitive dict to store user-visible data.

Just today I was using a defaultdict where the keys are stock symbols. They're case insensitive (at least for this particular application).

In this case I just str.upper everything, but it would be a nice feature to case-preserve the keys that I pre-populate. I care less about the keys from user data.

> In this case I just str.upper everything, but it would be a nice
> feature to case-preserve the keys that I pre-populate. I care less
> about the keys from user data.

Well, stock symbols are what I would call user data :-)

True enough!

I was trying to distinguish keys that I populate with initial values (mostly stock indexes) versus those where I just read values from a user-supplied file. When I populate the index values, I'd like to preserve the case I initially used. When I use user-supplied values, I don't know that the first value I use to populate the defaultdict has any more meaning that the last one I see.

It would just be a nice-to-have feature for which I have a real use case. It's not so critical that I can't work around it.

To the point, however, Eric's example would make use of both the defaultdict portion and the transformkey portion in a single dict.

mappeddict?

Re defaultdict, you could write a dict that does all of these things, called superdict! :-)

coercekeydict

I would like to shorten the proposals to "transformdict" and "coercedict".
(after all, transforming the values would have little sense: you can do it yourself trivially)

FYI os.environ uses something similar: keys and values are encoded and decoded using functions. So any transformation is supported.

http://hg.python.org/cpython/file/eac63e7ceb03/Lib/os.py#l636

On UNIX, the encoder and decoder are os.fsencode() and os.fsdecode() (not exactly, the real functions are more strict on the input type).

On Windows, the encoder converts the key to uppercase.

> FYI os.environ uses something similar: keys and values are encoded and
> decoded using functions. So any transformation is supported.

I don't think this is the same situation. os.environ has bijective
transformations, which don't pose any implementation challenge.

The whole point of a "transformdict" is to allow for multiple keys to
actually map to the same dict entry.

+1 For the general idea
+1 For the more generic approach of which "lowercase" is just one special case

+10 to make this a PEP so that more people have a chance to express their opinion (currently only those who noticed it on the issues mailing list). I find the issue tracker a very bad medium for any kind of brain-storming or bikeshedding.

> +10 to make this a PEP so that more people have a chance to express
> their opinion (currently only those who noticed it on the issues
> mailing list). I find the issue tracker a very bad medium for any kind
> of brain-storming or bikeshedding.

Well, I don't think there is a lot of brainstorming to be done here: we
are talking about a single well-defined functionality. I don't mind
writing a PEP if other people ask for it, but I'd rather spend my time
on more important things.

As for bikeshedding, there is no "good medium" for it, really :-) At
worse we could ask python-dev for their naming contributions (a great
idea, surely).

Indeed.  Although there was apparently some call for it, it doesn't sound from a quick google like defaultdict was deemed to require a PEP.  Presumably the informed audience should be wider than this issue, though.

I also note that defaultdict is implemented via a special method on dict itself (__missing__), and if this one was implemented the same way it would be easy to combine the features.

Precisely what I was thinking.  :)

> I also note that defaultdict is implemented via a special method on
> dict itself (__missing__), and if this one was implemented the same
> way it would be easy to combine the features.

It's not that simple: to remember the original casing you need either a
second container, or to use (original_key, value) tuples as values. Both
approaches have non-trivial repercussions on the implementation of many
methods.

(as a sidenote, you might want a case-insensitive OrderedDict as well, I see no reason to make a special case for defaultdict here)

See also discussion on a topic: http://comments.gmane.org/gmane.comp.python.ideas/18469 .

Proposed names: custom_dict, KeyedDictionary, Dictionary.

It will be confused if this dict will not be compatible with PyDict API. It is possible to add such feature directly into the dict class (I experimented with IdentityDict).

> Proposed names: custom_dict, KeyedDictionary, Dictionary.

Sounds much too vague and un-specific.

> It will be confused if this dict will not be compatible with PyDict
> API.

Why? Many custom dict-like classes aren't.

> It is possible to add such feature directly into the dict class (I
> experimented with IdentityDict).

Can you explain how?

> Why? Many custom dict-like classes aren't.

And this is weird (issue10977).

> Can you explain how?

By patching Objects/dictobject.c of course. I suppose it should require changing about 400 lines of code, a little more than for IdentityDict. When you provides the specification perhaps I will provide a patch.

> By patching Objects/dictobject.c of course.

Am I stupid :-)

> I suppose it should require changing about 400 lines of code, a little
> more than for IdentityDict. When you provides the specification
> perhaps I will provide a patch.

Well, take a look at the code I've pointed to above. I'm curious to know
how you'll integrate it in dictobject.c without slowing down normal dict
objects, and without making them bigger.

> I'm curious to know how you'll integrate it in dictobject.c without slowing down normal dict objects, and without making them bigger.

It of course will make a size of source file bigger, but shouldn't affect a size or performance of normal dicts. A dict object contains dk_lookup. Constructor for keyed dict (subclass of ) should initialize it with specialized function which calls the "key" function and recalculate a hash (yes, with this simple approach a hash will be calculated twice, for original and for transformed keys). Hmm, actually it can be even simpler than for IdentityDict (for which not calculating a hash was important). Also some other methods which relies on dict implementation details (e.g. making a copy of dict) should be modified.

The most cumbersome part is the tests. Unfortunately I lost my tests for IdentityDict (used hg diff without --git). It will be good if your provide complete test suite.

> It of course will make a size of source file bigger, but shouldn't
> affect a size or performance of normal dicts. A dict object contains
> dk_lookup.

You need to keep both the original keys and the transformed keys. It's not only about transforming keys on lookup. Otherwise, yes, it's trivial.

> It will be good if your provide complete test suite.

... Again, take a look at the code I've posted above ...

(now relayed on python-dev)

I'm uploading a pure Python transformdict implementation + tests, for Serhiy's benefits (and others') :-)

With the current patch __repr__() will fail if the untransformed key is unhashable:

>>> d = collections.transformdict(id)
>>> L = [1,2,3]
>>> d[L] = None 
>>> d.keys()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Repos\cpython-dirty\lib\collections\abc.py", line 444, in __repr__
    return '{0.__class__.__name__}({0._mapping!r})'.format(self)
  File "C:\Repos\cpython-dirty\lib\collections\__init__.py", line 944, in __repr__
    self._transform, repr(dict(self)))
TypeError: unhashable type: 'list'

> With the current patch __repr__() will fail if the untransformed key
> is unhashable:

Yeah, the __repr__() implementation will be a bit annoying to get right :-)

FTR this is one message from the previous thread about this: https://mail.python.org/pipermail/python-ideas/2010-June/007332.html

Updated patch: fixes repr(), and retains the first key not the last.

I would *really* like for this to start outside the standard library.
It needs to mature with user feedback before being dumped
in the collections module (which was never intended to be a
giant pile of every collection a person could think of).  

Adding yet more dictionary variants is an example of
way-too-many-ways-to-do-it.

Here is a preliminary C implementation.

Serhiy, any benchmarks for your implementation? Does it slow down regular dicts?

On 09/11/2013 02:39 PM, Tim Delaney wrote on PyDev:
> 
> I would think that retrieving the keys from the dict would return the transformed keys (I'd 
> call them canonical keys).

The more I think about this the more I agree.  A canonicaldict with a key function that simply stored the transformed key and it's value would seem to be a lot simpler:

  - no need to store a separate "presentation" key
  - no confusion about which of the first key/last key seen is stored
  - no mistakes with the "first" key not being added before real data
    and getting the presentation key wrong

Further, in order to store the non-canonical keys a separate list must be kept of the keys to preseed the canonicaldict; if we store the canonical keys a separate list must be kept for presentation purposes -- so worst case scenario we're keeping the same amount of information and best-case scenario the presentation of the keys doesn't matter and we just saved ourselves an extra data structure.

It would be simpler, but it would also be useless for the actual use case for which this issue was opened.

True, but how big a deal is that?

For one, it seems questionable to have the presentation portion of the data be part of the key.

For two, when presentation is important a separate list must be kept anyway to preseed the dict; so just use that list to cycle through the canonicaldict:

--> some_dict = some_function_that_returns_a_conanicaldict()
--> presentation_list = ['IBM','Intel','AMD']
--> for company in presentation_list:
...     key = some_dict.key[company]  # demo purposes only
...     value = some_dict[company]
...     print(key, company, value)
ibm IBM 2172
intel Intel 3210
amd AMD 4399

Ethan, please don't post the same message *both* on the tracker
and on the mailing-list. I'm sure most people here also read
the ML thread.

Right, sorry.

You are conceptualizing this very differently.  In our view, this data structure is for cases where the original key is the most important piece of information (about the keys).  The transformation in the lookup process is entirely in the service of looking up the value paired with that original key when there is more than one possible representation of that key.  It is the original key that is critical when re-serializing the data or otherwise making use of the keys for anything other than lookup.  So this is about making the data structure succinctly model the problem domain, which is what OO is supposed to be good at :)

> R. David Murray added the comment:
> 
> You are conceptualizing this very differently.  In our view, this
> data structure is for cases where the original key is the most
> important piece of information (about the keys).  The transformation
> in the lookup process is entirely in the service of looking up the
> value paired with that original key when there is more than one
> possible representation of that key.  It is the original key that is
> critical when re-serializing the data or otherwise making use of the
> keys for anything other than lookup.  So this is about making the
> data structure succinctly model the problem domain, which is what OO
> is supposed to be good at :)

Thanks for putting it much more convincingly than my python-dev
response :-)

Here is an alternative C implementation. It adds to the dict class support of the __transform__() method. If this method is defined in dict subclass it used to transforming keys. collections.TransformDict is just utilizes this feature as collections.defaultdict utilizes __missing__(). This patch changes twice less C code than previous one (227 vs 474 lines).

> Does it slow down regular dicts?

I were surprized, but yes. The ComplexPythonFunctionCalls test from pybench is 40% slower with ctransformdict.patch (and I still don't known why). With dict__transform__.patch it is only 2% slower. All other pybench tests are approximately equal.

> > Does it slow down regular dicts?
> 
> I were surprized, but yes. The ComplexPythonFunctionCalls test from
> pybench is 40% slower with ctransformdict.patch (and I still don't
> known why). With dict__transform__.patch it is only 2% slower. All
> other pybench tests are approximately equal.

Did you try any other microbenchmarks? Your approach sounds promising.

> Did you try any other microbenchmarks? Your approach sounds promising.

Any microbenchmarks which I tried did not show any interesting. Until I found the cause of slowing down ComplexPythonFunctionCalls I have no idea which tests can be representable.

Of course you can run benchmarks yourself.

Updated patch adding the getitem() method.

Note: I haven't renamed transformdict to TransformDict yet.

Uploading new patch with added transform_func property.

Note that I'm strongly against this name of the getitem() method.

> Georg Brandl added the comment:
> 
> Note that I'm strongly against this name of the getitem() method.

Any suggestion?

Not really. Would "entry" be acceptable instead of "item"?

> Georg Brandl added the comment:
> 
> Not really. Would "entry" be acceptable instead of "item"?

getentry() sounds decent to me, but it loses the parallel to popitem() and items().

Hmm, I didn't consider popitem().  Maybe I'm too paranoid about users confusing __getitem__() and getitem() after all :)

But why not getkey()? Why you need return value too?

> But why not getkey()? Why you need return value too?

Because it's more useful to return both.

Sorry, I don't understand why it's more useful. We need create a tuple and then index it or unpack it and drop one of elements. This only muddles away a time and programmer's attention.

Because most often the time at which you want the original key is the point at which you are about to re-serialize the data...so you need the value too.

I do think getitem is the most natural name for the method.

Oh, could anyone borrow Guido's time machine and rename either __getitem__() to __getvalue__() or items() to entries()?

On 09/17/2013 09:34 AM, R. David Murray wrote:
> 
> R. David Murray added the comment:
> 
> Because most often the time at which you want the original key is the point at which you are about to re-serialize the data...so you need the value too.

I can't think of a case where I'd need (original_key, value) where I
wouldn't also be iterating over items(). Especially so if I'm serializing.

On the other hand, I don't have a use case for the original key, anyway.
So I don't have a strong feeling about this, other than it feels odd
that the answer to the original question (I think on python-dev) "how do
we get the original key back?" is answered by "by giving you the
original key and its value".

On 09/17/2013 10:12 AM, Eric V. Smith wrote:
> On the other hand, I don't have a use case for the original key, anyway.
> So I don't have a strong feeling about this, other than it feels odd
> that the answer to the original question (I think on python-dev) "how do
> we get the original key back?" is answered by "by giving you the
> original key and its value".

I meant: I don't have a use case for finding the original key outside of
iterating over items().

I just want to say thanks for working on this. I also have needed this functionality for various needs in the past. To fulfill my needs, I wrote this implementation:

https://bitbucket.org/jaraco/jaraco.util/src/1ab3e7061f96bc5e179b6b2c46b06d1c20f87129/jaraco/util/dictlib.py?at=default#cl-221

That implementation is used in the irc library for a case-insensitive dict, but using the IRC-specific standard for case insensitivity (https://bitbucket.org/jaraco/irc/src/1576b10dc2923d4d7234319d2d1e11a5080e1f7d/irc/dict.py?at=default#cl-49).

I share this just to add a +1 for the need and to provide additional use cases and implementations for reference.

Raymond, have you had time to look at this?

Antoine, is the PEP ready for review?

> Antoine, is the PEP ready for review?

Well, I think it is. Do you think other points should be addressed in it?
We still have some time.

+1 for this (for Python 3.5, now, I guess). I've just found another place where I'd use it.

Looking at the implementation, one thing surprises me a bit:  I'd expect the KeyError from a 'del' or 'pop' operation to have the untransformed key rather than the transformed key in its .args.

How about '_keys' and '_values' for the slot names, in place of '_original' and '_data'?

Mark, what was the use case you found?

> Mark, what was the use case you found?

It's essentially an IdentityDict, though I've found other more specific transforms useful.

I was writing a tool to find reference cycles between Python objects (we have a customer application that's working in a multithreaded COM environment and has to ensure that COM objects are released on the same types of threads they were created on, so we have to be careful about cyclic garbage and delayed garbage collection).

The graph of Python objects (class 'ObjectGraph') is modelled as a fairly standard directed graph (set of vertices, set of edges, two dictionaries mapping each edge to its head and tail), but of course for this application the dict and set have to be based on object identity rather than normal equality.  Using a TransformDict (and an IdentitySet) lets me write the standard graph algorithms (e.g., for finding strongly connected components) in a natural way, leaving it to the TransformDict and IdentitySet to do the necessary id() conversions under the hood.)

I also have a similar AnnotatedGraph object (a sort of offline version of the ObjectGraph), where the edges and vertices carry additional information and it's convenient to be able to use a lightweight ID rather than an entire vertex or edge as a dictionary key.  Again, using a TransformDict lets one hide the details and present the graph manipulation code readably and naturally.

Some code here, if you're interested:

https://github.com/mdickinson/refcycle/blob/refactor/refcycle/object_graph.py

Caveat: it's work in progress.

[Mark Dickinson]
> It's essentially an IdentityDict, though I've found other 
> more specific transforms useful.

Have any of the applications had use for the part of the API that looks up the original, untransformed key?

Not my applications, no.

3.5 is almost here; Raymond, care to make a ruling?

Yes.

I intend to button this one up before long.

For the record, this is related to PEP 455 (key-transforming dictionary)

The API is simple and well defined, the addition is small, I don't understand what is the problem with this enhancement.

Some refactoring that I'm working on for http.client could use this (currently I have it as part of my patch set). I haven't run into any issues using it and it's definitely useful. Would be nice to get this merged.

FYI, the PEP for this isn't going to be accepted (I'm working on the write-up for the reasons why and will post on python-dev).   That said, it would be great if the code continues to be improved and then posted on the Python Package Index.

I will be interested to see those reasons. Another way to do a similar thing might be using a Key(value, transform) class, somewhat along the lines of Issue 20632, but as a separate class rather than part of the core type system. But I have not thought that idea through very much.

> I will be interested to see those reasons.

+1. Something like what this PEP proposed would be beneficial in a few places throughout the library (header and cookie implementations would definitely benefit rather than having to deal with buggy normalization themselves). It’s unfortunate that this isn’t going to be approved.

I'm also eager to hear what limitations prevented the acceptance. Please do link back here when you've posted.

I have to say, I'm not entirely surprised. In my implementation, I struggled with some cases, and it certainly doesn't feel like a fully safe implementation.

That said, since I mentioned the implementation in jaraco.util earlier, I wanted to announce that those implementations (FoldedCase and FoldedCaseKeyedDict) have been moved to two libraries (jaraco.text and jaraco.collections).

From https://mail.python.org/pipermail/python-dev/2015-May/140003.html
======================================================================
Before the Python 3.5 feature freeze, I should step-up and
formally reject PEP 455 for "Adding a key-transforming
dictionary to collections".

I had completed an involved review effort a long time ago
and I apologize for the delay in making the pronouncement.

What made it a interesting choice from the outset is that the
idea of a "transformation" is an enticing concept that seems
full of possibility.  I spent a good deal of time exploring
what could be done with it but found that it mostly fell short
of its promise.

There were many issues.  Here are some that were at the top:

* Most use cases don't need or want the reverse lookup feature
  (what is wanted is a set of one-way canonicalization functions).
  Those that do would want to have a choice of what is saved
  (first stored, last stored, n most recent, a set of all inputs,
  a list of all inputs, nothing, etc).  In database terms, it
  models a many-to-one table (the canonicalization or
  transformation function) with the one being a primary key into
  another possibly surjective table of two columns (the
  key/value store).  A surjection into another surjection isn't
  inherently reversible in a useful way, nor does it seem to be a
  common way to model data.

* People are creative at coming up with using cases for the TD
  but then find that the resulting code is less clear, slower,
  less intuitive, more memory intensive, and harder to debug than
  just using a plain dict with a function call before the lookup:
  d[func(key)].  It was challenging to find any existing code
  that would be made better by the availability of the TD.

* The TD seems to be all about combining data scrubbing
  (case-folding, unicode canonicalization, type-folding, object
  identity, unit-conversion, or finding a canonical member of an
  equivalence class) with a mapping (looking-up a value for a
  given key).  Those two operations are conceptually orthogonal.
  The former doesn't get easier when hidden behind a mapping API
  and the latter loses the flexibility of choosing your preferred
  mapping (an ordereddict, a persistentdict, a chainmap, etc) and
  the flexibility of establishing your own rules for whether and
  how to do a reverse lookup.


Raymond Hettinger


P.S.  Besides the core conceptual issues listed above, there
are a number of smaller issues with the TD that surfaced
during design review sessions.  In no particular order, here
are a few of the observations:

* It seems to require above average skill to figure-out what
  can be used as a transform function.  It is more
  expert-friendly than beginner friendly.  It takes a little
  while to get used to it.  It wasn't self-evident that
  transformations happen both when a key is stored and again
  when it is looked-up (contrast this with key-functions for
  sorting which are called at most once per key).

* The name, TransformDict, suggests that it might transform the
  value instead of the key or that it might transform the
  dictionary into something else.  The name TransformDict is so
  general that it would be hard to discover when faced with a
  specific problem.  The name also limits perception of what
  could be done with it (i.e. a function that logs accesses
  but doesn't actually change the key).

* The tool doesn't self describe itself well.  Looking at the
  help(), or the __repr__(), or the tooltips did not provide
  much insight or clarity.  The dir() shows many of the
  _abc implementation details rather than the API itself.

* The original key is stored and if you change it, the change
  isn't stored.  The _original dict is private (perhaps to
  reduce the risk of putting the TD in an inconsistent state)
  but this limits access to the stored data.

* The TD is unsuitable for bijections because the API is
  inherently biased with a rich group of operators and methods
  for forward lookup but has only one method for reverse lookup.

* The reverse feature is hard to find (getitem vs __getitem__)
  and its output pair is surprising and a bit awkward to use.
  It provides only one accessor method rather that the full
  dict API that would be given by a second dictionary.  The
  API hides the fact that there are two underlying dictionaries.

* It was surprising that when d[k] failed, it failed with
  transformation exception rather than a KeyError, violating
  the expectations of the calling code (for example, if the
  transformation function is int(), the call d["12"]
  transforms to d[12] and either succeeds in returning a value
  or in raising a KeyError, but the call d["12.0"] fails with
  a TypeError).  The latter issue limits its substitutability
  into existing code that expects real mappings and for
  exposing to end-users as if it were a normal dictionary.

* There were other issues with dict invariants as well and
  these affected substitutability in a sometimes subtle way.
  For example, the TD does not work with __missing__().
  Also, "k in td" does not imply that "k in list(td.keys())".

* The API is at odds with wanting to access the transformations.
  You pay a transformation cost both when storing and when
  looking up, but you can't access the transformed value itself.
  For example, if the transformation is a function that scrubs
  hand entered mailing addresses and puts them into a standard
  format with standard abbreviations, you have no way of getting
  back to the cleaned-up address.

* One design reviewer summarized her thoughts like this:
  "There is a learning curve to be climbed to figure out what
  it does, how to use it, and what the applications [are].
  But, the [working out the same] examplea with plain dicts
  requires only basic knowledge."  -- Patricia