This is a simple enhancement that may reduce memory consumption by a bit (unfortunately, this is difficult to measure using standard tools).

Dict has a special case for string keys. Is set have such optimization?

> Dict has a special case for string keys. Is set have such optimization?

Yes (see set_lookkey_unicode).

Note that the resizing heuristic is slightly different for sets and dicts, so a small-to-middle-size dict can sometimesbe smaller than a set of the same len().

However, for large dicts (>= 50000) it seems the corresponding set is most always 66% smaller.

If sets need to be hacked-up to accommodate this, I would like to draft the patch for it that I think will be cleaner.

FWIW, I'm dubious that there will be any benefit from this at all.  The savings of one-pointer is the dictionary is likely to be insignificant compared to the size of the string object themselves.

Also, I think sets conceptually are not the right choice of data structure because they are primarily about membership testing and not about looking up values.   Incorporating this kind of hack will make my other set optimization efforts more difficult.

I agree the benefit is likely to be small and very minor. Victor has experience measuring memory consumption of various programs, I would like to know about his measurements.

This exact issue was discussed in #19187 and #7224 many years ago.

The first reference above should have been to #1507011.

Raymond> FWIW, I'm dubious that there will be any benefit from this at all.  The savings of one-pointer is the dictionary is likely to be insignificant compared to the size of the string object themselves.

As I wrote in python-dev, the dictionary is usually the largest memory block, at least at Python startup. The dictionary (without counting the string, just the dict) is between 192 KB and 1.5 MB on x86_64.

In the implementation of the PEP 454, issue #18874, I added a function to get the length and size of the dictionary of Unicode interned strings.

Objects/unicodeobject.c:

PyObject*
_PyUnicode_GetInterned(void)
{
    return interned;
}

tracemalloc.get_unicode_interned():

http://hg.python.org/features/tracemalloc/file/b797779940a5/Modules/_tracemalloc.c#l4606

You can use this function to see how many KB are saved. In embedded systems, every byte of memory counts :-)

> In embedded systems, every byte of memory counts 

It is not just embedded systems.  The range 192 KB to 1.5 MB is where typical L2 cache sizes are these days.  I would expect that the intern dictionary is accessed very often and much more often than the actual strings.  In theory, if it fits in L2 cache, the performance will be much higher than if it is even slightly over.

This said, when I was looking at this 3-4 years ago, I did not see measurable improvements in my programs.

I afraid that adding new parameter to the set_contains_key function will slow down other set operations. It will be better return a key as the result of the function (NULL if not found).

Based on the discussion so far, I'm going to close this one.  I don't think it is an appropriate use of sets (they are about membership testing, not about looking up values). 

The upside is minimal.  The downside is hacking up the set implementation and making it more difficult to implement some of the other alterations I'm planning for sets -- I want to tune the performance to optimize set operations without worrying about slowing down all the rest of Python because of interned strings.  Martin's reason's for rejecting this previously still hold true.