Micro-optimisation: Special-case dictionaries for CaseInsensitiveMapping.

[Keryn Knight]

Currently, all data given to HttpHeaders and ResponseHeaders go through CaseInsensitiveMapping._unpack_items to unwind in case it's not a dictionary.

In the case of HttpHeaders though, it's always a dictionary (and I've got a plan for ResponseHeaders in general, as a separate patch, should this one go OK as a pre-requisite), so for the common cases of CaseInsensitiveMapping being used, we can optimise for the expected use-case by introducing an additional isinstance check to avoid going through _unpack_items, __instancecheck__ and _abc._abc_instancecheck

The change itself (I'll add a PR after I get a ticket number) is thus:

-        self._store = {k.lower(): (k, v) for k, v in self._unpack_items(data)}
+        data = data.items() if isinstance(data, dict) else self._unpack_items(data)
+        self._store = {k.lower(): (k, v) for k, v in data}

The isinstance check will add roughly 70ns (for me) to non-dict usages of CaseInsensitiveMapping (and it's subclasses), but will save
roughly 500ns (for me) when using dictionaries.

Baseline, main as of b626c5a9798b045b655d085d59efdd60b5d7a0e3:

In [1]: from django.http.request import HttpHeaders

In [2]: from django.utils.datastructures import CaseInsensitiveMapping

In [3]: %timeit CaseInsensitiveMapping({'Name': 'Jane'})
1.4 µs ± 7.94 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

In [4]: %timeit HttpHeaders({"HTTP_EXAMPLE": 1})
2.79 µs ± 20.1 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

In [5]: %timeit HttpHeaders({"CONTENT_LENGTH": 1})
2.56 µs ± 40 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

In [6]: %prun for _ in range(100000): CaseInsensitiveMapping({"Name": "Jane"})
   900009 function calls (900007 primitive calls) in 0.356 seconds
   Ordered by: internal time
   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
   100000    0.073    0.000    0.231    0.000 datastructures.py:306(<dictcomp>)
   200000    0.066    0.000    0.147    0.000 datastructures.py:328(_unpack_items)
        1    0.063    0.063    0.356    0.356 <string>:1(<module>)
   100000    0.063    0.000    0.293    0.000 datastructures.py:305(__init__)
   100000    0.030    0.000    0.072    0.000 {built-in method builtins.isinstance}
   100000    0.022    0.000    0.041    0.000 abc.py:117(__instancecheck__)
   100000    0.019    0.000    0.020    0.000 {built-in method _abc._abc_instancecheck}
   100000    0.011    0.000    0.011    0.000 {method 'lower' of 'str' objects}
   100000    0.010    0.000    0.010    0.000 {method 'items' of 'dict' objects}
        1    0.000    0.000    0.356    0.356 {built-in method builtins.exec}
      2/1    0.000    0.000    0.000    0.000 {built-in method _abc._abc_subclasscheck}
      2/1    0.000    0.000    0.000    0.000 abc.py:121(__subclasscheck__)
        2    0.000    0.000    0.000    0.000 _collections_abc.py:409(__subclasshook__)
        1    0.000    0.000    0.000    0.000 {method 'disable' of '_lsprof.Profiler' objects}

After adding the patch suggested above:

In [1]: from django.http.request import HttpHeaders

In [2]: from django.utils.datastructures import CaseInsensitiveMapping

In [3]: %timeit CaseInsensitiveMapping({'Name': 'Jane'})
862 ns ± 9.4 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

In [4]: %timeit HttpHeaders({"HTTP_EXAMPLE": 1})
2.05 µs ± 31 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

In [5]: %timeit HttpHeaders({"CONTENT_LENGTH": 1})
1.98 µs ± 18.1 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

In [6]: %prun for _ in range(100000): CaseInsensitiveMapping({"Name": "Jane"})
   500003 function calls in 0.204 seconds
   Ordered by: internal time
   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
   100000    0.084    0.000    0.148    0.000 datastructures.py:305(__init__)
        1    0.056    0.056    0.204    0.204 <string>:1(<module>)
   100000    0.038    0.000    0.047    0.000 datastructures.py:307(<dictcomp>)
   100000    0.009    0.000    0.009    0.000 {method 'lower' of 'str' objects}
   100000    0.009    0.000    0.009    0.000 {method 'items' of 'dict' objects}
   100000    0.008    0.000    0.008    0.000 {built-in method builtins.isinstance}
        1    0.000    0.000    0.204    0.204 {built-in method builtins.exec}
        1    0.000    0.000    0.000    0.000 {method 'disable' of '_lsprof.Profiler' objects}

As you can see from the cProfile timing + number of function calls, we've saved a not insubstantial amount of work from being done, and that looks consistent over both classes and both likely paths through an HTTP Header.

Tests all pass for me locally...

[Carlton Gibson]

OK, I'll accept pending discussion on the PR. Thanks.

[Mariusz Felisiak]

​PR

[GitHub <noreply@…>]

In 3de787a7:

Fixed #33532 -- Optimized CaseInsensitiveMapping instantiation for dicts.

Internal usages of this class (e.g. HttpHeaders) provide it with a dict,
so testing for that type first avoids the cost of going through the
potential instancecheck + _abc_instancecheck to establish it's
a Mapping.

Co-authored-by: Nick Pope <nick@…>