Should we cache some small properties? #3514

Activity

1. shoyer commented on Nov 12, 2019

   

   shoyer

   on Nov 12, 2019

   Member

   More actions

   I think this would be totally fine to add. A variant on cache_readonly appears in Python 3.8 standard library as functools.cached_property. I think we could backport that pretty easily:
   https://github.com/python/cpython/blob/d360346640e19231032b072216195484fa2450b4/Lib/functools.py#L924
2. max-sixty commented on Nov 12, 2019

   

   max-sixty

   on Nov 12, 2019

   Last edited by max-sixty

   CollaboratorAuthor

   More actions

   Great, I didn't know about that, thanks

   It does remind me, though, that I'm not sure it's possible given we're using __slots__, at least without adding an entry to __slots__...
   https://github.com/python/cpython/blob/d360346640e19231032b072216195484fa2450b4/Lib/functools.py#L954-L958
3. crusaderky commented on Nov 12, 2019

   

   crusaderky

   on Nov 12, 2019

   Contributor

   More actions

   By reading the implementation of cachedproperty, it needs a __dict__. It should be straightforward to write a variant that uses slots though.

   👍React with 👍1Reacted by Maximilian Roos
4. max-sixty commented on Nov 13, 2019

   

   max-sixty

   on Nov 13, 2019

   CollaboratorAuthor

   More actions

   By reading the implementation of cachedproperty, it needs a __dict__. It should be straightforward to write a variant that uses slots though.

   Is there a way of doing that which doesn't nullify the benefits of slots? i.e. if we add a _cache_dict to the class, is that the same overhead as having a __dict__?
5. crusaderky commented on Nov 14, 2019

   

   crusaderky

   on Nov 14, 2019

   Contributor

   More actions

   @max-sixty afraid so. But as I said it should be straightforward to use a variant that instead of self.__dict__[self.name] uses getattr/setattr.
6. crusaderky commented on Nov 14, 2019

   

   crusaderky

   on Nov 14, 2019

   Contributor

   More actions

   I prefer pandas.utils.cache_readonly to the Python 3.8 cachedproperty. Besides not needing a dict, it is also written in cython, which should make a substantial difference.

   max-sixty could you post your benchmark where you measure 150us? I tried caching that property with @cache_readonly and I only get a boost of 7us.

   import xarray
   ds = xarray.Dataset({'d': ('x', [1, 2]), 'x': [10, 20]})
   ds
   ds.to_netcdf("foo.nc")
   ds.close()
   ds = xarray.open_dataset("foo.nc")
   
   %timeit ds.isel(x=[0])

   With @property: 187us
   With @cache_readonly: 180us
7. crusaderky commented on Nov 14, 2019

   

   crusaderky

   on Nov 14, 2019

   Contributor

   More actions

   %%prun -s cumulative
   for _ in range(10000):
       ds.isel(x=[0])

   Output (uncached):

      ncalls  tottime  percall  cumtime  percall filename:lineno(function)
       10000    0.148    0.000    3.317    0.000 dataset.py:1854(isel)
       [...]
       60000    0.092    0.000    0.122    0.000 indexing.py:535(shape)
   
8. dcherian

   added

   topic-performance

   on Nov 16, 2019
9. max-sixty

   closed this as completedon Aug 28, 2024