Hacker News Reader

Comment by tcoder | Hacker News Reader

Namedtuple CPU-speed performance for some common operations is TERRIBLE compared to dictionaries, at least on 2.7. Orders of magnitude difference. So bad that it really matters. It's shocking how bad the implementors got this.

View on HN

namedtuple is just a thin wrapper around tuple. It literally constructs a class definition that sub-classes tuple as a string and then executes it. You can see the string template that it uses here[1]. If you're interested in something like namedtuple there are other[2] things[3] you can use depending on your use-case.

[1] https://github.com/python/cpython/blob/master/Lib/collection...

[2] http://stackoverflow.com/a/2648186

[3] https://pypi.python.org/pypi/frozendict/

It's really weird that immutable tuples take a performance hit, compared to mutable lists (almost 50% on access!?!).

That's because Python 2 has a special case for lists[1]. This does not exist in Python 3[2] (and therefore neither does the performance difference).

[1] https://github.com/python/cpython/blob/2.7/Python/ceval.c#L1...

[2] https://github.com/python/cpython/blob/master/Python/ceval.c...

It looks like list access got slower as a result though.

Curious what the operations you're talking about are. Get?

Pickling, for one.

  >>> nt = namedtuple('nt', 'a b')
  >>> nt10k = [nt(1, 2) for i in range(10000)]
  >>> dict10k = [{'a':1, 'b':2} for i in range(10000)]
  >>> timeit pickle.dumps(nt10k)
  10 loops, best of 3: 26.3 ms per loop
  >>> timeit pickle.dumps(dict10k)
  100 loops, best of 3: 2.49 ms per loop

If you look at the the StackOverflow in my sibling comment to yours, running 'obj.attrname' on a namedtuple takes longer because it needs to translate 'attrname' to an integer index and then run (e.g.) 'obj[0]'. Outside of that, I'm not sure.

Reply on news.ycombinator.com