The implementation of nsmallest in heapq contains an optimization for when n is an order of magnitude less than the size of the data, which uses bisect to find the n-smallest elements. This optimization is guarded by a check to ensure that the data iterable has a length method.

This method is then decorated to add support for the key kwarg. The decorator creates a zip object and passes the zip object to the decorated nsmallest. As zip objects are generators, they do not have a __len__ attribute, and the bisect optimization is never used.

The attached patch file detects whether the data passed to the decorator has a length attribute, and if it does, it creates a list with the data before passing it to the decorated nsmallest. This is my first patch, so if I've done something wrong please let me know. Thanks!

Just so you know, in CPython the bisect version of nsmallest() isn't  called at all.  Instead, the C implementation is used and it employs the heap based algorithm rather than the bisect algorithm.

Instead of mucking with code that call nsmallest(), I'll likely just update nsmallest() at some point to use the same algorithm as the code for nlargest().  That with bring the C code and pure Python code fully in-sync.

> This is my first patch, so if I've done something wrong please let me know.
The Python Development Guide helps you with all this and more: http://docs.python.org/devguide/

Thanks for the clarification Raymond!

Raymond said he would make a code change relevant to this at some point, so we should probably leave this issue open until he does.  (Or if he doesn't want the issue open, we can let him close it).

I've to say that I also ran into that by following coverage numbers and as a newbie is not so easy to see what Raymond told. IMHO that information could be put as a comment.

Regards

francis

New changeset 362826298fdb by Raymond Hettinger in branch 'default':
Issue #16098:  Update heapq.nsmallest to use the same algorithm as nlargest.
http://hg.python.org/cpython/rev/362826298fdb

New changeset de6627cf81f4 by Raymond Hettinger in branch '3.3':
Issue #16098:  Update heapq.nsmallest to use the same algorithm as nlargest.
http://hg.python.org/cpython/rev/de6627cf81f4

New changeset aa9ed98a49cb by Raymond Hettinger in branch '2.7':
Issue #16098:  Update heapq.nsmallest to use the same algorithm as nlargest.
http://hg.python.org/cpython/rev/aa9ed98a49cb