Here is a patch which allows pickling of recursive OrderedDicts. Actually it simplifies __reduce__() code.

In additional the patch decreases the size of the pickled data, speeds up pickling and unpickling.

$ ./python -c "import collections, pickle; od = collections.OrderedDict((i, i) for i in range(100000)); print(len(pickle.dumps(od)))"

Without patch: 1536827.
With patch: 737578.

$ ./python -m timeit -s "from pickle import dumps; import collections; od = collections.OrderedDict((i, i) for i in range(100000))"  "dumps(od)"

Without patch: 454 msec.
With patch: 361 msec.

$ ./python -m timeit -s "from pickle import dumps, loads; import collections; od = collections.OrderedDict((i, i) for i in range(100000)); x = dumps(od)"  "loads(x)"

Without patch: 1.41 sec.
With patch: 1.15 sec.

Ping.

The patch looks good to me. So go ahead and submit it. :)

New changeset 56f25569ba86 by Serhiy Storchaka in branch 'default':
Issue #17900: Allowed pickling of recursive OrderedDicts.  Decreased pickled
http://hg.python.org/cpython/rev/56f25569ba86

This patch caused a regression by breaking designed-in support for pickling using PyYAML (note, there was a test for yaml support, but this patch defeated the test).  As a result of the patch, ordering now gets lost during a dump/load roundtrip.

This change should not have been committed without my review (I would have caught the problem instantly).

Can you explain what is broken independently of PyYAML?

Oh, sorry, I did not notice test_yaml_linkage. And in any case this test was broken (should be rv[1][0] instead of rv[1]).

I don't like to revert these changes, and the reversion can break a code written for 3.4 which pickles recursive OrderedDicts. I suggest following solution. PyYAML calls __reduce_ex_(2) or __reduce__(). Therefore we can return to old behavior with protocol < 3 and left current behavior with protocol >= 3. Recursive OrderedDicts will be pickleable with default protocol and with highest protocol, and PyYAML will be happy.

Does this patch look good to your?

I think it is too late to revert this change. Is it worth to add a workaround for PyYAML if it still hasn't fixed serialization of OrderedDicts?

At this point, we can drop it.  PyYAML will just have to deal with it.

Reported PyYAML issue: https://bitbucket.org/xi/pyyaml/issues/61/ordereddict-doesnt-round-trip.