This question is old, but I landed here when searching for a "deep merge" solution. The answers above inspired what follows. I ended up writing my own because there were bugs in all the versions I tested. The critical point missed was, at some arbitrary depth of the two input dicts, for some key, k, the decision tree when d[k] or u[k] is not a dict was faulty.
Also, this solution does not require recursion, which is more symmetric with how dict.update() works, and returns None.
import collections
def deep_merge(d, u):
"""Do a deep merge of one dict into another.
This will update d with values in u, but will not delete keys in d
not found in u at some arbitrary depth of d. That is, u is deeply
merged into d.
Args -
d, u: dicts
Note: this is destructive to d, but not u.
Returns: None
"""
stack = [(d,u)]
while stack:
d,u = stack.pop(0)
for k,v in u.items():
if not isinstance(v, collections.Mapping):
# u[k] is not a dict, nothing to merge, so just set it,
# regardless if d[k] *was* a dict
d[k] = v
else:
# note: u[k] is a dict
if k not in d:
# add new key into d
d[k] = v
elif not isinstance(d[k], collections.Mapping):
# d[k] is not a dict, so just set it to u[k],
# overriding whatever it was
d[k] = v
else:
# both d[k] and u[k] are dicts, push them on the stack
# to merge
stack.append((d[k], v))