Dubslow added the comment: Regarding the current bug more specifically, I think a few comments would go a long way to helping readers understand the code. And also, I do think the (1, +1, -1) is less readable, simply because it doesn't follow the most common usage patterns of range, where your first version using (0,0,0) (with implicit zeroes of course) is cleaner. It's much more apparent how long the loop is at first glance ("loop once per iter" vs "loop... from what to what? oh, once per iter"). Perhaps not using reversed() but instead setting step=-1 would be a better case to use the off-by-1 variant. Such a version with both proposed changes (which are independent and can be considered or accepted separately) looks like this: def roundrobin(*iterables): "roundrobin('ABC', 'D', 'EF') --> A D E B F C" # This recipe is also called other names like "alternate", "interleave", or # "merge". "zip_flat" would also be an accurate name. # Algorithm: cycle around the __next__ methods of each iterable. when an # iter runs out, remove its __next__ from the cycle. nexts = cycle(iter(it).__next__ for it in iterables) for reduced_len in reversed(range(len(iterables))): try: for next in nexts: yield next() except StopIteration: nexts = cycle(islice(nexts, reduced_len)) # This skips one round of the cycle, starting the next round # without the current iter The last comment is probably the least valuable, though it does point out the slight quirk of how the last line works. I suppose this is the case for having the loop variable be named the actual length, but I think most Python users are accustomed to the last element of a list having index length-1. At any rate, I think the readability gain in the for statement is more than the readability loss in the islice(). ---------- _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Dubslow added the comment: Er, in my first message, make that "(yield from tup for tup in zip_longest(*iters, usefill=False))" ---------- _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Raymond Hettinger added the comment: IMO, George Sakkis's original version is better than the TJR's proposed revision which has three separate adjustments by one and adds to extraneous functions which aren't central to the example. Dubslow's alternative at least avoids the three offsets by one; however, it has the disadvantage that meaning of the loop induction variable is somewhat indirect and creates its own source of confusion. I agree that some algorithmic comment should be added, probably just for the last line. Let's not add any of the listed alternative names to the comments. Each one is either confusing or wrong. The word "merge" already has an established, different meaning as in "sort/merge". For example, heapq.merge('ABC', 'D', 'EF') gives ['A', 'B', 'C', 'D', 'E', 'F']. The word "alternate" tends to mean "toggle back-and-forth" in common usage. The term "zip_flat" has no meaning to me, it has no hits on Google, and the closest match is a recipe on StackOverflow that does something different. And "interleave" is not suggestive of looping back over the iterables until each is exhausted. Also, we may yet use interleave() to mean something else. In contrast, "round robin" has a well established meaning that matches what this recipe does. Until now, not a single reader has ever claimed to not know what it means. https://en.wikipedia.org/wiki/Round-robin_scheduling FWIW, the recipe has several benefit. 1) Give a way to implement round robin iteration without re-inventing the wheel, 2) Demonstrate ways to use cycle() and islice(). 3) Demonstrate useful optimization technique of factoring the try/except out of the for-loop, 4) Demonstrate the useful optimization technique of calling bound methods, and 5) Be concise (I've left longer or more complex recipes for the ASPN cookbook or StackOverflow). Ideally, I would prefer to not add two extra builtin lookups (the recipe is sometime used on short inputs which would be affected by the extra overhead). Also, I prefer the visual weight to be on the central message of tight inner loops that exploit itertools rather than having the visual weight shift to the for-loop which is the least important part. Can you a suggest a concise single line comment that would make the last line clearer about what it is doing? Also, I'm open to changing the name of the "pending" variable but think "current_len" and "reduced_len" are not improvements. ---------- _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Change by Raymond Hettinger : ---------- keywords: +patch pull_requests: +4424 stage: needs patch -> patch review _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Raymond Hettinger added the comment: Serhiy, I think you're focusing on an irrelevant edge case and reading too much into the recipe. You could create an underlying binary tree with O(n) iteration and O(log n) deletion but then that completely misses the point of the itertools recipes and would likely be pointless in the real-world and likely perform worse than the current recipe in common cases unless you wrote a C extension for it. ---------- _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Serhiy Storchaka added the comment: I agree that using binary tree would be excessively here. I just wondering if there is a simple way to get rid of the quadratic complexity. In any case this does not matter much until you work with many hundreds or thousands of iterables. ---------- _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Change by Roundup Robot : ---------- pull_requests: +4425 _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Serhiy Storchaka added the comment: Today I have published a similar recipe on Python-Ideas. It uses popleft/append instead of __getitem__/rotate. def roundrobin(*iterables): "roundrobin('ABC', 'D', 'EF') --> A D E B F C" nexts = deque(iter(it).__next__ for it in iterables) popleft = nexts.popleft append = nexts.append while nexts: next = popleft() try: yield next() except StopIteration: pass else: append(next) It is faster (10-25%) in all microbenchmarks that I did (Steven's benchmarks for small number of iterables and my examples for large number of iterables). ---------- _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Serhiy Storchaka added the comment: But why use less efficient code? Is my code worse? ---------- _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________

Raymond Hettinger added the comment: New changeset 0858495a50e19defde786a4ec050ec182e920f46 by Raymond Hettinger in branch 'master': bpo-32099 Add deque variant of roundrobin() recipe (#4497) https://github.com/python/cpython/commit/0858495a50e19defde786a4ec050ec182e9... ---------- _______________________________________ Python tracker https://bugs.python.org/issue32099 _______________________________________