iter_bucket fails on AWS Lambda when worker threads is not 1

[rodjunger]

I did some research before opening this issue so it's more of a "information" kind of issue.
AWS Lambda environments do not support multiprocessing.Queue or multiprocessing.Pool (which is used by smart open) so calling iter_bucket with default arguments results in this exception:

Traceback (most recent call last):
File "/var/task/handler.py", line 33, in hello
for data in zs.stream():
File "/var/task/zipstream/zipstream.py", line 298, in stream
for idx, source in enumerate(self._source_of_files):
File "/var/task/handler.py", line 11, in stream_files_from_s3_bucket
for key, content in s3_iter_bucket(bucket_name, prefix=prefix):
File "/var/task/smart_open/s3.py", line 603, in iter_bucket
with _create_process_pool(processes=workers) as pool:
File "/var/lang/lib/python3.6/contextlib.py", line 81, in enter
return next(self.gen)
File "/var/task/smart_open/s3.py", line 694, in _create_process_pool
pool = multiprocessing.pool.Pool(processes=processes)
File "/var/lang/lib/python3.6/multiprocessing/pool.py", line 156, in init
self._setup_queues()
File "/var/lang/lib/python3.6/multiprocessing/pool.py", line 249, in _setup_queues
self._inqueue = self._ctx.SimpleQueue()
File "/var/lang/lib/python3.6/multiprocessing/context.py", line 112, in SimpleQueue
return SimpleQueue(ctx=self.get_context())
File "/var/lang/lib/python3.6/multiprocessing/queues.py", line 315, in init
self._rlock = ctx.Lock()
File "/var/lang/lib/python3.6/multiprocessing/context.py", line 67, in Lock
return Lock(ctx=self.get_context())
File "/var/lang/lib/python3.6/multiprocessing/synchronize.py", line 162, in init
SemLock.init(self, SEMAPHORE, 1, 1, ctx=ctx)
File "/var/lang/lib/python3.6/multiprocessing/synchronize.py", line 59, in init
unlink_now)
OSError: [Errno 38] Function not implemented

I found this fix from another project which seems simple enough for me to try to apply to this project too

[piskvorky]

That might work. Which version of Python is needed for concurrent.futures?

[rodjunger]

That might work. Which version of Python is needed for concurrent.futures?

It does work locally, going to test it on lambda in a second. Here's the code i "wrote" (basically a copy of the commit i linked with some adaptations) :

with concurrent.futures.ThreadPoolExecutor(max_workers=workers) as executor:
        futures = [executor.submit(download_key, key) for key in key_iterator]
        for key_no, future in enumerate(concurrent.futures.as_completed(futures)):
            (key, content) = future.result()

            if True or key_no % 1000 == 0:
                logger.info(
                    "yielding key #%i: %s, size %i (total %.1fMB)", key_no, key, len(content), total_size / 1024.0 ** 2
                )

            yield key, content

            total_size += len(content)

            if key_limit is not None and key_no + 1 >= key_limit:
                # we were asked to output only a limited number of keys => we're done
                break

concurrent.futures was added on python 3.2 according to https://docs.python.org/3/library/concurrent.futures.html

[piskvorky]

smart_open supports 2.7 too, so that wouldn't work out-of-the-box. Although we are considering dropping py2 support in our open source projects.

@mpenkov WDYT? According to https://pypistats.org/packages/smart-open , we still get more py2 downloads than py3 downloads. But god knows how many of these are bots / noise…

"Support iter_buckets on AWS lambda" doesn't seem a strong enough argument for us to drop py2. So I'm -1 on dropping py2, unless there are additional reasons to switch.

[rodjunger]

Yeah that doesn't seem like a worthy reason to completely drop python 2 support, but there's this: https://pypi.org/project/futures/

[mpenkov]

It sounds like the current smart_open version with default arguments does not work under Lambda at all, regardless of the Python version you're using. The fix suggested by @rodjunger improves that situation for Py3.2+ users only. As long as nothing changes for Py2.7 users (they continue to see OSError: [Errno 38] Function not implemented when using default arguments under Lambda), I think it's a worthwhile fix.

@rodjunger Can you make a PR that fixes the problem if concurrent.futures is available, but leaves everything as-is if it isn't?

@piskvorky I think we'll have to drop Py2.7 support eventually, for the same reasons as in gensim. Our Py2.7 users can continue to download smart_open<2.0. We'll develop new features for smart_open 2.0 and beyond, but those will support Py3.5 only (same as gensim).

[rodjunger]

@mpenkov Yeah I can do that, I'll submit the PR this week if possible.

[derpferd]

@mpenkov Assumed that since you marked this as Hackoberfest and @rodjunger hasn't respond that this needed work. I submitted a PR for the work @rodjunger said they were going to do.
Let me know what you think.

[mpenkov]

Closed via #368