werkzeug.formparser is really slow with large binary uploads

[sekrause]

When I perform a multipart/form-data upload of any large binary file in Flask, those uploads are very easily CPU bound (with Python consuming 100% CPU) instead of I/O bound on any reasonably fast network connection.

A little bit of CPU profiling reveals that almost all CPU time during these uploads is spent in werkzeug.formparser.MultiPartParser.parse_parts(). The reason this that the method parse_lines() yields a lot of very small chunks, sometimes even just single bytes:

# we have something in the buffer from the last iteration.
# this is usually a newline delimiter.
if buf:
    yield _cont, buf
    buf = b''

So parse_parts() goes through a lot of small iterations (more than 2 million for a 100 MB file) processing single "lines", always writing just very short chunks or even single bytes into the output stream. This adds a lot of overhead slowing down those whole process and making it CPU bound very quickly.

A quick test shows that a speed-up is very easily possible by first collecting the data in a bytearray in parse_lines() and only yielding that data back into parse_parts() when self.buffer_size is exceeded. Something like this:

buf = b''
collect = bytearray()
for line in iterator:
    if not line:
        self.fail('unexpected end of stream')

    if line[:2] == b'--':
        terminator = line.rstrip()
        if terminator in (next_part, last_part):
            # yield remaining collected data
            if collect:
                yield _cont, collect
            break

    if transfer_encoding is not None:
        if transfer_encoding == 'base64':
            transfer_encoding = 'base64_codec'
        try:
            line = codecs.decode(line, transfer_encoding)
        except Exception:
            self.fail('could not decode transfer encoded chunk')

    # we have something in the buffer from the last iteration.
    # this is usually a newline delimiter.
    if buf:
        collect += buf
        buf = b''

    # If the line ends with windows CRLF we write everything except
    # the last two bytes.  In all other cases however we write
    # everything except the last byte.  If it was a newline, that's
    # fine, otherwise it does not matter because we will write it
    # the next iteration.  this ensures we do not write the
    # final newline into the stream.  That way we do not have to
    # truncate the stream.  However we do have to make sure that
    # if something else than a newline is in there we write it
    # out.
    if line[-2:] == b'\r\n':
        buf = b'\r\n'
        cutoff = -2
    else:
        buf = line[-1:]
        cutoff = -1

    collect += line[:cutoff]

    if len(collect) >= self.buffer_size:
        yield _cont, collect
        collect.clear()

This change alone reduces the upload time for my 34 MB test file from 4200 ms to around 1100 ms over localhost on my machine, that's almost a 4X increase in performance. All tests are done on Windows (64-bit Python 3.4), I'm not sure if it's as much of a problem on Linux.

It's still mostly CPU bound, so I'm sure there is even more potential for optimization. I think I'll look into it when I find a bit more time.

[languanghao]

I also have same problem, when I upload an iso file(200m), the first call to request.form will take 7s

[RonnyPfannschmidt]

2 things seem interesting for further optimization - experimenting with cython, and experimenting with interpreting the content-site headers for smarter mime message parsing

(no need to scan for lines if you know the content-length of a sub-message)

[lnielsen]

Just a quick note, that if you stream the file directly in the request body (i.e. no application/multipart-formdata), you completely bypass the form parser and read the file directly from request.stream.

[carbn]

I have the same issue with slow upload speeds with multipart uploads when using jQuery-File-Upload's chunked upload method. When using small chunks (~10MB), the transfer speed jumps between 0 and 12MB/s while the network and server are fully capable of speeds over 50MB/s. The slowdown is caused by the cpu bound multipart parsing which takes about the same time as the actual upload. Sadly, using streaming uploads to bypass the multipart parsing is not really an option as I must support iOS devices that can't do streaming in the background.

The patch provided by @sekrause looks nice but doesn't work in python 2.7.

[cuibonobo]

@carbn: I was able to get the patch to work in Python 2.7 by changing the last line to collect = bytearray(). This just creates a new bytearray instead of clearing the existing one.

[carbn]

@cuibonobo: That's the first thing I changed but still had another error. I can't check the working patch at the moment, but IIRC the yields had to be changed from yield _cont, collect to yield _cont, str(collect). This allowed the code to be tested and the patch yielded about 30% increase in the multipart processing speed. It's a nice speedup, but the performance is still pretty bad.

[sekrause]

A little further investigation shows that werkzeug.wsgi.make_line_iter is already too much of a bottleneck to really be able to optimize parse_lines(). Look at this Python 3 test script:

import io
import time
from werkzeug.wsgi import make_line_iter

filename = 'test.bin' # Large binary file
lines = 0

# load a large binary file into memory
with open(filename, 'rb') as f:
    data = f.read()
    stream = io.BytesIO(data)
    filesize = len(data) / 2**20 # MB

start = time.perf_counter()
for _ in make_line_iter(stream):
    lines += 1
stop = time.perf_counter()
delta = stop - start

print('File size: %.2f MB' % filesize)
print('Time: %.1f seconds' % delta)
print('Read speed: %.2f MB/s' % (filesize / delta))
print('Number of lines yielded by make_line_iter: %d' % lines)

For a 923 MB video file with Python 3.5 the output look something like this on my laptop:

File size: 926.89 MB
Time: 20.6 seconds
Read speed: 44.97 MB/s
Number of lines yielded by make_line_iter: 7562905

So even if you apply my optimization above and optimize it further until perfection you'll still be limited to ~45 MB/s for large binary uploads simply because make_line_iter can't give you the data fast enough and you'll be doing 7.5 million iterations for 923 MB of data in your loop that checks for the boundary.

I guess the only great optimization will be to completely replace parse_lines() with something else. A possible solution that comes to mind is to read a reasonably large chunk of the stream into memory then use string.find() (or bytes.find() in Python 3) to check if the boundary is in the chunk. In Python find() is a highly optimized string search algorithm written in C, so that should give you some performance. You would just have to take care of the case where the boundary might be right between two chunks.

[sdizazzo]

I wanted to mention doing the parsing on the stream in chunks as it is received. @siddhantgoel wrote this great little parser for us. It's working great for me. https://github.com/siddhantgoel/streaming-form-data

[lambdaq]

I guess the only great optimization will be to completely replace parse_lines()

+1 for this.

I am writing a bridge to stream user's upload directly to S3 without any intermediate temp files, possibly with backpressure, and I find werkzeug and flask situation frustrating. You can't move data directly between two pipes.

[davidism]

@lambdaq I agree it's a problem that needs to be fixed. If this is important to you, I'd be happy to review a patch changing the behavior.

[lnielsen]

@lambdaq Note that if you just stream data directly in the request body and use application/octet-stream then the form parser doesn't kick in at all and you can use request.stream (i.e. no temp files etc).

The only problem we had is the werkzeug form parser is eagerly checking content length against the allowed max content length before knowing if it should actually parse the request body.

This prevents you from setting max content length on normal form data, but also allow very large file uploads.

We fixed it by reordering the check the function a bit. Not sure if it makes sense to provide this upstream as some apps might rely on the existing behaviour.