I've done the implementation for both non-chunked and chunked reads. I haven't benchmarked chunked reads because I don't currently have a convenient way to generate a high-bandwidth chunked stream, but I don't see any reason that it shouldn't enjoy the same benefits that the non-chunked case does. I've used the benchmark attached to the bpo bug to verify that performance now matches the unsized read case.
Automerge-Triggered-By: @methane
* Move socket related functions from test.support to socket_helper.
* Import socket, nntplib and urllib.error lazily in transient_internet().
* Remove importing multiprocess.
* bpo-38216: Allow bypassing input validation
* bpo-36274: Also allow the URL encoding to be overridden.
* bpo-38216, bpo-36274: Add tests demonstrating a hook for overriding validation, test demonstrating override encoding, and a test to capture expectation of the interface for the URL.
* Call with skip_host to avoid tripping on the host checking in the URL.
* Remove obsolete comment.
* Make _prepare_path_encoding its own attr.
This makes overriding just that simpler.
Also, don't use the := operator to make backporting easier.
* Add a news entry.
* _prepare_path_encoding -> _encode_prepared_path()
* Once again separate the path validation and request encoding, drastically simplifying the behavior. Drop the guarantee that all processing happens in _prepare_path.
Modern Linux distros such as Debian Buster have default OpenSSL system
configurations that reject connections to servers with weak certificates
by default. This causes our test suite run with external networking
resources enabled to skip these tests when they encounter such a failure.
Fixing the network servers is a separate issue.
with debuglevel=1 only the header keys got printed. With
this change the header values get printed as well and the single
header entries get '\n' as a separator.
blocksize was hardcoded to 8192, preventing efficient upload when using
file-like body. Add blocksize argument to __init__, so users can
configure the blocksize to fit their needs.
I tested this uploading data from /dev/zero to a web server dropping the
received data, to test the overhead of the HTTPConnection.send() with a
file-like object.
Here is an example 10g upload with the default buffer size (8192):
$ time ~/src/cpython/release/python upload-httplib.py 10 https://localhost:8000/
Uploaded 10.00g in 17.53 seconds (584.00m/s)
real 0m17.574s
user 0m8.887s
sys 0m5.971s
Same with 512k blocksize:
$ time ~/src/cpython/release/python upload-httplib.py 10 https://localhost:8000/
Uploaded 10.00g in 6.60 seconds (1551.15m/s)
real 0m6.641s
user 0m3.426s
sys 0m2.162s
In real world usage the difference will be smaller, depending on the
local and remote storage and the network.
See https://github.com/nirs/http-bench for more info.
The previous attempt to determine the file’s Content-Length gave a false
positive for pipes on Windows.
Also, drop the special case for sending zero-length iterable bodies.
When the body object is a file, its size is no longer determined with
fstat(), since that can report the wrong result (e.g. reading from a pipe).
Instead, determine the size using seek(), or fall back to chunked encoding
for unseekable files.
Also, change the logic for detecting text files to check for TextIOBase
inheritance, rather than inspecting the “mode” attribute, which may not
exist (e.g. BytesIO and StringIO). The Content-Length for text files is no
longer determined ahead of time, because the original logic could have been
wrong depending on the codec and newline translation settings.
Patch by Demian Brecht and Rolf Krahl, with a few tweaks by me.
The @reap_threads decorator made the test wait (for up to 1 s) until
background threads have finished. Calling join() with a timeout should be
equivalent.
* No longer attempts to close already freed socket file descriptor
* Use socket object to be compatible with Windows
* Do not use a timeout to avoid complication with non-blocking mode
* Use internal localhost server rather than depending on a third party
* Avoid trouble with buffered HTTP data by testing tunnelled CONNECT data
This changeset does two things: introduces a new RemoteDisconnected exception
(that subclasses ConnectionResetError and BadStatusLine) so that a remote
server disconnection can be detected by client code (and provides a better
error message for debugging purposes), and ensures that the client socket is
closed if a ConnectionError happens, so that the automatic re-connection code
can work if the application handles the error and continues on.
Tests are added that confirm that a connection is re-used or not re-used
as appropriate to the various combinations of protocol version and headers.
Patch by Martin Panter, reviewed by Demian Brecht. (Tweaked only slightly by
me.)
Some http servers will reject PUT, POST, and PATCH requests if they
do not have a Content-Length header.
Patch by James Rutherford, with additional cleaning up of the
'request' documentation by me.
Angelin BOOZ
Victor Stinner
AMIR
Hai Shi
Bruce Merry
Serhiy Storchaka
Ross
Ashwin Ramaswami
Dong-hee Na
Jason R. Coombs
Raymond Hettinger
Pablo Galindo
Christian Heimes
Gregory P. Smith
Giampaolo Rodola
Matt Houglum
Marco Strigl
Nir Soffer
Vitor Pereira
Antoine Pitrou
Martin Panter
R David Murray
Terry Jan Reedy
Victor Stinner
Berker Peksag
Benjamin Peterson