File test.py is#!/usr/bin/env pythonimport urllib2print urllib2.urlopen('ftp://ftp.ripe.net/pub/stats/ripencc/delegated-ripencc-extended-latest').read()When I issuepython test.py > out.txt, I get file about 100KB in size, the beginning of the actual file. I repeated it a hundred times, and almost every time I get 98305 byte file, and a couple of times a 49153 bytes or 188417 bytes file.When I replace urllib2 with urllib in test.py, I get full (4 MB) file.I have Ubuntu 12.04 64-bit, Python 2.7.3 (from default Ubuntu repository, up-to-date as of 4-june-2012) and slow, 64KB/s, Internet connection.However, I asked my friend with Windows and faster connection to check it, and he got partial download as well, while he had another size of partial file (50109 bytes). I do not know his OS ant Python versions.

The same problem exists in Python3.

That's surprising! I shall test it with http debug mode and see what's happening.

I've tested this on head, and the issue appears to be buggy ftp code in python.From the attached tcpdump for fetching delegated-ripencc-20120706:12:57:19.933607 IP myhost.39627 > ftp.ripe.net.ftp: Flags [.], ack 511, win 115, options [nop,nop,TS val 129353190 ecr1632444059], length 012:57:19.934853 IP myhost.39627 > ftp.ripe.net.ftp: Flags [F.], seq 97, ack 511, win 115, options [nop,nop,TS val 129353191 ecr1632444059], length 0and a bit later:12:57:20.043701 IP ftp.ripe.net.42707 > myhost.50818: Flags [.], seq 46337:47785, ack 1, win 227, options [nop,nop,TS val2552550247 ecr 129353204], length 144812:57:20.043717 IP myhost.50818 > ftp.ripe.net.42707: Flags [.], ack 47785, win 353, options [nop,nop,TS val 129353218 ecr2552550247], length 012:57:20.043816 IP ftp.ripe.net.42707 > myhost.50818: Flags [FP.], seq 47785:49153, ack 1, win 227, options [nop,nop,TS val2552550247 ecr 129353204], length 136812:57:20.043992 IP myhost.50818 > ftp.ripe.net.42707: Flags [F.], seq 1, ack 49154, win 376, options [nop,nop,TS val 129353218 ecr2552550247], length 012:57:20.094067 IP ftp.ripe.net.42707 > myhost.50818: Flags [.], ack 2, win 227, options [nop,nop,TS val2552550299 ecr 129353218], length 0As you can see we're sending a FIN without sending a close command to the control connection, and in response the server stops sending data about 49k in. Per RFC 959 section 2.3: "The server may abort data transfer if the control connections are closed without command." so this is acceptable behaviour on the part of the server, and means we need to keep the control connection open for longer.

More particularly, the ftpwrapper's ftp member is being GCed sometime after FtpHandler.ftp_open returns.

Looking into this. It seems that it doesn't happen for all servers, I can download large files reliably from other sources.I'll make another wireshark recording to get more details for me to analyze.

> I'll make another wireshark recording to get more details for me to analyze.Thank you! That will be useful. Please test it against 3.x version as it has seen cleanups recently.

This is actually the same problem as#18879.Changing the sample to keep a reference to the addinfourl object avoids this issue.This is even worse than#18879 in the sense that the error goes undetected and just leaves you with partial data.Looking at the solution in#18879 I think we can reuse that, maybe even better by refactoring that to a common file proxy object.

I wasn't able to come up with a good testcase. :(I tried similar approaches as in#18879 but I wasn't able to make them trigger the behaviour as it also seems to be an issue regarding actual network performance ... :/Backport to 2.7 is currently missing as I'd need#18879 to be backported. If that is OK (I'd like to have this in 2.7) then I'd be happy to port both.

Antoine, I'm adding you here as I'm leveraging your patch from#18879. I'd need some feedback about the backport, but this patch should be OK for 3.4. Also, if you had an idea how to test this - I tried, but failed so far.

Antoine, could you check my last comment in here?(The nosy list got reset accidentally when I made that comment and got a conflict from the tracker).

Christian , with respect to patch, I agree with the logic (using something similar to#18879). Does all current unittests succeed with this? (I suspect not) A unittest for coverage would be helpful.

Well, this looks ok on the principle, but I haven't investigated the urllib-specific parts, so I'll let Senthil delve into this :)

Christian's patch is good.It helps in setting the socket.makefile file descriptor to a well behaving well file close wrapper and thus will help us prevent some tricky fd close issues.I added tests for coverage to ensure that we are asserting the type and covering all the methods in urllib.response. Attaching the patch built on top of Chritians one.

New changesetbb71b71322a3 by Senthil Kumaran in branch '3.4':urllib.response object to use _TemporaryFileWrapper (and _TemporaryFileCloser)http://hg.python.org/cpython/rev/bb71b71322a3New changeset72fe23edfec6 by Senthil Kumaran in branch '3.4':NEWS entry for#15002http://hg.python.org/cpython/rev/72fe23edfec6New changeset8c8315bac6a8 by Senthil Kumaran in branch 'default':merge 3.4http://hg.python.org/cpython/rev/8c8315bac6a8

This is fixed in 3.4 and 3.5. I will backport to 2.7 ( I think, it is worth it).

2.7 backport is no longer relevant.