Hi all, I found a nasty regression in dev23, it enforces the client/server timeout during HTTP transfers and does not report any issue because it believes it reached the end and closes the connection. This means that if you have a 30 second timeout and a visitor downloads a large file that needs more than 30 seconds to come, it will be truncated. This has been fixed in the git repository and will be in the next snapshot in a few hours.
This bug has also caused some breakage for some downloads from the main site, I could see in the logs that a number of 1.5-dev23 and 1.4.25 tarballs were truncated (which can easily be detected by checking their md5sum). I'll issue another version quickly after the snapshot, please test it as I did other changes in it : I fixed the bug that Igor reported long ago with the stats page in chunked mode, in fact it was not the stats page but a bug at a much lower layer in the buffer management. It could have caused some peers to occasionally disconnect/reconnect as well. And after fixing the stats bug, I enabled keep-alive and compression on the page. I already noticed that gitweb and the stats page compress at about 80%, this already saves some bandwidth on my poor line :-) Regards, Willy
