According to the RFC, the server is allowed to send back any encoding it likes when no Accept-Encoding header is supplied, but all the examples I can find of urllib2.urlopen usage assume they're getting plain text back.  I think it would be better to inject an Accept-Encoding header when none is explicitly supplied so that nobody else trips over this issue.

See http://support.github.com/discussions/site/1510

HTTP Ref says that Server can send any encoding, if client does not
specify Accept-Encoding header. But if 'identity' is one of the
encoding that server recognizes (?), then it should send it as
identity, which indicates untransformed content.

I also see in the httplib that Accept-Encoding = 'identity' is added in the
request level to the headers. I shall see what is missing here, if it
is not being sent for all requests.

BTW, I could not figure out the problem you are facing from the url
mentioned. I specifically do not see any interleaving gzip and no-gzip
request behaviours at different points.

How many tests did you run?  My two tests were minutes apart.  I have the feeling that this has something to do with cacheing behavior on the server.