I've released v0.95 of bz2file, which incorporates all the optimizations discussed here. The performance should be similar to 2.x's bz2 in most cases.

It is still a lot slower when calling read(10) or read(1), but I hope no-one is doing that anywhere where performance is important ;-)

One other note: bz2file's readline() is faster when running on 3.x than on 2.x (and in some cases faster than the 2.x stdlib version). This is probably due to improvements made to io.BufferedIOBase.readline() since 2.7, but I haven't had a chance to investigate this.

Let me know if you have any issues with the new release.