@Martin:

Yes, that's right, but only for cases where the TarFile.fileobj attribute is an actual file object. But, most of the time it is something special, e.g. GzipFile or sys.stdin, which makes random seeking either impossible or perform very badly.

But thanks for your objection, I have to withdraw the statement I made under option 2.: compressed archives are much more common than uncompressed ones. We probably wouldn't lose too much if we no longer use seek() but read() in TarFile.next(). Reading in an uncompressed file is fast anyway. I have to think about this.