Re: Useless error message?

Aghiles wrote:

Please, fix it. :)

The problem is this: as far as I can tell, the git protocols are designed around the success case. Sometimes if there is an error or other interesting event, the servers are kind enough to notify the user “on the side”. But in the end, all too often, they do not bother to inform the client _program_ that a fatal error occured.

We can’t just throw away this hang-up message because sometimes when the remote host hangs up it really was unexpected.

So the trick is to make it expected more often. See the side-band-64k capability in Documentation/technical/protocol-capabilities.txt: the goal is to have fatal error messages for as many failure modes as possible.

Examples (I could be missing nuances; I am just trying to convey the idea):

upload-archive:
 - a pipe(), write(), or poll() failure when communicating over
   the wire or between local processes will result in an unexpected
   hangup.  I have not checked, but I suspect a SIGPIPE can kill
   upload-archive, too.
 - On the bright side, all other error conditions are properly
   handled.  The code for this is very nice and worth imitating.

upload-pack:
 - a missing or shallow repo, HEAD or some other ref pointing to
   a nonexistent object, early protocol error, or failure to start
   rev-list or pack-objects will result in an unexpected hangup.
 - errors from rev-list, pack-objects, or the transition of the
   generated pack are correctly handled.

receive-pack:
 - all errors result in unexpected hup as far as I can tell.

daemon:
 - hangs up without an explanation (except to syslog) for invalid
   or disabled repositories
 - if the underlying service hangs up, hangs up.  If the underlying
   service writes a message to stderr, writes that message to
   syslog.  Surely the client is not interested...

If any other information would help, please let me know. Jonathan