Re: Improving merge failure message

From: Johannes Sixt <j.sixt@viscovery.net>
Date: Sep 8, 2009, 08:24 UTC
Message-ID: <4AA614A2.2060908@viscovery.net>
In-Reply-To: <7vbplmhr0i.fsf@alter.siamese.dyndns.org>

Junio C Hamano schrieb:

Show 14 quoted lines

> Nanako Shiraishi <nanako3@lavabit.com> writes:
> 
>> [2]% git merge feature
>> error: Entry 'cool' not uptodate. Cannot merge.
>> fatal: merging of trees 8ec1d96451ff05451720e4e8968812c46b35e5e4 and aad8d5cef3915ab78b3227abaaac99b62db9eb54 failed
>>
>> ... the messages look unnecessarily scary, with two
>> "error" and "fatal" comments, and long sha1 commit names.
> 
>> It would be nice if the message in the latter case can be toned down.
> 
> Yeah, it would be nice.  This actually was something that bothered me as
> well while trying to explain the recovery procedure for these two cases.
> Give me half an hour or so to cook up something...

I think that this is a symptom of a much more involved issue about error handling.

Currently, it is customary in the code to report an error at each location where an exceptional condition is detected:

	if (frotz())
		ret = error("frotz messed up");

If frotz() is a "low-level" routine, like a C library function, then this is the right thing to do. However, if frotz() is our own function ("high-level") which itself calls low-level functions with the same pattern, then this is obviously the wrong thing to do, because it results into two error messages.

We need a guideline how errors are reported. This guideline could be:

(1) Low-level functions do not print error messages, but set an error code and leave the reporting to the caller.

(2) High-level functions must print an error message (through error() or die()) when they detect a failure of a low-level function that was called directly.

(3) High-level functions must not print an error message when they detect a failure of another high-level function that was called directly.

There remains to classify our functions into high-level or low-level. But I think we won't have a lot of low-level functions.

BTW, I applied this guideline when I worked on {start,finish,run}_command recently, and IMHO it worked out quite nicely because exactly one error is reported after a failure.

-- Hannes