Open Stack

On 02/06/2017 08:45 AM, Brian Rosmaita wrote:
> On 2/6/17 5:51 AM, Jordan Pittier wrote:
> [super-enormous snip -- Chris, Ken, and Jordan make good points, I
> encourage you to read the entire thread; I just want to concentrate on
> one point]
>>
>> I would say we should make compromise, not solve dilemma. I can live in a
>> world where we sometimes allow an API change and sometimes prevent it.
>>
> +1000
> 
> I agree with Jordan.  We need to look at the context of each specific
> case and decide whether a change is OK based on the details.  We've
> already got the guideline that says "in general", you shouldn't change
> the response code, and we respect that.  The Glance team isn't claiming
> that the guideline is incorrect--we're just saying that given the
> context of this specific bug (that is, it's been documented for a long
> time to return a 204, all other metadefs DELETE calls are documented to
> return a 204, all the other metadefs DELETE calls do in fact return a
> 204, etc.), it makes sense that this case is an exception.

Agree.

I also agree with something from back in the thread - consumers are much
more likely to be checking for non-2xx statuses. In fact,
python-requests exception_from_request does exactly this- it checks to
see if the code is 4xx or 5xx and if so throws an appropriate exception.
Folks not using python requests are much more likely to be using a
similar approach (checking for return codes generally before looking for
specific reasons in switching logic) than to actually have an explicit
implementation for each REST call with an active check that the status
code was 200 vs 204.

> Granting an exception here doesn't mean that the floodgates have opened
> for an "anything goes" approach to API changes.  It just means that an
> exception is appropriate in this particular case.  I am being a bit
> disingenuous there because if an exception is appropriate in this case,
> then it will be appropriate in other relevantly similar cases.  But
> "relevant similarity" will include the entire context of the case, for
> example, whether there was a published API contract, whether the other
> similar calls behave as documented, etc.  From 10,000 meters, it looks
> like what we're advocating is "It's OK to change a response code".  But
> when you look more closely, our claim is that given the details of this
> particular bug, it makes sense to fix it in the code and not in the docs.
> 
> To summarize, my point is that we shouldn't be worried that this case is
> going to set a precedent.  It would be worrisome if it were going to set
> a *bad* precedent, but when you look at the details of the situation, I
> don't think it will.  So it looks to me, anyway, that a compromise is in
> order here.  (In case I'm being too obscure, what I mean is: we should
> agree that it's OK for the Glance team to fix this bug in the code with
> patch https://review.openstack.org/#/c/420038/.)

Yup. Totally agree.