bug#26489: [PATCH] substitute: Ignore bad responses.

[Ludovic Courtès]

Hi!

Tobias Geerinckx-Rice <address@hidden> skribis:

> On 14/04/17 11:54, Ludovic Courtès wrote:
>> Tobias Geerinckx-Rice <address@hidden> skribis:
>>> One weird HTTP response from a server will kill ‘guix substitute’:
>>> 
>>> updating list of substitutes from 'https://foo'... 50.0%Backtrace:
>>>  ... guix/ui.scm:1229:8: In procedure run-guix-command: 
>>> guix/ui.scm:1229:8: Throw to key `bad-response' with args `("Bad 
>>> Response-Line: ~s" (""))'. error: build failed: substituter 
>>> `substitute' died unexpectedly
>>> 
>>> Attached is a patch to ignore such bad responses. The offending 
>>> .narinfo will be ignored for that session, and not cached at all. 
>>> The result:
>> 
>> I’m sure you expect this question: what bad responses did you get in 
>> practice?  :-)
>
> In fact, not really. The error message looked unambiguous to me: the
> HTTP response (the first line returned to the client, e.g. "HTTP/1.1 200
> OK") was simply empty, throwing an exception.
>
> Interestingly, a newline seems to be required.
>
> Using http://bad.http.response.tobias.gr as a substitute server triggers
> it. http://no.http.response.tobias.gr does not.
>
>> Usually that is a sign of a broken HTTP server.
>
> I think it's actually something in-between me and the server. I'll take
> a closer look next time this happens.
>
>> Of course it’s  widespread enough, we’d better handle it, either in 
>> Guix or directly in (web client) in Guile;
>
> As I read it, (web client) considers throwing a BAD-RESPONSE exception
> the best or only way to deal with an error like this. I agree.
>
>> OTOH, if it’s a genuine problem, we’d better not hide it.
>
> Well, we don't hide it, per se. Hence the error message.
>
> I think throwing an unhandled exception is definitely the wrong thing to
> do here — this kills even ‘guix --keep-going --fallback’. I'm less sure
> about the right place to do it

Oh right, if that kills --fallback, that’s a problem.

Back to your initial patch, what about moving ‘bad-response’ handling to
the call site of ‘http-multiple-get’ instead of having it in
‘http-multiple-get’?  (That way, ‘http-multiple-get’ would behave like
‘http-get’ in this respect.)

Upon a ‘bad-response’, ‘fetch-narinfos’ would return #f or the empty
list or the partial narinfo list it has built so far.

WDYT?

I’d be happy with a patch along these lines!

Thank you,
Ludo’.