[PATCH] HTTP/2: expose function to push single resource to modules

[Alessandro Ghedini]

On Fri, Feb 09, 2018 at 10:35:59AM +0300, Ruslan Ermilov wrote:
> On Thu, Feb 08, 2018 at 07:48:25PM +0000, Alessandro Ghedini wrote:
> > On Thu, Feb 08, 2018 at 10:00:27PM +0300, Maxim Dounin wrote:
> > > On Thu, Feb 08, 2018 at 04:52:59PM +0000, Alessandro Ghedini wrote:
> > > 
> > > > # HG changeset patch
> > > > # User Alessandro Ghedini <alessandro at ghedini.me>
> > > > # Date 1518108716 0
> > > > #      Thu Feb 08 16:51:56 2018 +0000
> > > > # Branch expose-push
> > > > # Node ID 1bb98b06d5536dfc80a407aabd8d06f9309f8df6
> > > > # Parent  a49af443656f2b65ca5de9d8cad5594f44e18ff7
> > > > HTTP/2: expose function to push single resource to modules.
> > > > 
> > > > This makes it possible for 3rd party modules to implement alternative
> > > > methods for deciding which resources to push to clients on a per-request
> > > > basis (e.g. by parsing HTML from the response body, by using a custom
> > > > Link header parser, ...).
> > > > 
> > > > No functional changes.
> > > 
> > > Not sure this is a good idea.
> > > 
> > > You may consider exposing a variable to be used in http2_push 
> > > instead.
> > 
> > Right, the problem is that as far as I can tell http2_push only supports a
> > single resource, even when a variable is used, so it wouldn't be possible to
> > push multiple resources without specifying multiple http2_push directives,
> > each with its own variable, and even then you'd only have a fixed number of
> > resources that can be pushed, which wouldn't work well when the number of
> > resources changes depending on each request/response.
> > 
> > So in the end exposing the internal functions to modules seemed better than
> > just trying to make http2_push support multiple resources per directive,
> > which would add complexity to NGINX itself rather than the external modules
> > (though I can do that if you think it would be a better solution).
> 
> We've also considered adding support for the X-Accel-Push header, but
> decided not to implement it at this time.  If implemented, there could
> be multiple X-Accel-Push headers in the proxied response.

That might work for us, but it's a somewhat awkward interface to use from
inside a module, so I'd still prefer something more direct, and as I said,
exposing the function to modules seemed the least invasive change to NGINX.

Could you please expand a bit on why you think this might be a bad idea?

In any case I can look into implementing X-Accel-Push support if you don't
plan on doing it yourself.

Cheers