RFC: new vcl_lookup{} proposal

[Poul-Henning Kamp]

Right now we go directly from vcl_hash{} to either vcl_hit{}, vcl_miss{}
or vcl_pass{} depending on the outcome, and that means, amongst other
things, that a VCL implementation of "PURGE" as to add code to all those.

Furthermore we have some semi-magic things going on behind the scenes,
such as grace and saint modes.

Assume:

* Backend fetches will be done by a different thread than the one
  processing the clients request.

* vcl_hash{} always is followed by vcl_lookup{} and that the above mentioned
  logic gets moved to VCL, rather than C code.

* vcl_hit{} goes away.

Then a plausible default vcl_lookup{} could then look like:

	sub vcl_lookup {

		if (!obj) {
			// Miss:  Start a background fetch in another thread
			// We come back here, when the status of that fetch
			// is available. (ie: we go on the waiting list)
			return (miss);
		}

		if (obj.ttl <= 0s &&
		    obj.ttl > -req.grace && obj.ttl > -obj.grace) {
			// We have a grace situation
			if (!obj.busy) {
				// Nobody is fetching yet, so start a
				// background fetch
				background_fetch();
			}
			// NB: First client does not hang on the fetch
			return (deliver);
		}

		if (obj.ttl <= 0s) {
			// object is too old, fetch new one (see miss)
			return (miss);
		}

		return (deliver);
	}

Possible customizations:

Purging:

	if (req.request == "PURGE") {
		// We pressume access control was in vcl_recv{}
		return (purge);
		return (purge_all);
	}

Prefetching:

	if (obj.ttl < 2s && !obj.busy) {
		// Close to deadline, pre-fetch
		background_fetch();
	}

Don't stream, only deliver complete objects:

	if (!obj.complete) {
		wait_complete(obj);
	}

I'm seriously pondering if saintmode should become a vmod at the same
time, in which case it would probably look something like:

	sub vcl_lookup {
		if (obj.ttl <= 0s && !saint.ok(req.backend, req.url)) {
			return (deliviver)
		}
	}

	sub vcl_fetch {
		if (beresp.status >= 400) {
			saint.bad(bereq.backend, bereq.url);
		}
	}

-- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
phk at FreeBSD.ORG         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe
Never attribute to malice what can adequately be explained by incompetence.