HTTP/2 Push is dead
One of the hot features that came with HTTP/2 was PUSH frames. The main idea is that if the server can predict what requests a client might want to make, the server can preemptively send request/response pairs to the client and warm its cache.
This is a feature I’ve been very interested in for a long time. I feel that it can be incredibly useful for APIs to invalidate & warm caches, remove the need for compound requests (which I feel is a hack, although sometimes a necessary one).
To help advance this idea, I’ve worked on a Internet Draft to let API
clients tell the server what resources they would like to have pushed, I built
a Node.js framework with first-class, deeply integrated Push support, and
added support for
Prefer-Push to Ketting.
Unfortunately HTTP/2 push always felt like feature that wasn’t quite there yet. It’s usefulness was stunted due to Cache-Digest for HTTP/2 being killed off, and no browser APIs to hook into push events.
The Chrome team has considered removing Push support since at least 2018. The main reason being that they hadn’t really seen the performance benefit for pushing static assets. At the time, I tried to defend the feature for the API use-case.
Yesterday, the Chrome team announced to remove the feature from their HTTP/2 and HTTP/3 protocol implementations.
I’m a little sad that it never got to its full potential, but with this step,
I no longer think it’s worthwhile to invest in this feature. So I’m ceasing
my work on
Prefer-Push, and will also remove support from the next major
On a positive note, I spent a lot of time thinking and working on this, but it is sometimes nice to just be able to close a chapter, rather than to wait and let it sizzle out. It’s not an ideal conclusion, but it’s a conclusion nonetheless.
Lacking server-driver push, we’re back to many small HTTP request, or compound requests. This means you either have the N+1 problem, or (in the case of compound requests) poor integration with HTTP caches. More on this here.
If you’re going the ‘compound request’ route, there is a draft of a similar
Prefer: transclude, which Ketting also
A feature that I hoped would work well in the future with Server Push was server-initiated cache invalidation. We never quite got that. To work around this, we use Websockets and will keep doing this for the forseeable future.
To reduce general latency of fetching many things, the
103 Early Hints
can help, although this is not supported yet in Chrome, and this is also
only really useful for speeding up delivery of assets like images, css and