HTTP/2 Server Push

HTTP/2 server push is a mechanism to allow a server to return multiple documents in response to a single HTTP request. Server push can be used to improve page load performance when the server knows that a client requesting one document will also request another document; for example, when the client requests an HTML page, the server might also include the page's assets in its response. This avoids the need for the client to download and parse the HTML before requesting the assets, reducing the total number of HTTP transactions and improving the page load time.

However, server push has the significant disadvantage that the pushed documents are always sent to the client, even if the client already has them in its cache and wouldn't have requested them otherwise. Thus, pushing every asset used on a page can easily reduce performance (and waste bandwidth) by forcing the client to download objects it already has. We may look at ways of mitigating this on the server side in a future release (by detecting which assets the client might already have), but there is currently no perfect solution to this problem.

Deciding when to use server push is a complicated and application-specific issue, which we won't cover here. The following article may be helpful:

• Smashing Magazine: "A Comprehensive Guide To HTTP/2 Server Push"

Configuring Server Push

Server push is enabled by default. If you want to disable server push on an Ingress, use the ingress.kubernetes.io/server-push annotation:

metadata:
  annotations:
    ingress.kubernetes.io/server-push: "false"

Like other implementations, Traffic Server uses the Link header field in the HTTP response to determine what content to push to the client. Processing is done as soon as the response header is received, so if your application is slow to generate the response body, it can still begin pushing assets to the client once the header is sent.

To push a document, include a Link header field in your response with the rel=preload attribute:

Link: </path/to/my.css>; rel=preload; as=style

The argument to the as attribute should be one of the standard request destinations ("audio", "document", "embed", "font", "image", "manifest", "object", "report", "script", "serviceworker", "sharedworker", "style", "track", "video", "worker", or "xslt") although TS does not enforce this and will push the object anyway.

If you want to include a rel=preload link but not use server push for that resource, add the nopush attribute:

Link: </path/to/my/image.jpeg>; rel=preload; as=image; nopush

Server Push and caching

Before TS can push an object to the client, it must have a copy of the object itself. This means you should almost always enable caching on assets you intend to push; otherwise, the client will have to wait while TS goes back to the application to fetch each pushed asset, and you have lost a significant fraction of the performance you might gain from using server push.