CNP batch requests (draft)

Overview

This is a draft of batch CNP requests and responses.

Batch requests could be used to send multiple requests for content without making multiple connections and CNP round-trips.

This solves the problem that keep-alive connections or request pipelining would solve while still keeping the entire request/response a single CNP round trip where the endpoints send or receive their entire side of transmission at once.

This solution is significantly simpler to implement than the previous solutions and solves most of the same problems. The main downside is that all content to be requested must be known ahead of time, so loading pages with embedded content means first making a normal request connection for the page and then a batch request connection for all the embedded content to be loaded immediately instead of using a single keep-alive connection. However, the count of round-trips remains the same as a pipelined keep-alive connection.

Specification

A batch request would use the keyword batch instead of the request intent, the count of batch requests in the count parameter and CNP requests as the data. The length parameter is not necessary when count is provided in a supported context, such as the batch request; however, each sub-request must use the length parameter if it has any data.

Note that the data of a batch request and the next batch request are not separated by a line feed; as such, a request in the batch does not always start on a new line but instead after the last request's length of data has been processed. All requests in the batch still end with a line feed.

Only one request for each individual host+path pair (i.e. request intent) may be present in the batch request. If multiple requests in the batch have the same intent, it is an invalid error.

Requests in the batch may not be batch requests themselves. If any of the sub-requests is a batch request, an invalid error should be returned.

cnp/0.4 batch count=5
cnp/0.4 example.com/logo.png
cnp/0.4 static.example.com/img/image.png if_modified=1970-01-01T00:00:00Z
cnp/0.4 example.com/submit length=6
foobarcnp/0.4 example.com/page/helloworld.cnm
cnp/0.4 static.example.com/img/image2.webp if_modified=1970-01-01T00:00:00Z

The response would use the keyword batch as the intent with the count parameter set to the count of responses. Each response in the batch must include the length parameter, even if the length is zero, and the intent parameter set to the parameter of the corresponding request in the batch request. Individual responses in the batch are parsed the same way as batch requests.

Batch responses do not have to come in the same order as the requests and may be processed in arbitrary order, including concurrently. The client must not rely on the requests being executed in order and thus should not issue non-idempotent batch requests that interact with other requested resources.

A batch response is not required to contain responses to all batch requests, but is recommended to. Any omitted responses can be requested again with individual requests to ensure the server answers them.

cnp/0.4 batch count=5
cnp/0.4 error reason=not_found length=0 intent=example.com/logo.png
cnp/0.4 not_modified length=0 intent=static.example.com/img/image2.webp
cnp/0.4 not_modified length=0 intent=static.example.com/img/image.png
cnp/0.4 ok length=30 intent=example.com/page/helloworld.cnm
content
	text
		Hello, world!
cnp/0.4 ok length=16 type=application/json intent=example.com/submit
{'status': 'ok'}

Possible changes

  • The CNP version field is unnecessary in messages in the batch and could be elided. However, this may require special handling by parsers, so it may be simpler to just leave it there.

  • If a message in the batch has a non-zero length body, a line feed could be appended to the end. This would produce more human-readable requests and responses, but may be slightly more complicated to parse, since the line feed character would have to be consumed before the next message is parsed.

  • Batch requests with body could be forbidden. This would solve the problem of non-idempotent requests conflicting with each other, but the same can happen with multiple concurrent connections anyway.

  • Force the response to have the same number of messages as the request, with any omitted responses being error (perhaps with a new reason parameter value). This risks discarding mostly complete responses, though.

  • Force the batch response to retain the ordering of requests. This may make responses easier to parse, but would introduce head-of-line blocking. The intent parameter would still be required, since it ensures that the response can be understood even without the context of the request body (i.e. the batch of requests).