3 releases

0.1.2 May 24, 2021
0.1.1 Apr 27, 2021
0.1.0 Apr 26, 2021

#365 in HTTP client

MIT/Apache

32KB
588 lines

CI

Rust library to parse and serialize async multipart/x-mixed-replace streams, suitable for use with reqwest, hyper, and tokio.

Note multipart/x-mixed-replace is different than multipart/form-data; you might be interested in the multipart crate for that.

What's a multipart stream for?

A multipart stream is a sequence of parts in one HTTP response, each part having its own headers and body. A stream might last forever, serving parts that didn't exist at the start of the request. This is a type of "hanging GET" or Comet request. Each part might represent the latest state and conceptually replace previous ones, thus the MIME type multipart/x-mixed-replace.

It's a simple HTTP/1.1 way of accomplishing what otherwise might require fancier server- and client-side technologies, such as:

Never-ending multipart streams seem popular in the IP camera space:

  • Dahua IP cameras provide a multipart/x-mixed-replace stream of events such as motion detection changes. (spec)
  • Hikvision IP cameras provide a multipart/mixed stream of events, as described here.
  • wikipedia mentions that IP cameras use this format for MJPEG streams.

There's a big limitation, however, which is that browsers have fairly low limits on the number of concurrent connections. In Chrome's case, six per host. For this reason, multipart streams are only suitable in HTTP APIs where the clients are not web browsers.

What is a multipart stream exactly?

A multipart response might look like this:

Content-Type: multipart/x-mixed-replace: boundary=B

--B
Content-Type: text/plain
Content-Length: 3

foo

--B
Content-Type: text/plain
Content-Length: 3

bar

and is typically paired with Transfer-Encoding: chunked or Connection: close to allow sending a response whose size is infinite or not known until the end.

I can't find a good specification. This WHATWG document describes multipart/x-mixed-replace loosely. It refers to RFC 2046 which defines multipart encodings originally used for rich emails. I don't think these HTTP multipart streams quite follow that RFC. My library currently requires:

  • Any content type—the caller should validate this to taste and pass the boundary parameter to the multipart-stream library.
  • No preamble. That is, no arbitrary bytes to discard before the first part's boundary. (Newlines are okay.)
  • Zero or more newlines (to be precise: \r\n sequences) between each part and the next part's boundary.
  • A Content-Length line for each part. This is a much cleaner approach than producers attempting to choose a boundary that doesn't appear in any part and consumers having to search through the part body.
  • No extra -- suffix on the final part's boundary. I've never seen one.

Please open a github issue if you encounter a multipart stream which doesn't match these requirements.

What does this library do?

It takes a stream of Bytes (such as those returned by reqwest or hyper) and returns a stream of multipart_stream::Parts, or vice versa. (For client or server use cases, respectively.) Each Part packages together headers and a body.

Author

Scott Lamb <slamb@slamb.org>

License

SPDX-License-Identifier: MIT OR Apache-2.0

See LICENSE-MIT.txt or LICENSE-APACHE, respectively.

Dependencies

~1.5–2.3MB
~44K SLoC