Add RFC 006: Format Negotiation Protocol

docs/rfc/006-format-negotiation.md

@@ -0,0 +1,170 @@
+# RFC: Conjure Format Negotiation
+
+2018-10-04
+
+Today, Conjure supports one wire format: PLAIN for primitive data types in headers, path parameters, and query
+parameters, and JSON for complex objects (and primitives) in HTTP bodies. In the future, we may want to introduce new
+versions of this wire format, or entirely new wire formats like CBOR. This RFC defines a negotiation strategy by which
+clients and servers agree on the wire format used for requests and responses.
+
+We have the following requirements for the format negotiation protocol:
+- The protocol should in almost all cases not incur additional round-trips
+- Clients and servers will eventually agree on the protocol that is (1) most preferred by the client and (2)
+  supported by the server. That, clients drive the negotiation, constrained by server capabilities.
+
+### Definitions
+
+- A *version* is a positive integer
+- A *Conjure format identifier* is a string of the `application/<format>; conjure=<version>` where `<format>` is a
+  non-empty string over `[a-z-]` (e.g., `json`) and `<version>` is a version string (as above)
+- A *Conjure format list* is a comma-separated, ordered list of Conjure format identifiers
+- The `Accept` and `Content-Type` HTTP headers are defined as per the
+  [HTTP 1.1 spec](https://www.w3.org/Protocols/rfc2616/rfc2616-sec14.html). Note that `Accept: <format list>` is a
+  valid HTTP 1.1 Accept header, and that `Content-Type: <format identifier>` is a valid HTTP 1.1 Content-Type header.
+
+### Wire format versioning
+
+A Conjure wire format comprises specifications for Conjure parameters and return values. Today, the canonical format
+uses JSON objects for `body` request and response objects and PLAIN for all other parameters.
+
+We propose that every revision of a Conjure wire format be labelled with a format identifier. The current PLAIN+JSON
+format shall be labelled `application/json; conjure=1`. To allow for backwards compatibility, clients and servers
+should interpret a `application/json` header without a `conjure=<version>` parameter as the `application/json;
+conjure=1` format.
+
+### Format capabilities
+
+We assume that clients and servers implement both serialization and deserialization for a given format:
+
+- A server can consume requests with a given format iff it can produce responses with the same format.
+- A client can produce requests with a given format iff it can consume responses with the same format.
+
+### Negotiation protocol
+
+**Requests.** 
+Clients send the format identifier used to encode the request as a `Content-Type` header, and a format list as an
+`Accept` header. The format indicates the preference-ordered list of formats that the client supports. The supported
+formats must include the format use to encode the request, i.e., the format specified in the `Content-Type` request
+header.
+
+**Responses.**
+Servers that do not support the request format respond with HTTP status code `415 Unsupported Media Type`. Otherwise, if
+the server does support the request format, it uses the most-preferred (as per `Accept` request header) format to encode
+the response and advertise the chosen format in the response `Content-Type` header. Upon receiving a `415` response,
+clients may choose to repeat the request encoded with a more widely supported (i.e., typically older) format.
+
+**Binary data** Conjure supports endpoints accepting `binary` request parameters or returning `binary` responses. In
+accordance with the wire spec, a request carrying binary data in the request body must include a `Content-Type:
+application/octet-stream` request header, and a server returning a `binary` response must include a `Content-Type:
+application/octet-stream` response header. Further, when calling endpoints returning a `binary` response, clients must
+include in the `Accept` format list:
+
+- `application/octet-stream`
+- the list of supported Conjure formats and versions (e.g., `application/json; version=2`)
+
+The rationale for including both the `octet-stream` and the `json` (or `cbor`, etc.) formats is that the former is used
+when returning binary data and the later for returning structured Conjure errors. The `application/octet-stream` format
+is not versioned.
+
+
+### Discussion
+
+**Clients in control.** Clients are in control of the format negotiation in the sense that rank the list of acceptable
+protocols based on their preference and let the server merely "chooses" based on its support for the most preferred
+formats. Further, clients are in control of the trade-off between choosing the newest or most preferred versus an older
+or more widely supported format. The former approach unlocks new formats and features more quickly, but may result in an
+additional round-trip when the client has to reissue the request encoded with a more widely supported format.
+
+**OPTIONS.** This RFC does not propose to make the set of server-supported formats available as through an OPTIONS
+request. A client receiving a `415` error from a server can instead retry the request with the maximal list of
+client-supports formats and let the server pick the most preferred one.
+
+**Blue/green, server vs service.** This RFC does not propose any explicit mechanism for distinguishing between the
+different servers making up a remote service. For example, a service undergoing a blue/green migration between versions
+that support different sets of formats may pick different response formats depending on which server handles the
+request. We submit that this will rarely cause issues since clients and servers would still be able to agree on a format
+and even switching between different formats for subsequent requests would be acceptable.
+
+**Example: Conservative client.** The following sequence of two requests and corresponding responses are between a client
+that supports CBOR version 2, CBOR version 1, and JSON version 1, and that prefers formats in that order, and a server
+that supports CBOR version 1 and JSON version 1. To bootstrap a session, the client makes conservative assumptions about
+the server's capabilities and encodes the first request with a widely supported JSON format, version 1. The servers
+encodes the response with the format most preferred by the client that it also supports itself, CBOR version 1. The
+second request is encoded with the format most preferred by the client that the server supports, CBOR version 1.
+
+```text
+Request
+  Content-Type: application/json; conjure=1
+  Accept: application/cbor; conjure=2, application/cbor; conjure=1, application/json; conjure=1
+
+Response
+  Content-Type: application/cbor; conjure=1
+
+Request
+  Content-Type: application/cbor; conjure=1
+  Accept: application/cbor; conjure=2, application/cbor; conjure=1, application/json; conjure=1
+
+Response
+  Content-Type: application/cbor; conjure=1
+```
+
+**Example: Cutting edge client.** If a client accepts only the most cutting edge version of a format, an older server
+may not be able to pick a supported format. The client may choose to error, or can alternatively retry the request
+with an older format.
+
+```text
+Request
+  Content-Type: application/json; conjure=2
+  Accept: application/json; conjure=2
+
+Response
+  Status: 415
+
+Request
+  Content-Type: application/json; conjure=1
+  Accept: application/json; conjure=1
+
+Response
+  Content-Type: application/json; conjure=1
+```
+
+**Example: Old server.** A server that predates this RFC may ignore the format instructions and decode the request
+w.r.t. the canonical JSON format (i.e., `application/json; conjure=1`). The response should carry the appropriate
+content type as per the wire spec.
+
+```text
+Request
+  Content-Type: application/json; conjure=2
+  Accept: application/json; conjure=2
+  # Note that the server will interpret the version 2 request as if it were version 1
+
+Response
+  Content-Type: application/json
+  # Clients may assume that application/json is the version 1 JSON format
+```
+
+**Example: JSON request, binary response.** A request to an endpoint returning `binary` data must include
+both the `octet-stream` and a standard Conjure format list in the `Accept` header.
+
+```text
+Request
+  Content-Type: application/json; conjure=2
+  Accept: application/octet-stream, application/json; conjure=2
+
+Response
+  Content-Type: application/octet-stream
+```
+
+In case the server returns a structured Conjure error, it formats the error according to the most prefered of the
+accepted Conjure formats.
+
+```text
+Request
+  Content-Type: application/json; conjure=2
+  Accept: application/octet-stream, application/json; conjure=2
+
+Response
+  Content-Type: application/json; conjure=2
+  Status: 4xx or 5xx
+  # body contains Conjure error in JSON version 2 format
+```