Kuadrant / authorino

K8s-native AuthN/AuthZ service to protect your APIs.
Apache License 2.0
201 stars 32 forks source link

Well-known attribute `request.time` is not really usable as it #500

Open alexsnaps opened 3 weeks ago

alexsnaps commented 3 weeks ago

This has become more painful because of the introduction of CEL, but request.time is made available as the struct that backs it: map[nanos:68000 seconds:1.730319299e+09]. With CEL having first class support it is probably expected (and documented) to be available as a proper Timestamp. Even in JSON tho, it would be probably be best represented in RFC3339 format.

Proposal

protojson deals with these conversions properly. Which is already used to produce the intermediary JSON result coming back from a CEL Value, as well as feeding out WellKnownAttributes into the interpreter. request itself is protobuf as it comes from envoy/service/auth/v3's AttributeContext which is autogenerated from a .proto.

I'm thinking we should have our own WellKnownAttributes spec'ed as protobuf and use protojson to marshal that data. With that approach, no json marshaling would actually to inject the WellKnownAttributes into the CEL interpreter, but they would be made available as protobuf Messages directly.

So the proposal is to refactor all of WellKnownAttributes into .proto, and use that generated code to produce proper typed JSON/CEL Values. wdyt?

alexsnaps commented 3 weeks ago

As part of the refactoring Value.ResolveFor would be changed to not take a JSON literal in anymore, but the WellKnownAttributes directly... so they still can be "cached" (kept around) on the call site, but the callee would now be the one to decide how to marshal that data...

type Value interface {
    ResolveFor(attributes *WellKnownAttributes) (interface{}, error)
}

// where WellKnownAttributes are obtained from the AuthPipeline, just as the JSON currently is with GetAuthorizationJSON
value.ResolveFor(authPipeline.GetWellKnowAttributes())
alexsnaps commented 3 weeks ago

Also note that currently, because of how data "gets out of CEL", we have this

exp, err := NewExpression("timestamp(\"2024-12-25T12:00:00Z\") + duration('123456m')")
v, err := exp.ResolveFor("{}") 
// v is a `string` of value `2025-03-21T05:36:00Z` , i.e. a RFC3339 date format

which I think is a desirable outcome here... but only given this.