The .NET Service Bus REST protocol for Queues - Some comments, some answers

Stefan Tilkov has several interesting remarks regarding our .NET Service Bus REST Queue Protocol that are worth addressing.

      Putting a password in the URI to get an identity token seems to expose information unnecessarily

That’s an area where we know that we’re going to change the protocol. We’ve already labeled that protocol as temporary in the documentation for the PDC CTP and we didn’t get all the pieces in the .NET Access Control service together, yet. Since it’s a HTTPS call, the data doesn’t get exposed on the wire, though.

      Queue creation seems fine, even though I feel a little uneasy about wrapping this in an Atom entry

Using Atom 1.0 and the Atom Publishing Protocol as the framework for managing the namespace is very intentional - for several reasons. First of all, it’s a standardized protocol for managing generic lists and the elements in those lists. With that we have a stable and accepted protocol framework and there’s plenty of tooling and framework support around it. That’s worth something. All we need to do is to add some simple extensions – the policies – on top of that stack. Beats having to define, version, and maintain a whole protocol.

      On the other hand, Atom seems reasonable considering you get an Atom feed from the queue’s “parent” resource

That’s what I mean. All sorts of tools know how to navigate and display Atom.

      Very nice to see the use of link/rel to get to the detailed Queue URIs; it would be even better if the rel values themselves were URIs in an MS namespace

I don’t see much potential for collision here and I would find it odd to have something as simple “self” and “alternate” and then add some unsightly QName for my rel expressions. Simple is good.

      Using “alternate” for the tail seems strange

“self” refers to the location where the Atom <entry> resides. “alternate” is what the entry points to. Since the Queue gets mapped into the namespace by “sticking its tail out”, the choice of the alternate link is the simplest possible mapping I could think of.

      “The queue’s tail endpoint accepts requests with any HTTP Content-Type and any HTTP method except GET, HEAD, and OPTIONS.” The tail does not accept GET. Why not? Can’t think of a reason. Also, I don’t get why any method is allowed. It seems to me a POST/PUT combination or an idempotent POST variant would be the way to go.

The way to look at this is that the Queue’s tail is acting on behalf of the receiver/resource that’ll eventually pick up and process the messages. POST, PUT, DELETE, and BOOYAH are all operations that cleanly map to processing operations and can often be delivered asynchronously with a 202 receipt reply. GET and HEAD don’t make much sense when executed in an asynchronous fashion without getting a reply that’s backed by a response for the receiver. OPTIONS is simply reserved for future use.

      “DELETE performing a destructive read” – huh? What happens if I don’t get back a response? How can I retry (as I can’t tell whether I get the same or a different message)? Maybe it’s intended to be used only if you don’t care about losing messages.

DELETE is indeed the dequeue operation variant that you’d use if you are ok with occasional message loss and want to trade transfer reliability for fewer roundtrips.

      The POST lock/delete approach, on the other hand, is very nice. Maybe it should be made idempotent, again e.g. using POE

POST lock/delete is the dequeue variant where you are doing the trade the other way around. More reliability bought with an extra roundtrip. In my view, idempotent access to individual messages isn’t much of a practical priority for a competing consumer queue. You’ll get a message or a set of messages to look at under a lock and if you walk away from the message(s), those messages pop back into the queue for someone else to look at. There are obviously scenarios where you want to look at a message sequence in other ways than an ordered queue where you can only get at messages as they appear on the head of the sequence – direct message access and per-message idempotent access matter in those scenarios and we’re looking to give you a capability of that sort in a different kind of messaging primitive.

      “The Delete request is sent without any entity body and MUST have a Content-Length header that is set to zero (0)”; although my immediate reaction was to question whether DELETE ever carries a body, the HTTP spec indeed doesn’t say anything about this

One of the great things about working here is that there are all sorts of interesting people around. I’ve discussed the use of DELETE and whether you can provide an entity body in either direction with Henrik Frystyk Nielsen, who works as an architect on the WCF team and is one of the co-authors of HTTP 1.1. Henrik’s stance is that all operations allow entity-bodies unless it’s explicitly forbidden in the spec. I don’t have a better authority to talk to.

      “The DELETE and POST operation have a set of options that are expressed as query parameters appended to the queue’s head URI” - the wording is worse than the actual approach.

I’m sorry that my writing is so clumsy ;-)