For the record, I also like option (2), for the reasons that Justin
articulated.
-Ted
On 10/26/2012 11:24 AM, Justin Ross wrote:
I like option 2) for two reasons. One, it produces very
straightforward semantics for the various levels of delivery
guarantee. Two, it's easy to ignore if you'd like: if you just want
fire-and-forget messaging, you don't have to engage delivery as an api
concept; if you later change your mind, it's not hard to level up.
Justin
On Fri, 26 Oct 2012, Rafael Schloming wrote:
I'm taking a look at expanding the messenger API to support
reliability and so far there seem to be two directions to explore
which I'll attempt to describe below:
Option 1)
Messenger.ack(Message) or possibly Message.ack()
I'll describe this as the simple/expected/conservative option, and
those really are its strong points. Some less desirable points are
that it takes the Message concept in a bit of a different direction
from where it is now. Message is no longer simply a holder of
content, but it is also now (at least internally) tracking some kind
of delivery state. This is undesirable from a design perspective
since you're really merging two separate concepts here, delivery
state and message content, e.g. you now end up holding onto the
message content to track the delivery state when arguablly the
common case is that an app will be done processing the content
quickly but may care about the delivery state for longer. Another
implication of this merging is that it makes messages harder to
reuse, e.g. imagine if you want to receive a message, mutate it a
bit, and then resend it or send a number of similar messages.
It's also potentially more work from an implementation perspective
as the underlying model treats Message as pure content and has
delivery factored out as a separate concept, so this would be the
start of a bit of an impedence missmatch between layers. It's
certainly doable, but might result in more overall code since we'd
be expressing similar concepts in two different ways.
Option 2)
Introduce/surface the notion of a delivery/tracking number through
the Messenger API, e.g.:
Messenger.put(Message) -> Delivery/Tracking-Number/Whatever
Messenger.get(Message) -> Delivery/Tracking-Number/Whatever
(I'll abbreviate Delivery/Tracking-Number/Whatever as DTW for now.)
There are a couple of choices with this option, DTW could be a value
object (i.e. similar to a handle), with actions on the messenger
itself, e.g. Messenger.ack(DTW), Messenger.status(DTW). This kind of
takes the tracking number analogy and runs with it, you ask your
messenger about the status of a given delivery via its tracking
number. Alternatively, DTW itself could be a more action oriented
interface with its own methods for acking/status/etc. This isn't
necessarily an either/or as there are reasons the C API might want
to use a handle approach even if we wish to conceptualize/surface
DTW as more of an independent thing in the object oriented
interfaces.
On the negative side this is less traditional/expected relative to
Option (1), and it does in total add more surface area to the API.
On the positive side however it does provide a lot more capability
since the same concept extends quite easily to track the state of
outgoing deliveries.
From a design perspective this is nicer for a couple of reasons,
unlike Option (1) by keeping Message as a pure holder of content you
have more flexibility with how you use the API, e.g. you can
discard/reuse the Message but still track the status of your
deliveries. Also, it seems likely that once the Option 1) path
includes the ability to track the status of outgoing deliveries, the
total surface area balance might fall more in favor of Option (2).
A possible addendum to Option (2) suggested by Rob in order to deal
with the surface area concerns is adding some kind of
Messenger.ack() that would acknowledge all unacked deliveries pulled
from the incoming queue. This would enable you to do basic acking
without ever needing to bother with DTWs if you don't care about
them. This could in fact be a nice place to start as it doesn't
necessarily commit us to either path initially.
FWIW, my bias right now is towards exploring Option (2). I think the
fact that it is less expected is sufficiently mitigated by the fact
that with the addendum there is a very gentle learning curve, and even
if you explain all the concepts up front, the whole tracking number
analogy makes it fairly intuitive. I can even imagine playing up the
difference as a selling point from a technical marketing perspective.
--Rafael