Support Message Batching

[jeremydmiller]

It comes up often enough.

So if you have a handler like:

public static void Handle(Message[] messages), and you publish await bus.PublishAsync(new Message()), Wolverine does the aggregation for you. Needs to be able to track the success/failure of each message, plus the batched message. For right now, only support this for local queues.

cc @BrianVallelunga

[BrianVallelunga]

I'll add the comment that I think the aggregation needs to be by Tenant, also.

[jeremydmiller]

Oh, good point

[jeremydmiller]

Notes

This might be a bit of over-analysis, but there's some hidden complexity to this

• Do we still store each incoming message as a separate row in the incoming table? There's value in doing that for the
  sake of the idempotency protection. Simpler mechanics in some way if we just start the batched message and go
• Batching will need to be explicitly configured some how. Control the batch size, buffer time, and whether it's tenanted or not
• Do you track the attempts at the single message level, or at the batch level? Easier at the batch level
• Do you persist at the batch level?
• Should it be strictly ordered, or not? Maybe that's configurable?

Concrete thoughts per endpoint type:

• W/ Azure Service Bus or AWS SQS listeners, do the batching at the point the message batching is done? Doesn't help if the batching has a delay

Implementation options

• Special IMessageHandler that just puts the data into a BufferBlock?
• Special BufferedBatchingQueue and DurableBatchingQueue? So you immediately publish to that special local queue
• Try to use a special message type that tracks the original envelopes? That would simplify the persistence

[jeremydmiller]

Latest Thoughts

Use two new classes that effectively implement IReceiver & ILocalQueue:

1. DurableBatchedReceiver
2. BatchedReceiver (basically Buffered)

Configure batching separately for each message type. Use LocalQueue for the configuration though. Build a separate receiver for each.

If receiving a message from the outside where there's the possibility of batching, use a MaybeBatchingReceiver that can do the splitting. If a listener endpoint has the MessageType set, use the batching queue as the receiver.

[jeremydmiller]

Punchlist

• End to end test with durable local queueing
• End to end test with durable local queueing and multi-tenancy
• Docs