ARTEMIS-5037: option to limit mirror propagation

[lavocatt]

Add a new option in the Mirror settings to prevent a broker from
propagating messages.

When working with a topology where 4 nodes are forming a square and
where each node in that square mirrors its two neighbors: a message
leaving a corner can reach the opposite corner of the square by two
different routes. This is causing the message ordering to get broken.

example:

1 <-> 2
^     ^
|     |
v     v
4 <-> 3

A message from 1 will reach 3 by 2 and 4. Message duplication checks
will prevent the message from being duplicated but won't help regarding
the order of the messages. This is because a either the route by 2 or 4
can be faster than the other, so whomever wins the race sets the message
first.

Fixing the example:
Using the new option to not forward messages coming from a link, we
break the possibilities to have two routes to reach the opposite corner.

The above example is updated as followed:

• 2 never forwards messages coming from 1
• 1 never forwards messages coming from 2
• 3 never forwards messages coming from 4
• 4 never forwards messages coming from 3

Now, when a messages leaves 1:

• it reaches 2 and stops there
• it reaches 4
• it reaches 3 through 4 and stops there

Now, when a messages leaves 2:

• it reaches 1 and stops there
• it reaches 3
• it reaches 4 through 3 and stops there

Now, when a messages leaves 3:

• it reaches 4 and stops there
• it reaches 2
• it reaches 1 through 2 and stops there

Now, when a messages leaves 4:

• it reaches 3 and stops there
• it reaches 1
• it reaches 2 through 1 and stops there

The new test AMQPSquareMirroringTest.java is testing this exact setup.

[lavocatt]

@tabish121 Thanks for the review. FYI this is in a very drafty state ATM. Many refactors might come after that, in any case I'll make sure to take your notes in.

[lavocatt]

Thanks @gemmellr for the review! There are still things that are not working with the current PR and some refactors in the added lines might happen. In any case I'll make sure to take your comments in.

[lavocatt]

I've exposed the new no-message-forwarding parameter (with @gemmellr's help) , now I need to test that the parameter works in a test.

[gemmellr]

I wondered if the RoutingContext could be used for the 'dont [re-]mirror this' handling. It already has functionality for doing that (see isMirrorDisabled() and related).

However it now also occurs to me, we possibly do need to mark the message like this simply because when that message is acknowledged later, we probably want to know then that it was 'no forward' now, so that we dont mirror acknowledgement for it either given we never mirrored the message itself to begin with.

[gemmellr]

I'd also try adding some testing using the ProtonJ2 test peers, rather than all brokers, to validate and exercise the actual broker behaviour at the protocol level. It gives a better accounting of what is really going on and when it changes unexpectedly.

E.g it seems all but certain these brokers are currently forwarding around mirrored acks for noForward messages they didn't originally mirror to begin with, but this test will not notice that one way or the other. Similarly with various other potentially unexpected behaviours that might occur.

[gemmellr]

I suspect a different check will be needed here, inspecting the specific message; this is ultimately called by the queue, which I expect usually wont have a mirror target controller set. It will also want to accommodate not sending acks for no-forward messages [that it didnt send to begin with] even across broker restarts, due to expiry, etc etc.

I expect the test-peer based testing I know you are working on will demonstrate this.

[lavocatt]

I've implemented some tests and they seem to work despite the assumption we had that it would not. Maybe something is worth a double check?

[gemmellr]

I think your changes last week have actually already stopped it sending some of the stuff I previously thought it would and checked that it actually did.

There are other scenarios that are less clear though. For example, if you consume the message from the middle broker instead of the first, its going to mirror the ack onward, even though it didnt mirror the message originally, which is marked as no-forward.