Clarify relationship between messaging, faas, and RPC

[anuraaga]

I want to clarify the relationship between messaging, RPC, and FaaS which is confusing to me right now. Let me give an example, based on my understanding of the current spec, and let me know if this aligns with expectations. This is an example of lambda + sqs, something I'm working on now and is sufficiently complicated that hopefully other cases are simpler.

SQS - an HTTP-based RPC API for publishing and receiving messages
Lambda - a FaaS that can be integrated with SQS. Lambda runtime polls for messages using SQS's RPC API, and when a message comes in, it runs a function with the received batch of messages

SQS Queue - NiceQueue
Lambda function - process_function

So I can invision many spans. These are all for the same request so up to the producer spans is all in the same trace
Request - Request span for the "trace", with current span lasts until end of production
Message1 - Message 1
Message2 - Message 2
Producer1 - Producer Span for Message 1
Producer2 - Producer Span for Message 2
APIPublish1 - Client span using SQS API to send Message1
APIPublish2 - Client span using SQS API to send Message2
PollMessages - Internal, infinite? span that corresponds to the lambda runtime process itself
APIReceive - Client span using SQS API to poll for messages and receive Message 1 and 2
FaaSInitialize - Span encompassing lambda function initialization and invocation. May have cold start
FaaSInvoke - Span encompassing lambda function invocation by runtime
FaasFunction - Span encmpassing lambda function execution in user's app
Receive - Consumer span encompassing the recived batch
Process1 - Processing span for Message1
Process2 - Processing span for Message2

Trace 1
|--------------------------OrderService.ProcessOrders-------------------------|
 |NiceQueue send|                                    |NiceQueue send|
                     |SQS.SendMessage|                               |SQS.SendMessage|


Trace 2
|-------------------------pollmessages------------------------------------------------------------------------------------------|
|APIReceive||APIReceive||APIReceive||APIReceive||APIReceive||APIReceive||APIReceive|
                                                                                    |-----FaaSInitialize-----------------|                                                                                                                  
                                                                                               |---FaasInvoke--------------|
                                                                                               |---FaasFunction------------|
                                                                                                  |--Receive--------------|
                                                                                                   |Process1||Process2|

Phew lots of spans. So does it look like this?

id	name	kind	duration	parent	link	semantic types	owner
Request	OrderService.ProcessOrders	SERVER	50ms	none	none	rpc,http	user
Producer1	NiceQueue send	PRODUCER	0	Request	none	messaging	user
ApiPublish1	SQS.SendMessage	CLIENT	20ms	Producer1	none	rpc,http	user
Producer2	NiceQueue send	PRODUCER	0	Request	none	messaging	user
ApiPublish2	SQS.SendMessage	CLIENT	20ms	Producer2	none	rpc, http	user
PollMessages	pollmessages	INTERNAL	infinity	none	none	none	infra
APIReceive	SQS.ReceiveMessage	CLIENT	30ms	PollMessages	none	rpc, http	infra
FaaSInitialize	Lambda::Initialize	INTERNAL	500ms	none	none	none	infra
FaaSInvoke	Lambda::Invoke	INTERNAL	300ms	FaasInitialize	none	none	infra
FaaSFunction	process_function	INTERNAL	300ms	FaaSInvoke	none	faas	user
Receive	NiceQueue receive	CONSUMER	250ms	FaaSFunction	none	messaging	user
Process1	NiceQueue process	CONSUMER	100ms	Receive	Producer1	messaging	user
Process2	NiceQueue process	CONSUMER	100ms	Receive	Producer2	messaging	user

My currently open questions

• This is an ideal case, but especially when using auto instrumentation, it may not be possible to have Process1 and Process2 - Lambda presents the entire batch to the user. Should Receive also have links to Producer1 and Producer2? Should Receive be a process span instead of a receive span?

• How to connect the API calls and the messaging spans. Completely separate like I listed? It means there are mysterious 0 duration producer spans and APIPublish are child or PRODUCER. Or combine e.g., Producer1 and ApiPublish1? Would the name be NiceQueue.send but we still include rpc.service == SQS, rpc.method == SendMessage to not lose that information? Is such a span's kind producer or client?

• Are FaaSInvoke and Receive separate or should be combined? In reality, if the runtime is a cloud provider, there may not be control and FaasInvoke is automatically created.

• APIReceive is a CLIENT span so I wouldn't expect it to be a parent of the rest - is it basically an orphan span? Anyways, while I may be able to model it as a parent of something if I was implementing all this myself, it wouldn't be for lambda = SQS.

Appreciate any guidance :)

[anuraaga]

/cc @Oberon00 @arminru Sorry for summoning you but I'm enjoying hearing about messaging from you and hope to enjoy more here too :)

[arminru]

Hi @anuraaga! It's really nice to see a real-life end-to-end example that is a bit more complex than the isolated examples we currently have in the spec, thanks for that!
In order to understand the whole compilation better, it would be helpful if you could mention which semantic conventions are applied in which span (I suppose it's only RPC in APIReceive and only Messaging in Receive, right?) and ideally arrange them in a timing diagram (either a simple ASCII chart or a quick drawing).

Regarding your questions:

1.) While the semantic convention does not prescribe which links should be applied, it could make sense to add links to the receive span to not lose that information if no (individual) processing spans can be created to preserve it as a parent. If - from the (auto-)instrumentation's and library's point of view - processing is finished when the messages are handed over to user code, just creating process spans instead of a receive span could make sense. I could imagine, however, that it does not necessarily make sense to look at each message independently in that function, which is why the semantic conventions suggests creating one receive span covering all messages as they are received in a batch. Therefore going with links might be the better solution.
I wonder if we should extend the messaging semantic conventions further to give guidance on links rather than showing possible solutions in the examples.

2.) If producing the message is an operation that might be interesting, it should be tracked independently. Given that you list them with a duration of zero, it looks like you only create the spans to fulfill the messaging semantic conventions separately but don't track any actual work, right? I don't think our "top-level" semantic conventions consider being mixed together in one span, as you can see from the incompatible span name and kind guidelines. I don't see that zero duration spans as a problem but I'm not sure if there are backends that would not be able to cope with that.
I would expect FaaSInitialize to either be a child of APIReceive or both to share common parent span in order to be connected. Can you illustrate in more detail how they are connected together?

3.) Both being separate looks correct to me. Why would you combine them? Those are different operations and concerns being tracked in my opinion.

4.) Adding to my last sentence on 2.), please give more detail on how this operation is invoked. I'd expect APIReceive to have a parent span rather than being an orphan. I would also expect it to be a parent or sibling of FaaSInitialize (or FaaSInvoke, if FaaSInitialize is not happening each time).

[anuraaga]

1. Let me give a few possible scenarios, based on the instrumentation restrictions, and see if they align with what you're saying

a) Use Java library instrumentation of lambda with code change. Instead of using official RequestHandler, which only supports receiving a batch of messages, users use a MessageHandler we provide.

Receive span created around the processing of the batch (instrumenting RequestHandler.handleRequest, which is implemented by us and instrumented by us). Receive span has no links

Processing span for each message (instrumenting opentelemetry-instrumentation-awslambda's special interface, MessageHandler.handleMessage, which is implemented by the user but instrumented by us). Each processing span has a link to its producer.

b) Use Java auto instrumentation with no code change at all. User is using RequestHandler.handleRequest.

Receive span created around the processing of the batch (instrumenting RequestHandler.handleRequest, which is implemented by the user but instrumented by us). Receive span has links to all the producers.

No processing spans

c) Use Java auto instrumentation, but also manually instrument user code to create processing spans. User is using RequestHandler.handleRequest but also calls effectively Tracer.startSpan (more likely, a helper we provide AwsLambdaMessagingTracer.startSpan which handles semantic conventions OK, while processing each message.

Receive span created around the processing of the batch (instrumenting RequestHandler.handleRequest, which is implemented by the user but instrumented by us). Receive span has links to all the producers. (must be same as b, it's auto instrumentation and we have no idea what the user is doing underneath)

Processing span for each message, created by user. Each processing span has a link to its producer.

Observations that make me not totally clear on these, but they're not that big of a deal either is

c) has duplication of links between receive span and producer span
a) isn't consistent with b)/c) aren't consistent since a) doesn't have links on receive span. Should it?

2. Yeah mainly to fulfill the messaging conventions. FWIW, these spans are what particularly matter when understanding the request in the trace since they are business logic. However, they are implemented as API calls in practice. Actually I'm starting to realize the produce span should probably have a duration, and encompass the API call itself, since if the API call fails the production failed too. I wonder if that makes sense as a model.

This brought me a possibly more radical thought - should APIReceive be a SERVER span? It's an HTTP client call - but as car as causal relationship, it's caused by the ApiPublish spans, and normally this looks like a client-server interaction. I'm guessing the answer is no though - there is a different server involved and if it was part of the trace, things are different. It's not part of the trace coincidentally because message queue happens to not be traced, at least in this example.

3. The function is invoked to receive the messages - they're one operation with a cause (message came in) and procedure (invoke function) and don't feel like two spans conceptually. I guess, I might not understand what the actual meaning of a receive span is.

4. Good callout. While I don't think we'd usually have a span for it in practice, since it's an infinite duration span, I added a pollmessages span which hopefully makes it more clear. The runtime is an app that is basically an infinite loop calling APIReceive, and that would be its parent. But since we don't really want an infinite trace I think, each APIReceive is fairly independent, no real cause relationship between each, except for the fact that it happens to be invoked in the same code fragment only. So figured APIReceive is supposed to be parentless in practice.

[anuraaga]

Also fleshed out table and added an ugly ascii time chart :) One key point in the table is a new column on whether a span is created in the user's app or infrastructure (e.g., cloud provider). Infrastructure spans are generally not available, though in my example in practice FaaSInitialize and FaaSInvoke are present. But they're not generated with OTel and it will be years before they follow semantic conventions and should be considered as just parent windows of time. So I realized I was missing yet another span, one in the user's app that is generated by instrumentation so it can have semantic conventions.

[anuraaga]

Noticed a possible corner case in our messaging spec - I don't think in lambda+SQS it's possible to robustly have a name like NiceQueue receive. A function can be registered to multiple queues, so a received batch can correspond to multiple queues, only the messages themselves have a reference to their queue. One option is to check the messages and if they all have the same event source, use it, otherwise something like multiple_sources receive?

[arminru]

@anuraaga

1.a) And the processing spans have the receive span as parent, right? This would be fine then.

1.b) Here the connections to the producing spans are lost. This could be solved by adding links from the receive span to the producing spans. You could also declare that span as processing span (since it actually encloses the processing of all messages) or declare no operation at all. Would you suggest the semantic convention to include these links and optionally also an attribute to define the kind of relationship?

1.c) That sounds legitimate to me and I think the duplication of links is acceptable as auto-instrumentation and user-instrumentation are colliding here.

Since links are allowed to be added for any reason and don't have any semantic defined at all (beyond that links specify spans as "related"), automated analysis could only determine that the span linked from the processing (or receiving) span is the producer of the message by looking at the attributes (messaging.* present) and kind (either producer or client). I think this could even be sufficient but I'm not entirely sure if we should rely on just that. What do you think?

[arminru]

@anuraaga

2.) Yeah I think that having the producing span encompass the lower-level send call (implementation detail of the produce operation) would make sense.

The messaging semantic conventions define whether it should be producer+consumer or client+server by whether it is part of a synchronous conversation (client sends request via a message, server responds to it via a message) or rather a "fire and forget" item of work (producer sends request via a message, consumer fulfills the request but does not render a (direct) response). Does that make sense to you?

[arminru]

@anuraaga

3.) The receive span is intended to track the time spent waiting on a message. If this time spent waiting is not of interest, which is, for example, the case when waiting/polling is hidden away in a framework and the instrumented routine is only invoked at the time a message arrives, the receive span can be omitted. Only a span with no operation attribute being set would be created in this case and would track the operation handling (i.e., processing) the message.

This is explained in the paragraph below the message.operation attribute here:
https://github.com/open-telemetry/opentelemetry-specification/blob/master/specification/trace/semantic_conventions/messaging.md#messaging-attributes

[arminru]

@anuraaga

4.) I see. If this is the main loop of that app, I also don't think tracking it with a span that spans over the entire lifetime of the process would make sense.

[arminru]

@anuraaga Thanks for expanding the table and for the pretty piece of art!

Infrastructure spans are generally not available, though in my example in practice FaaSInitialize and FaaSInvoke are present. But they're not generated with OTel and it will be years before they follow semantic conventions and should be considered as just parent windows of time.

What do you mean by that? Are you saying you made them up for your example but don't expect any spans like these to be created IRL (unless the cloud provider instruments their engine and deploys OTel in production while also exporting the spans to the same back end as the ones from the user domain)?

[arminru]

Noticed a possible corner case in our messaging spec - I don't think in lambda+SQS it's possible to robustly have a name like NiceQueue receive. A function can be registered to multiple queues, so a received batch can correspond to multiple queues, only the messages themselves have a reference to their queue. One option is to check the messages and if they all have the same event source, use it, otherwise something like multiple_sources receive?

@anuraaga I think it makes sense to discuss this separately. I could well imagine the same handler code being invoked at once for multiple sources in batch scenarios like you just described. We might need to consider this in the semconv.

[anuraaga]

Thanks @arminru

1.b) Here the connections to the producing spans are lost. This could be solved by adding links from the receive span to the producing spans. You could also declare that span as processing span (since it actually encloses the processing of all messages) or declare no operation at all. Would you suggest the semantic convention to include these links and optionally also an attribute to define the kind of relationship?

Ah - I wasn't thinking enough about the "no operation" option. I might understand what the receive is now (hopefully, heh). I think the important piece missing from the convention is to indicate this is a batch. I think there are options of

• No operation, messaging.scope = batch
• operation = process, messaging.scope = batch
• operation = process_batch

I lean towards the latter two, it is about processing so makes sense as an operation. But I think one of these is needed to make clear this span is indicating multiple pieces of work, which may have absolutely no relation to each other. Ideally we'll have spans below for individual work which is even better, but if not, we need to know that the span information is lossy.

The messaging semantic conventions define whether it should be producer+consumer or client+server by whether it is part of a synchronous conversation (client sends request via a message, server responds to it via a message) or rather a "fire and forget" item of work (producer sends request via a message, consumer fulfills the request but does not render a (direct) response). Does that make sense to you?

Yeah I think this makes sense, and the unclear part is about parenting. For example, we always consider a CLIENT span to be the parent of a SERVER span in another service. This is a reason for not allowing nested CLIENT spans in instrumentation. I guess we have no similar notion for PRODUCER / CONSUMER? In particular, in our spec we allow nested CONSUMER spans and I want to confirm that's ok (@iNikem does too :) Also, for practical purposes, what is the difference between these PRODUCER and CONSUMER spans, and INTERNAL spans? Since the semantic conventions indicate the messaging semantics, and they don't have parent/child relationship, I'm not sure what these kinds are enabling us to do, where as CLIENT / SERVER allows us to create dependency graphs. Somehow, I feel like messaging.operation is enough (as long as we fill in send of course) without having these span kinds.

The receive span is intended to track the time spent waiting on a message.

Aha! Does this mean in my chart, rather than Receive being a child of FaasFunction, it's supposed to be a child of pollmessages and encompass the time around all of the APIReceive calls? If this is true, a lot of things get clearer for me - however I have a worry that this receive span could be considered an "infinite span" too. Though maybe we have an expectation that a message will come at least once in the future.