http_server sync response brokered with an other output and batching

[sagikazarmark]

I have the following pipeline:

input:
  http_server:
    address: 0.0.0.0:4196
    path: /api/v1/events
    sync_response:
      status: '${! meta("http_response_status").or("204") }'

output:
  broker:
    pattern: fan_out
    outputs:
      - sync_response: {}
        processors:
          - mapping: root = null
      - http_client:
          # ...
          batching:
            count: 20

The problem I have is reaching the desired count may take a long time and HTTP requests time out.

The reason I use sync response in the first place is being able to modify the response code (also, I have a couple processors validating the messages and I return an error in the response if the request is invalid).

Ideally, I'd like to return the response without waiting for the http_client request to be sent.

Is that possible?

Thanks!

[mihaitodor]

Hey @sagikazarmark 👋 I wonder if the sync_response processor might be of use here. Make sure you place your processor pipeline outside of the broker output, maybe at the input.processors or pipeline.processors level.

[sagikazarmark]

@mihaitodor according to https://www.benthos.dev/docs/guides/sync_responses

However, it is important to keep in mind that due to Benthos' strict delivery guarantees the response message will not actually be returned until the message has reached its output destination and an acknowledgement can be made.

I found this issue: #1946

It seems to be working, but it feels a bit hacky

[mihaitodor]

Oh, good point. I forgot that you'll also need to toss a buffer in the mix (I think it might be doable in a single config if you have the sync_response processor at the input level, before the message reaches the buffer, but I haven't checked).

[sagikazarmark]

I tried adding a buffer, but I haven't tried adding processors at the input level.

I do need a couple of processors to validate and unarchive the input though. Are there any performance penalties for doing so on the input level?

[mihaitodor]

Just keep in mind that processors at the input level are not executed in parallel like the ones under pipeline.processors. I think the parallel processor can be used in such cases, but you'll have to configure input batching for it to be able to do its thing.