This message was deleted.

Hi! Do you mean custom spans? The key is to correctly use the trace API, especially parent spans and trace_id. cc @Aaron Pham

I believe spans should already be there. Would you mind shareing what type of trace you are looking for?

so, I have a service (it has no runner for now as we work out the interfaces/tracing etc) and a python function which looks as such (it's pseudo code, it works)

bentoml_logger = logging.getLogger("bentoml")
@svc.api(...)
async def predict(x):
    [...]
    <http://logger.info|logger.info>("blablabla")
    [...]
    return ...

I also have the jaeger all-in-one docker running besides the service and tracing enabled using grpc However, when I do a request to the service, the trace only contains 4 spans: • one for the whole predict function • one named /predict http receive • one named /predict http send presumably for the response headers • one named /predict http send presumably for the response body I would like the log outputted by logger.info("blablabla") to be part of the trace. The text version of the logs contains the trace id ie (

blablabla (trace=...

) but it doesn't show in the trace from the jaeger's interface. Does this make sense? I have seen that bentoml already uses tools for opentelemetry instrumentation but I would have expected it to work for the bentoml logger, from what i've seen in the logging docs. Thank you!

Hmm, I’m not sure if this is the expected behaviour for user to use our logger like this.

bentoml_logger

is reserved specifically for BentoML internal. To include tracing from within the API function call we probably need to do more support wrt to logs