undertow_http.request span does not always fully enclose it's child spans

[shiyanhui]

Hi, recently I tested the latest agent version, i.e. dd-java-agent-0.100.0.jar and found in the APM that the undertow-2.0 instrumentation seems have some problem.

Please see this screenshot below, the undertow_http.request span is created by undertow-2.0 instrumentation, and the StatsController.ready span is created by spring-webflux instrumentation. I understand that the correct behavior is that the undertow_http.request span should be longer than the StatsController.ready span, right?

Another problem is that we didn't find the servlet.request span which is created in undertow-2.0 instrumentation, why?

Thanks!

[devinsba]

1. We are aware of this timing discrepancy, we haven't been able to address this yet but it is on our radar. As for suitable for production, if those ~200microseconds are important to you and your metrics then not at this time, otherwise it should be fine.

2. You don't see a servlet.request span because you are using webflux which does not work on top of a servlet engine so would be added into undertow as a handler

[shiyanhui]

We are using undertow and webflux in our tech stack, so how can we get the servlet.request span or trace in the APM? I mean, is there any workaround way to get it?

[devinsba]

Assuming this is using spring-boot you would probably have to configure it differently. I can't find an example of that configuration though. See: https://docs.spring.io/spring-framework/docs/current/reference/html/web-reactive.html#webflux-server-choice

For Undertow, Spring WebFlux uses Undertow APIs directly without the Servlet API.

In this case we won't create a servlet.request span because the servlet handler will not be used. From what I can tell you would have to use embedded tomcat or jetty to get the servlet.request span. Is having a span with that operation name required for you, or can you instead use undertow_http.request?

[shiyanhui]

Most our services use spring-mvc+undertow and several services use webflux+undertow, and we configure the monitoring metrics in one place and we want the query syntax to follow the same pattern. So it's better to emit the servlet.request trace.