Examine possibility of starting all spans after routing decision

[anuraaga]

Currently, http.route or a derived span name cannot be used in sampling. This does seem quite limiting as route would probably be a useful, if not the most useful, attribute to sample on.

open-telemetry/opentelemetry-java#4498 (comment)

We did put a lot of effort into RouteHolder, etc, but I wonder if we ever considered starting spans only after route is known, or alternatively having a StartTimeHolder to keep the current timings but still create the span after routing. Or perhaps we just keep a SpanBuilder in context until the route is ready.

[trask]

do we know if this is mostly a Java instrumentation problem, or if the popular web frameworks in other languages also suffer this problem (routing decision occurring in a different component / disconnected from the request start)?

one downside to holding the span start is that we would lose correlation with any logs which occur between request start and routing decision (I've seen people enable logging in their web framework and expect those logs to be correlated).

[mateuszrzeszutek]

one downside to holding the span start is that we would lose correlation with any logs which occur between request start and routing decision (I've seen people enable logging in their web framework and expect those logs to be correlated).

Potentially, it's more than just logs: e.g. there could be some custom servlet Filter code that needs to be traced before Spring controller gets matched. Or, in case of JAX-RS, you might have nested controller annotations and only want to choose the last one (but obviously you don't know which one is the last before the request ends).

[trask]

potentially related: open-telemetry/opentelemetry-specification#3678 (comment)

Jaeger supports a way to defer a sampling decision until the first child span starts

[trask]

Closing, can revisit later if anyone has renewed interest.