ScopeManager for event-loop based frameworks

[ikurovsky]

Current interface for ScopeManager assumes that there is an implicit link to a certain context working as a storage for active scope/span. For thread-pool oriented applications the default ThreadLocalScopeManager storing the active scope in a thread local variable is a good fit.

How is it supposed to be implemented for models such as Netty/Vert.x where thread-locals are not an option? Usually such details are stored in some kind of request context, but to my knowledge there is no way in Netty to get a current context without an explicit reference and so it is not clear how to implement the ScopeManager#activeSpan() method assuming there's only one tracer and one scope manager.

[tylerbenson]

I don't know of a good way to do this without having a static reference to the context, likely via a threadlocal. Ratpack exposes something like this, Maybe you could add similar hooks into netty?

[ashleymercer]

There are two approaches that I've seen for passing context around (my use case is Scala, where a lot of code is written async, and callbacks aren't guaranteed to execute on the same thread):

• pass a separate "context" object around everywhere: this is the approach taken by e.g. the AWS Java SDK (at least in version 2). See for example their code to intercept and trace calls to AWS services: the ExecutionAttributes object is essentially a strongly-typed map, and can be used to store / retrieve objects across threads:

  • makes everything explicit, so client code has control
  • easier to test / mock
  • requires adding context object to method signatures everywhere you need it

• use some kind of AOP to copy trace data between threads. For example, the Scala kamon library uses aspectj to copy trace context data between threads

  • much less intrusive, doesn't require changing existing code
  • can be made to work with third-party libraries
  • trickier to set up / get working correctly
  • difficult to test / mock reliably

In short, you're always going to require some way of passing around a reference to the "current" trace. With ThreadLocal that knowledge is implicitly defined by the current thread, but this breaks down in async / event-based code.

[Hohenheimsenberg]

• much less intrusive, doesn't require changing existing code
• can be made to work with third-party libraries
• trickier to set up / get working correctly
• difficult to test / mock reliably

I like the second approach, the problem with it is that there are libraries, like couchbase client that uses the GlobalTracer to access the active span. Also implementations like Jaeger always builds spans as child of the active span if not explicitly specified not to.

I haven't tested it, but on Spring webflux I think that building a bean with a request scope could work.