search

opentracing / specification

A place to document (and discuss) the OpenTracing specification. 🛑 This project is DEPRECATED! https://github.com/opentracing/specification/issues/163
http://opentracing.io/spec
Apache License 2.0
1.17k stars 182 forks source link

Finish time of child Span (ChildOf reference) #139

Open lookfwd opened 5 years ago

lookfwd commented 5 years ago

image

Wondering about Child D from here. it's a bit weird, isn't it?

eyjohn commented 5 years ago

To me, it comes down to whether the data model sets a constraint on:

A child span can not outlive its parent

I do not see any references to this in the specification, and I would suggest that if this is not a requirement that it actually be explicitly stated that this is not guaranteed, rather than some-one inferring it from the diagram.

As a side-point, I do see the appeal of adding such a constraint on the data-model as it does capture the nature of request/response based systems well, i.e. you can guarantee that all child requests have returned by the time the response gets sent back (timeouts + gradual degradation being the exception perhaps).

However in-practice, enforcing such a constraint in the framework is a bit tricky, especially given that say clock skew between two processes can invalidate this when looking at timestamps. Likewise it imposes a potentially difficult implementation of the tracing model. As such I would suggest this be seen more as a Best Practice rather than data model constraint.

yurishkuro commented 5 years ago

There is no requirement that child finishes before parent. The most common use case for this is when parent is the client span and child is the server span. The client may have a timeout that would cause the parent client span to be closed before server completes processing of the request.

In other words, childOf reference implies parent's dependency on the outcome (and one of the outcomes is "server did not finish in time"), but not the causal relationship child.end happens-before parent.end

eyjohn commented 5 years ago

The client may have a timeout that would cause the parent client span to be closed before server completes processing of the request.

This as interesting example and I think it would be great to explicitly include the characteristic and example in the reference.

jmilkiewicz commented 3 years ago

Wouldn't be the same for Child Span A and Child Span B where the former is considered a parent?