Conversation
... while also moving the refcounting to the util package and creating a cleaner path to something different should we want it.
bhs
left a comment
bhs
left a comment
There was a problem hiding this comment.
(self-comments as a sort of guided tour)
| * RPC will be unfinished but blocked on IO while the RPC is still outstanding. A {@link Scope} defines when a given | ||
| * {@link Span} <em>is</em> scheduled and on the critical path. | ||
| */ | ||
| public interface Scope extends Closeable { |
There was a problem hiding this comment.
kinda/sorta replaces ActiveSpan
| * Defined as equivalent to calling {@code tracer.scopeManager().activate(this)} on the {@link Tracer} instance | ||
| * that started this {@link Span} instance. | ||
| */ | ||
| Scope activate(); |
There was a problem hiding this comment.
only meaningful change to the Span interface... this is just sugar and doesn't need to be here (but it does save a lot of typing and complexity).
| public interface Tracer { | ||
|
|
||
| ScopeManager scopeManager(); | ||
| void setScopeManager(ScopeManager scopeManager); |
There was a problem hiding this comment.
note that this can be set – def different than inheriting as we did before
| * // (Do work) | ||
| * span.setTag( ... ); // etc, etc | ||
| * } // Span finishes automatically unless deferred via {@link ActiveSpan#capture} | ||
| * } // XXX Span finishes automatically unless deferred via {@link ActiveSpan#capture} |
There was a problem hiding this comment.
NB: I did not update most comments
| /** | ||
| * Created by bhs on 8/5/17. | ||
| */ | ||
| public class AutoFinishScopeManager implements ScopeManager { |
There was a problem hiding this comment.
This class demonstrates that the refcounting can be built entirely outside of the core OT API
| * or implied. See the License for the specific language governing permissions and limitations under | ||
| * the License. | ||
| */ | ||
| package io.opentracing.usecases; |
There was a problem hiding this comment.
(misc: decided to get started with a "usecases" package/module)
| /** | ||
| * Created by bhs on 8/1/17. | ||
| */ | ||
| public class ThreadLocalScopeManager implements ScopeManager { |
There was a problem hiding this comment.
the simple case (note that this does not need to be part of the core OT API!)
| // Nothing to do. | ||
| } | ||
|
|
||
| // Make sure the Span got finish()ed this time. |
There was a problem hiding this comment.
i.e., the refcounting can still work.
| * @see ActiveSpan | ||
| */ | ||
| ActiveSpan startActive(); | ||
| Scope startActive(); |
There was a problem hiding this comment.
I was thinking; maybe with a decent observer definition, could it be feasible to let the ScopeManager implementation decide whether to auto-activate started spans?
If it can register itself as observer to started spans, it can make the decision to auto-activate or not (maybe even based on some inspection of the started span).
That way for the instrumentation the tracer API basically becomes simple again (pre-0.30), while affording adding such complexity if desired in the scope manager as implementation decision.
Just thoughts basically, would have to play with things to test feasibility...
There was a problem hiding this comment.
I suppose activation might be a decision of specific framework instrumentation that will not be aware if auto-activation is enabled or not.
| // We can't use 1.7 features like try-with-resources in this repo without meddling with pom details for tests. | ||
| try (AutoFinishScopeManager.AutoFinishScope scope = autoFinishActivator.activate(span)) { | ||
| // Take a reference... | ||
| continuation = scope.defer(); |
There was a problem hiding this comment.
Not having the defer and Continuation as part of the API will be a problem, as it means any async based framework instrumentations will be bound to the AutoFinishScopeManager implementation.
If we want to reduce the number of components, how about having Scope.defer() return a deferred scope which can then be activated in the new context? So the Scope is either activated immediately on startActive() or explicitly via its activate() method - but can only be activated once.
|
Summarized the changes and testing results of this PR here: https://gist.github.com/carlosalberto/28ba04ae3201d87f91eff924e7d427da That will get an overall idea for those who haven't been lately following the discussion, as well as for those interested in any further changes to this design - specially the And hopefully this will get things rolling again (feedback, for a start) :) |
|
Anyone interested in proposing changes to Java API, please use https://github.com/opentracing-contrib/java-examples to ensure all identified use cases are being addressed. BTW I recommend forking Carlos's branch if you are working on the current PR: https://github.com/carlosalberto/java-examples/tree/bhs/scopes-update |
| * {@link Scope} instance after this is only defined to do no harm to the process (i.e., not-crash). Again, this is | ||
| * only intended for use during initialization. | ||
| */ | ||
| void setScopeManager(ScopeManager scopeManager); |
There was a problem hiding this comment.
It would be nice not to require that in the API, since it can always be a part of implementation-specific initialization sequence. Do we know of a use case where this method is absolutely required in the Tracer API?
6 participants
(deliberately not opening up a PR against the OT version of this repo... for casual inspection only)
This change is