WIP - InteractionProfilerGrain

[KSemenenko]

First step for #344

Filter + Grain

[SebastianStehle]

Can we start to use records for that?

[SebastianStehle]

Can we not extend the current filter for that? Because we have some logic there that we should keep. Like custom method names and attributes to opt out from the profiler.

[SebastianStehle]

In the current profiler we use batching to reduce the grain calls. I think we should use the same approach here.

[SebastianStehle]

Hi, I added a few comments.

[richorama]

I think I would prefer not to concern the back end with front end configuration, such as colours. Could we just return pure data as JSON to the front end?

[richorama]

I think I would like a way to control (switch on/off) this profiling, as it may get slow?

[SebastianStehle]

Can we close this?

[KSemenenko]

I can't figure out how to do it well to store the history of calls

[SebastianStehle]

Why do you want to store the history? The Dashboard is only designed to show either the present or data over the full lifetime of the silo. If you can provide a view that shows the interaction over the full lifetime it would provide value.

For detailed analysis you should use telemetry anyway.

[KSemenenko]

I'd like to show who's calling whom.
For debugging, so to speak.

I hadn't thought of this as a long term storage.

my case is to look for deadlocks or delays.
Sometimes there are too many grain calls to each other and I would like to see this call tree.

[KSemenenko]

Here, most likely, by storage, I mean how to collect statistics and pass them to the UI @SebastianStehle

[SebastianStehle]

I think we cannot analyze single calls. This would be responsibility of new relic or application insights. We can only collect general call structure.

So basically the following table:

SourceGrain | SourceMethod (if possible) | TargetGrain | TargetMethod | Count | CountSuccess | CountFailed | TotalTimeNs

So basically the following dictionary

record ProfilerKey (
  string SourceGrain,
   string SourceMethod,
   string TargetGrain,
   string TargetMethod);
   
record ProfileValue (
  long CountSuccess,
  long CountFailed,
   long TotalTimeNs); // Could store 9 years of compute time or 9000 years for ms)
   
Dictionary<ProfilerKey, ProfileValue>

[KSemenenko]

I found an interesting side effect
when using these filters, if requests were between different silo instances, the request will not be executed due to an error because there is no context

[KSemenenko]

Seems I found Idea for it