Watch interrupted during high performance

[HermansG]

We have an Actor that calculates results in approx. 10ms.
The calculation was started with .NET's Task.StartNew combined with PipeTo(Self).
The FrontEnd has a Watch with the parent of the Actor.
We encountered situations with high perfomance > 1000 calculations in a row where the Watch failed at the FrontEnd.
We upgraded 1.4.12 to 1.4.14 with the same result.
We changed the Task.StartNew and used child actors instead.
Our idea was that it would be better to use resources inside of the Akka system instead of outside with Windows Task's.
Observation was that the calculation was a fraction slower but the stability improved. But still the errors could occur.
We then added watch-failure-detector to the Akka configuration.
threshold = 100.0, acceptable-heartbeat-pause = 20 s
This improved the behaviour enough.
Still it is theoretically possible that under high pressure (high CPU and memory) the Watch can fail.
What do you think about this our try to improve stability?
Are there other improvements that you can think of?

[Aaronontheweb]

@HermansG does the watch fail because it never gets sent across the wire? Or because the watching actor never receives a Terminated?

(I know it might be tough to gather that information, but it would help us understand where the source of the problem might be)

[HermansG]

I saw it, I was not able to find the answer. Currently the system is stable again, I am not experienced enough to tackle the underlying mechanisms.

[Aaronontheweb]

You might have been affected by this issue: #4849

Those fixes will be included as part of Akka.NET v1.4.19.

[Aaronontheweb]

Was this issue resolved for you eventually @HermansG ?

[HermansG]

Indeed resolved: no problems currently. It was difficult to understand then (now almost a year ago).

[Aaronontheweb]

Some of these issues with the protocol can be difficult to track down but I'm glad this is resolved.