How tracing telemetry fit in runc to capture performance and behavior metrics (in DEBUG mode)

[Dentrax]

While traversing around the issues, I noticed an issue #3181. And one of the notes that pointed in there is: figure out why the difference is so high.

So, I thought that how could we trace performance and behavior more easily for single command in DEBUG mode? There are some kinds of tracing methods:

• OpenTelemetry
• OpenCensus
• OpenTracing

I think tracing can be easily handled by higher level runtimes ¹. Such as CreateContainer, KillContainer, etc. are high level metrics which we do not know what's happening in deeper levels.

Would it be a bit overengineering to trace syscalls, IO operaions, etc.? Do we already have a tracing strategy? Since we already have benchmark strategies, having a tracing method in entire runtime really worthy? Waiting your thoughts.

cc: @dashpole @kolyshkin

Footnotes

1. https://github.com/containerd/containerd/issues/3057 ↩

[kolyshkin]

Do we already have a tracing strategy?

I am mostly using strace, and have played with Go's pkg/trace (together with go tool trace).

I'm not sure we need distributed tracing in runc, and the projects you refer to are all about distributed tracing AFAICS.