Previously, we allow for BPF to be optionally enabled and provide
transparent fall-back to non-BPF samplers if BPF initialization fails.
This can lead to uncertainty around metric meaning since it makes
the data provenance unclear. Further, the sampler overhead depends
on whether the BPF or fall-back sampler is used with much higher
costs for non-BPF samplers.
Solution
Changes BPF to be non-optional on Linux. Drops fall-back samplers
so that metric provenance is clear and so sampler overhead is
consistent.
Result
A simpler configuration and consistent data provenance and
sampler overheads.
Problem
Previously, we allow for BPF to be optionally enabled and provide transparent fall-back to non-BPF samplers if BPF initialization fails. This can lead to uncertainty around metric meaning since it makes the data provenance unclear. Further, the sampler overhead depends on whether the BPF or fall-back sampler is used with much higher costs for non-BPF samplers.
Solution
Changes BPF to be non-optional on Linux. Drops fall-back samplers so that metric provenance is clear and so sampler overhead is consistent.
Result
A simpler configuration and consistent data provenance and sampler overheads.