[dalehamel]

Stack

1. #73 → #50
2. dale/offcpu-omnibus-probe-api — minimal Probe runtime foundation
3. This PR — duration-filtered off-CPU probe + task-state labels

What

Add off_cpu_time as a dynamic custom probe on the LuaJIT omnibus stack. It retains the upstream built-in profile's duration-valued semantics, while adding:

• duration-aware min_duration filtering before unwind/export;
• completion-side sample rate;
• migration-safe global LRU correlation;
• configurable map capacity and a safe 1ms default minimum;
• per-path health/drop metrics;
• thread.state labels separating voluntary blocking from runnable/preempted scheduler delay.

Important semantic clarification

The built-in upstream off_cpu profile already emits one selected wakeup event weighted by exact elapsed nanoseconds. This probe does not change count→duration. Its differences are selection timing, migration correctness, state visibility, and operational accounting. The README documents both paths and the no-inverse-scaling caveat.

Minimal ABI adaptation

The Hack Days implementation carried thread state through the async I/O ABI. This clean stack instead adds one generic probe_user_data field and an optional per-origin TraceLabeler. It has no dependency on async I/O, process observers, lock-wait, memory profiling, or other Hack Days probes. Probe-private ELF resources follow the <configured name>_* prefix contract and are removed from the main collection by the caller-supplied dynamic resource prefix.

Configuration

custom_probes:
  off_cpu_time:
    min_duration: 5ms  # safe default is 1ms; 0s is explicit exhaustive mode
    sample_rate: 1
    max_entries: 65536

Validation

• clang-17 deterministic amd64 + arm64 BPF rebuild; format lint green.
• Unit/build/vet sweep across offcpu, generic probes, tracer, support, reporter, traceutil, controller.
• Privileged arm64 integration: positive duration+hybrid-stack+thread.state=interruptible path in 0.92s; 10-minute threshold emits no short waits.
• Existing one-node staging tophat of the equivalent HD40 implementation: 585,084 interruptible rows / 140,162.52s, plus idle, uninterruptible, running, and preempted states, filterable as process.context.label.thread.state.

Upstream convergence

This remains a separate profile while we evaluate upstreaming the useful pieces into built-in off_cpu. Upstream history already requests duration-threshold sampling (open-telemetry#144) and documents the per-CPU migration amplification (open-telemetry#1348). Do not enable both built-in off_cpu and off_cpu_time simultaneously unless deliberately comparing them.

[github-actions]

This PR was marked stale due to lack of activity. It will be closed in 14 days.

[github-actions]

Closed as inactive. Feel free to reopen if this PR is still being worked on.