// the find
parca-dev/parca-agent
eBPF based always-on CPU/GPU profiler auto-discovering targets in Kubernetes and systemd, zero code changes or restarts needed!
Parca Agent is an eBPF-based sampling profiler that attaches to processes on a Linux host or Kubernetes cluster without any code changes or restarts, producing pprof-compatible CPU profiles. It's aimed at platform/SRE teams who want always-on production profiling across a mix of languages (Go, Rust, Java, Python, Ruby, Node, etc.) rather than developers profiling a single app locally.
Zero-instrumentation profiling via eBPF at 19Hz is the real draw — it covers native code, JIT runtimes, and interpreted languages with per-language unwinders instead of the usual 'Go/JVM only' limitation most eBPF profilers ship with. The relabeling system is modeled directly on Prometheus's relabel_config and supports both legacy Parca-style labels and OpenTelemetry semantic convention names, so you're not locked into one metadata vocabulary. Test coverage includes actual benchmarks for profile encoding/decoding and conformance tests against the pprof format, which is more rigor than most profiling agents bother with.
The feature list oversells current reality: network usage and allocation profiling are marked 'soon' in the README, so today it's CPU-only despite being pitched as a general profiler. It needs root or CAP_SYS_ADMIN plus a 5.3+ kernel with BTF, which is real friction on hardened nodes or older managed Kubernetes offerings. The CLI flags are littered with deprecated options still present and undocumented as to whether removal is imminent (memlock-rlimit, dwarf-unwinding-disable, object-file-pool-*), making it hard to know what's actually load-bearing versus legacy cruft. And it's really half of a two-component system — without a Parca server to send profiles to, you're limited to local HTTP endpoints with no historical querying.