// the find
performancecopilot/pcp
Performance Co-Pilot
Performance Co-Pilot is a mature, Linux-first framework for collecting, archiving, and analyzing system performance metrics across distributed hosts. It predates most modern observability stacks by decades and shows it — the architecture is built around its own daemon (pmcd), its own archive format, and a large ecosystem of PMDAs (Performance Metric Domain Agents) for collecting from specific subsystems. The target audience is sysadmins and performance engineers dealing with bare-metal or VM workloads, not teams already invested in Prometheus/OpenTelemetry.
The archive format is genuinely useful — you can record metrics locally and replay them later with the full tool suite, which beats most modern stacks for post-incident analysis on air-gapped or resource-constrained systems. The PMDA model is well-designed: each data source is an isolated agent with a defined wire protocol, so adding a new metric source doesn't require touching core daemons. Grafana integration via grafana-pcp is production-quality, not an afterthought — dashboards and a native datasource exist and are actively maintained. The build matrix is unusually thorough: the CI covers Amazon Linux, CentOS Stream, Debian, Fedora, Ubuntu, and macOS, which matters for a low-level C project.
The learning curve is steep and the documentation assumes you already speak PCP. The conceptual model (pmcd, pmlogger, pmproxy, PMDAs, PCP archives, metric namespaces) requires significant orientation before you can do anything useful, and the quick-start guides don't really shorten that path. The autoconf/GNUmakefile build system is a time capsule — no CMake, no Meson, just configure scripts and recursive makefiles that are painful to navigate if you're used to modern tooling. If your stack is already on Prometheus, adopting PCP means running a parallel collection infrastructure with its own storage format; the export bridges (pcp2json, pcp2influxdb, etc.) exist but they're one-way and add operational overhead rather than replacing anything. Star count relative to age (1064 stars for a project this old and capable) signals that it's valued by a small loyal community, not widely discovered.