Thanks for this, @dentiny — enabling profiling to debug OOM (#2016) is genuinely useful and something we want. I have to request changes, though, because as written this exposes /debug/pprof in a way that is a security regression for every existing install. Details and a recommended path below.

Why this can't merge as-is

import _ "net/http/pprof" runs the package init(), which registers /debug/pprof/* on http.DefaultServeMux. Both HTTP servers here serve that mux via the nil handler:

• pkg/metrics/operator/machinery.go:110 — operator metrics, default :9999, bound 0.0.0.0 (all interfaces) unconditionally.
• pkg/metrics/clickhouse/rest_server.go:207 — metrics-exporter, default :8888, 0.0.0.0 in the default (Plain) IPC mode.

Both ports are published cluster-wide by the clickhouse-operator-metrics Service and scraped by Prometheus, and we ship no NetworkPolicy. So after this change any pod in any namespace can reach /debug/pprof on the operator, unauthenticated:

• Info disclosure — /debug/pprof/heap and /allocs dump live operator memory, which holds ClickHouse credentials, IPC tokens, and Kubernetes API bearer tokens; /cmdline leaks all flags; /goroutine?debug=2 leaks full stacks.
• DoS — /debug/pprof/profile?seconds=N and /trace?seconds=N take an attacker-controlled duration and share the single listener/mux with /metrics, degrading the reconcile loop.
• Bypasses IPC Secure mode — the token + loopback middleware wraps only the /chi handler; pprof is registered at the mux root, outside that protection, even in Secure mode.
• On by default — a blank import active at init(), with no flag/config/opt-in, so every install silently gains this endpoint on upgrade.

A "just use a separate mux" fix is not sufficient

Moving pprof to a private *http.ServeMux alone does not help: net/http/pprof's init() registers on the global DefaultServeMux regardless of whether the import is blank or named, so as long as machinery.go / rest_server.go serve the nil handler they keep exposing it. (Verified: a named import still leaves DefaultServeMux responding 200 on /debug/pprof/.)

Recommended approach for a mergeable version

1. Don't register on DefaultServeMux. Simplest is to use runtime/pprof (no init() side-effect) to serve heap/allocs/goroutine — the OOM-relevant profiles — from a private mux. Alternatively keep net/http/pprof but convert the :8888/:9999 servers to explicit, non-nil muxes so the polluted default mux is never served.
2. Bind loopback-only (127.0.0.1), never on the metrics ports and never behind the Service — reachable via kubectl port-forward.
3. Default OFF, opt-in via a flag, consistent with the existing -debug / -fips-info flags.
4. Consider omitting profile/trace (the CPU/trace DoS vectors) unless separately gated, and log a warning when pprof is enabled so it is visible in kubectl logs.

Happy to review a revised version along these lines — thanks again for tackling the OOM-debuggability gap.