15.3 — Publish memory-per-target numbers

[vrabbi]

Phase: Phase 15 — Performance and scale
Proposal: docs/proposals/next-phases.md §15
Priority: P2 · Size: M

---

Problem

docs/operations/capacity.md covers CPU shape thoroughly — compile vs leaves, Convert vs array length, JSONPatch marshal cost, spoke-to-spoke overhead. It says nothing about memory, which is the dimension that actually decides how many targets a replica can hold and what resources.limits to set.

An operator sizing this today is guessing.

Goal

"How much memory per target?" has a published answer.

Design

Measure and publish:

• Bytes per compiled plan, as a function of schema size and rule count. A benchmark with -benchmem over the existing fixture ladder gets most of this.
• Registry footprint at 10 / 100 / 1000 targets.
• Informer footprint before and after the scoping work in 14.1 — Scope the manager's informer caches (it caches every Secret in the cluster) #146 and 14.2 — Scope the webhook-server's XRD and CRD informers #147 — this is the term that dominates today and the reason those two issues matter.
• Peak vs steady state during InitialSync, since compile allocates heavily (Phase 9 measured 4.7 MiB for a single 1000-leaf compile) and the transient peak may exceed the steady footprint by a lot.

Publish a sizing table with a worked example ("1000 targets averaging 200 leaves ⇒ approximately X MiB steady, Y MiB peak ⇒ set limits to Z"), and state the hardware and Go version the numbers came from, as the existing page does.

Add a make bench-mem target so the numbers are reproducible rather than folklore.

Acceptance criteria

• Bytes-per-plan benchmark over the fixture ladder
• Registry footprint at three target counts
• Informer footprint measured before/after scoping, coordinated with 14.1 — Scope the manager's informer caches (it caches every Secret in the cluster) #146 and 14.2 — Scope the webhook-server's XRD and CRD informers #147
• Transient peak during initial sync measured
• Sizing table with a worked example in docs/operations/capacity.md
• make bench-mem target
• Default chart resources reviewed against the findings and adjusted if they are wrong

---

Generated by Claude Code