Service level objectives

Issue #17. Numbers marked Target are design goals. Numbers marked Measured come from benchmarks/ on the hardware named in the baseline file. FluxVM-dependent figures must be labeled as such.

Control plane (zyvor-fabricd)

SLO	Target	Window	Measured
GET /health availability	99.9%	30 d	not yet from production telemetry
GET /readyz success when store + FluxVM up	99.9%	30 d	CI smoke only
API p99 (auth'd JSON, in-process / local)	≤ 50 ms	5 min bench	see benchmarks/baselines/
Concurrent inventory list (GET /api/vms) p99	≤ 200 ms at 500 VMs	bench	harness inventory mode

VM lifecycle (FluxVM-dependent — label every published number)

SLO	Target	Notes
Cold start (pre-pulled qcow2, 2 vCPU / 2 GiB)	p50 ≤ 8 s, p99 ≤ 20 s	FluxVM + image cache
Stop → stopped	p99 ≤ 10 s	ACPI then force
Snapshot disk-only (qcow2 internal)	p99 ≤ 30 s	not full memory dump
Backup create 10 GiB disk, local, gzip	p99 ≤ 180 s	backup crate path

Failover / maintenance

Objective	Target	Evidence
Host evacuate plan (no I/O)	deterministic, largest-first	host-lifecycle unit tests
Mid-migration failure	job failed, source stays cordoned	migration_failure_marks_job_failed_and_leaves_host_cordoned
Quorum loss (file heartbeat majority)	check_quorum → false at ≤ majority	fault-tolerance::quorum tests
RTO after node death with shared storage + FT enabled	unmeasured — do not quote	needs chaos lab (#15 live track)
RPO for incremental backup	last successful backup timestamp	metadata in backup crate

Error budget policy

• Publish a baseline JSON before claiming a number in README or sales decks.
• FluxVM-dependent rows must include fluxvm_version and labeled_fluxvm_dependent: true.
• If a live bench cannot run (no daemon), the harness writes status: offline and must not be copied into FEATURES.md as a measurement.