SLO Culture

An SLO that lives only in Prometheus changes nothing. The recording rules from Topic 50 and the burn-rate ladder from Topic 52 are machinery; what turns the number into decisions is the machinery around the machinery — a monthly report the whole org reads, a written rule for what happens when the budget runs out, and the discipline to start with one SLO instead of twenty.

This closing topic is about that cultural layer, and about its two hardest sells. The first is that 100% is the wrong target — not an ambitious one, a wrong one. The second is that a single honest SLO beats twenty mediocre ones, which runs against every instinct a program rollout has.

The Monthly SLO Report

One page per SLO: the target, the attainment, the budget spent and what spent it, the trend against the previous window. It goes to engineering, product, and leadership on the same day. Topic 51's whole argument was that the budget works as negotiation currency — and currency only functions when both sides of the negotiation can see the balance. A report that circulates inside the SRE channel is telemetry without an audience: product keeps making roadmap calls without the one number that was supposed to inform them.

The Decision Rule, Written Down

Harborline's policy fits on a page. Budget exhausted: feature releases for the checkout journey pause, and the next sprint's capacity goes to reliability work until the SLI holds its target for seven consecutive days. Below 50% remaining mid-window: risky deploys need a second reviewer. That is the whole document — and its value comes from when it was written, in calm weather, signed by engineering and product before the first window ever opened.

The alternative is negotiating consequences during the incident that spent the budget, which produces whatever the loudest person in the room wants. With the rule pre-agreed, enforcement is mechanical rather than political: nobody is punishing anyone, the policy is simply executing, the way a merge freeze executes when CI goes red.

Budget healthy, plenty remaining→Ship the risky change

Below 50% remaining mid-window→Risky deploys need a second reviewer

Budget exhausted→Freeze checkout features; sprint buys reliability work until the SLI holds target 7 days

The Case Against 100%

Each added nine costs roughly ten times the engineering effort of the last — redundancy, failover drills, multi-region state — while the improvement shrinks below what users can perceive. A customer booking a ferry on a flaky phone over terminal Wi-Fi experiences last-mile failures at rates well above 0.1%; past a point, extra nines vanish into that noise. And 100% exactly forbids every deploy, every maintenance window, and every dependency failure, so the first incident turns the target into a joke and takes the SLO program's credibility down with it.

The Dependency Ceiling

Availability also has a hard ceiling that no amount of engineering moves: the things checkout must synchronously call. Two serial 99.9% dependencies already cap the caller near 99.8% — 0.999 × 0.999 — before its own code fails once. Harborline's card processor publishes a 99.95% SLA, an allowance of about 20 minutes per 28 days. A 99.99% checkout SLO would grant itself 4 minutes of budget while depending on a service allowed to be down for 20: the processor's own allowance could spend the budget five times over, leaving the bookings team accountable for a number they cannot influence. The SLO sits below the dependency ceiling, or it stops being an engineering target at all.

One SLO First

The tempting rollout is an SLO per endpoint — twenty targets, a wall of dashboards, a program launch. Twenty simultaneous SLOs produce twenty arguments about SLI definitions and zero changed decisions: each one needs its recording rules, its alert ladder, and its report maintained, no team sustains that at twenty, and the stale ones poison trust in the accurate ones.

Harborline runs checkout alone for two full 28-day windows. The first report moves a feature release behind the follow-through on the Chapter 8 connection-pool fix — the Topic 51 meeting — and that one changed decision, not a slide deck, is the argument for expansion. Search gets its SLO next, carried by evidence that the loop demonstrably works.