Delivery runs on weekly dashboards and quarterly rhythms, and both are too close to the ground to answer the year-scale questions: is the architecture still right? Is the money going where the value is? Are the same incidents repeating? Once a year, the estate deserves a full accounting — not a report that gets filed, but a working session that ends in decisions with owners. This guide is that session: the evidence pack, the honest reading, and the loop into next year. It is also, fittingly, where this hundred-guide library’s threads all meet.
What the review is for (and not)
The annual review answers five questions no weekly view can: Did reality match plan — traffic, spend, reliability — and where it didn’t, was the plan or the execution wrong? Are incident and cost patterns telling a structural story that individual postmortems missed? Is the architecture — single or multi-CDN, the current providers, the current split — still the right answer to this year’s facts rather than the facts of whenever it was chosen? Is the commercial position (contracts, commits, leverage) where the coming year needs it? And is the team’s knowledge and tooling keeping pace with the estate? What the review is not: a blame exercise (it reads systems, not people), a status meeting (nothing weekly belongs in it), or a document deliverable (the deliverable is section four). Half a day, the delivery team plus the budget owner, once a year — scheduled to land a quarter before your biggest contract’s renewal notice date, for reasons that become obvious.
The evidence pack: a year in six exhibits
If the year’s rhythms ran, the pack assembles itself in an afternoon. Exhibit one, demand: the year’s traffic against the capacity plan — baseline, peaks, regional drift, forecast error. Exhibit two, money: spend against forecast, the allocation table by product, effective rates over the year, and what cost work actually banked. Exhibit three, reliability: SLO attainment by month, the incident log grouped by theme (not chronology — themes are the point), and the credit ledger against measured impact. Exhibit four, security and hygiene: the four quarterly audits’ findings and their closure rate, plus the drift log if you run multiple platforms. Exhibit five, vendors: the QBR registers’ completion rates, performance trends per provider, and the contract calendar with every renewal and notice date. Exhibit six, the estate itself: the one-pager’s year of diffs — what got added, what got odd, what nobody would build that way today.
The session: reading the year honestly
Walk the exhibits in order, but read for patterns, not lines — the review’s value is precisely the correlations the year’s close-up views couldn’t see. The recurring kinds: forecast misses that repeat in one direction (a calibration debt, not bad luck); incident themes that cluster around one component or one kind of change (the structural finding individual postmortems each explained away); cost growth concentrated in one product whose allocation line nobody owns in practice; a provider whose QBR completion rate and performance trend have both been drifting for three quarters (a relationship verdict, quietly already rendered); an estate diff full of oddities accumulating around one legacy decision. For each pattern, force the honest classification: plan wrong, execution wrong, world changed, or acceptable noise — because the classification decides whether the response is a better model, a process fix, an architecture question, or nothing. The architecture question itself gets asked explicitly, once, with this year’s evidence on the table: re-run the scorecard (in whichever direction — estates outgrow single-CDN and outgrow multi-CDN both), and let the answer be allowed to differ from last year’s.
Decisions: the only output that counts
The session ends by converting findings into a short decision list — and short is a discipline: five to ten items, each with an owner, a date, and the finding it answers. The natural families: commercial moves (renegotiate X with the leverage exhibit five documented, resize the commit per the maths against exhibit one’s real distribution, start shadow-testing a candidate before the renewal clock runs); architectural moves (the scorecard’s verdict, an origin change per the cost exhibit, a migration if a verdict came in hard); reliability investments (the funded work behind any SLO tightening, the runbook or drill gaps the incident themes exposed); and organizational ones (the training plan for the bus-factor finding, ownership fixes for the unowned cost line). Anything that didn’t earn an owner and a date was an observation, not a decision — move it to a parking list and keep the real list credible.
Closing the loop into next year
Three closing motions turn the review into a cycle rather than an event. Reset the instruments: next year’s forecast, capacity plan, SLO targets and commit positions, each updated from this year’s calibration data — the review is exactly where last year’s error becomes next year’s accuracy. Schedule the machinery: the four audits, the QBRs, the drills, the notice-date calendar entries, and next year’s review itself, booked now, a quarter ahead of the big renewal, while everyone still means it. And write the one-paragraph narrative — what kind of year delivery had, in five sentences a future team member could read — appended to the estate one-pager where institutional memory actually survives. Estates that run this loop for a few years develop a quality no individual practice produces: decisions reference evidence, evidence accumulates by habit, and the whole delivery function — performance, security, media, architecture, money, people — audits and improves itself on a calendar. Which was, from guide one to guide one hundred, the point.