{"id":1172,"date":"2026-11-16T09:40:00","date_gmt":"2026-11-16T09:40:00","guid":{"rendered":"https:\/\/flyaiaa.com\/blog\/?p=1172"},"modified":"2026-11-18T18:40:00","modified_gmt":"2026-11-18T18:40:00","slug":"observabilityops-three-operating-models","status":"publish","type":"post","link":"https:\/\/flyaiaa.com\/blog\/observabilityops-three-operating-models\/","title":{"rendered":"Three Ways to Run ObservabilityOps, Compared Side by Side"},"content":{"rendered":"<p>When a flight department compares aircraft, nobody asks which one is simply better. They ask which one fits the mission profile, the crew, the budget and the maintenance support available. The same question applies to how an organisation looks after its monitoring system, the discipline the Ops Atlas calls ObservabilityOps. The Atlas records three very different real-world setups, and setting them side by side is more instructive than any definition. This comparison draws on those cases and on the entry&#8217;s own decision rules, with the analogy of charter fleets in mind. It describes operating practices only and takes no view on any product.<\/p>\n<h2>The three setups in brief<\/h2>\n<p>The entry gives one example each of a large self-built platform, a tiny team using a hosted back end and a public-sector shared service. They differ in size by orders of magnitude, which is what makes the comparison useful.<\/p>\n<ul>\n<li><strong>Case A: a very large internal estate.<\/strong> A web-infrastructure company, writing in March 2023, described a fleet of 916 Prometheus instances that together held close to 4.9 billion series. Its engineers changed Prometheus so every instance would limit how many series it took in, rather than discarding an entire scrape when a target misbehaved. The promise to internal teams was simple: if your CI checks pass, the capacity you asked for exists.<\/li>\n<li><strong>Case B: one engineer, high traffic.<\/strong> A federated social-network organisation of about twenty staff serves around ten million requests each minute. A single engineer is responsible for observability, with one Collector in each namespace, rolled out via an operator plus a GitOps controller, a commercial hosted back end, and a tail-sampling policy that stores every error alongside roughly 0.1 percent of successful traces. The sheet records no Collector trouble across two years or so.<\/li>\n<li><strong>Case C: a shared government platform.<\/strong> A British justice ministry runs an Observability Platform on top of its Modernisation Platform, offering logs, metrics and alerting, with notifications sent to on-call paging and chat tools. The Atlas notes the service is due to be retired, although no date has been given.<\/li>\n<\/ul>\n<h2>Comparison one: who carries the state<\/h2>\n<p>The most important difference is where complexity lives. Case A absorbs it internally, patching its own software, and so it needs a team able to do that. Case B hands the heaviest part, the stateful back end, to a supplier and keeps only the lightweight pieces, which is why one person can cope. Case C centralises the work for many departments, trading autonomy for economy.<\/p>\n<p>The Atlas draws this line in its team-shape heuristics. Where stateful back ends run in-house, it leans toward a platform team; where the back end is bought, a smaller standards function usually suffices. For a company with one or two teams, assigning a named person on call is worth it chiefly when you host the stores yourself or answer to a regulator. Fleet planners will recognise the logic: owning aircraft means owning the maintenance organisation, whereas chartering means paying someone else to hold that burden. If you are weighing the cost of chartering rather than owning, our list of <a href=\"https:\/\/flyaiaa.com\/blog\/hidden-costs-to-check-before-booking-charter-flights\/\">hidden costs worth checking before a charter booking<\/a> applies a similar discipline to trips.<\/p>\n<h2>Comparison two: how each case handles overload<\/h2>\n<p>Every setup eventually receives more data than it planned for. The three answers differ. Case A&#8217;s answer is a hard ceiling, enforced in software and checked before release, so that one team&#8217;s mistake cannot take away everyone&#8217;s visibility. The Atlas explains the motive: with a stock Prometheus setup, going over a configured sample or label cap discards that target&#8217;s entire scrape, not just the excess. Case B&#8217;s answer is selection, with tail sampling keeping the failures and slow requests while discarding most of the routine successes. Case C&#8217;s answer, as recorded, is a standard offering with defined integrations, though the entry does not describe its limits.<\/p>\n<p>Sampling choices carry trade-offs that are easy to miss. Deciding at the start of a trace is cheap but cannot guarantee that error traces survive. Deciding at the end preserves the interesting traces at the price of routing every span of one trace to one Collector instance, which the guidance describes as difficult to operate. Case B proves it can be done, and the entry&#8217;s wording leaves no doubt that it takes care.<\/p>\n<h2>Comparison three: how they prove it works<\/h2>\n<p>Evidence of health also differs. Case A offers a promise that is checkable: CI passes, so capacity exists. That is a contract between platform and users. Case B has the discipline of a single owner and, apparently, the benefit of simplicity. Case C documents integrations but not outcomes.<\/p>\n<p>The Atlas suggests a few measurements that any of the three could adopt. One is the success rate of exported telemetry, illustrated with a calculation of 9,950,000 delivered against roughly ten million attempted, giving 99.5 percent, below a 99.9 percent target. Another is freshness, measured by a synthetic canary at the 95th percentile. A third is the heartbeat of the alerting path. A group booking planner tracks something similar when confirming each leg before the party travels, as in our <a href=\"https:\/\/flyaiaa.com\/blog\/group-charter-flight-planning-checklist-for-organizers\/\">checklist for organisers of group charters<\/a>.<\/p>\n<h2>Comparison four: cost and unit economics<\/h2>\n<p>The entry&#8217;s survey context shows why cost matters in all three. The entry cites a 2025 vendor survey of 1,255 respondents in which the average observability bill came to 17 percent of compute spending while the middle respondent paid 10 percent. A mean well above the median hints at a group of heavy spenders pulling the average up. Nothing in the entry says which of our three cases sits where, and it would be wrong to guess.<\/p>\n<p>What the Atlas does provide is a method: convert the bill to a unit price, such as spend per million requests, and watch whether it falls while coverage holds. A worked example computes 20 per million from 18,000 a month and 900 million requests. In a comparison like this one, unit price is the fairest yardstick, because it adjusts for size. Case B probably has the smallest absolute bill and Case A the largest, but their unit prices could be anywhere.<\/p>\n<h2>Comparison five: what happens when the person leaves<\/h2>\n<p>Continuity is the least glamorous axis and often the decisive one. Case B concentrates knowledge in one engineer, which is efficient until that engineer takes leave. The entry does not say how the organisation covers this, so the sensible reading is that it has to be solved by documentation, templates and configuration kept in version control, the kind of habits the Atlas associates with its third maturity level. Case A has depth but pays for it with coordination effort. Case C depends on the sponsoring department staying committed, which the retirement notice shows is not guaranteed.<\/p>\n<p>Fleet planners ask the same thing about maintenance providers: if the lead technician is away, can another one pick up the work card? Any model you adopt should have an answer. Even a single-owner setup can be made resilient by writing the runbook as if a stranger will use it next quarter.<\/p>\n<h2>A compact scorecard<\/h2>\n<ul>\n<li><strong>Staffing need:<\/strong> A high, B very low, C moderate.<\/li>\n<li><strong>Control over internals:<\/strong> A high, B low, C shared.<\/li>\n<li><strong>Exposure to vendor decisions:<\/strong> A low, B high, C tied to organisational policy.<\/li>\n<li><strong>Resilience of the model itself:<\/strong> C carries a stated retirement plan, which is a reminder that organisational risk matters too.<\/li>\n<\/ul>\n<p>This table is an interpretation, not a finding of the Atlas, and it is deliberately coarse. Use it as a prompt for discussion rather than a verdict.<\/p>\n<h2>What the comparison suggests<\/h2>\n<p>None of the three setups is a template to copy. Each solves a particular problem under particular constraints. The shared lesson is that someone, somewhere, owns the observability system itself, whether as a team, an engineer or a service. The anti-patterns the Atlas names, &#8220;collect everything and decide later&#8221;, &#8220;self-hosting is free&#8221; and a central team owning everyone else&#8217;s alerts, all describe situations where that ownership is missing or misplaced.<\/p>\n<p>For decision-makers, the practical test is simple. Choose the model whose weaknesses you can afford. If you cannot staff a platform team, avoid designs that require one. If you cannot tolerate dependence on a supplier, budget for the people to run your own stores. For the underlying text, including the maturity levels, see the <a href=\"https:\/\/www.xopsschool.com\/observabilityops\/\">Ops Atlas entry comparing practices around ObservabilityOps<\/a>. A final note from the planning desk: whichever model you choose, write it down, name the owner and review it once a quarter, the way a sensible flight department reviews its fleet plan.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>A fleet-planner&#8217;s comparison of three documented ways to operate a monitoring platform: a huge self-built estate, one engineer, and a shared public service.<\/p>\n","protected":false},"author":6,"featured_media":1173,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[634],"tags":[],"class_list":["post-1172","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-flight-training-tech"],"_links":{"self":[{"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/posts\/1172","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/users\/6"}],"replies":[{"embeddable":true,"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/comments?post=1172"}],"version-history":[{"count":1,"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/posts\/1172\/revisions"}],"predecessor-version":[{"id":1294,"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/posts\/1172\/revisions\/1294"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/media\/1173"}],"wp:attachment":[{"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/media?parent=1172"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/categories?post=1172"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/flyaiaa.com\/blog\/wp-json\/wp\/v2\/tags?post=1172"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}