Emerging technology gets a lot of attention for what it can do and very little for who watches it while it runs. Research clouds, lab networks and early production services all generate logs, and someone has to read those logs for signs of trouble. The profession that does so has a name, SOC Ops, and a surprisingly mathematical core. This explainer sets that core out in plain language, for readers who enjoy a good model more than a product tour.
The one-paragraph version
A security operations centre receives alerts, decides what each one means and records the decision. SOC Ops is the craft of doing that dependably, shift after shift. The Ops Atlas, in its explanatory entry on SOC Ops, puts the emphasis on turning alerts into timely, recorded decisions, using tiers, written procedures, search platforms, shift handovers and measurement. It calls the label informal; the work is well established.
Think of it as a queue
The cleanest way to see the problem is as a waiting line, the same kind studied in telephone exchanges and supermarkets. Three quantities set how long an item waits: the rate at which alerts show up, the minutes each one consumes, and the headcount on duty. If arrivals outpace handling capacity, the line grows without limit, and no amount of enthusiasm repairs it.
The Atlas applies a standard queueing calculation, known as Erlang C, to a worked example. In that illustration, weekly triage demand of 386 seat-hours works out to about 11.6 full-time triage staff. Separately, covering a seat around the clock needs about 5.03 people, which agrees with MITRE’s rough figure of five. The numbers belong to an illustration, not to your organisation, but the method is portable: measure arrival rate and handling time, choose a target wait, and solve for staffing.
Why searching logs is its own problem
The raw material is event data, and analysts need to search it in seconds. A platform that collects and indexes logs is the usual answer, and readers curious about how search speed might change in future will enjoy our explanation of quantum speed-ups for database searching. For today’s teams, the practical limit is rarely raw speed. It is the quality of what gets collected. Drawing on CIS Controls v8.1, the Atlas wants logs gathered in one place, held for a quarter or longer and reviewed on a weekly rhythm.
Why the maths of staffing resembles an optimisation puzzle
Rotas are optimisation problems in disguise: cover every hour, respect rest rules, keep at least two people on at any moment, and avoid burning people out. It is no accident that scheduling is a classic example in discussions of computing methods, and the quantum-algorithms piece about fixing hard optimisation tasks covers problems from that same family. Whatever tools eventually help, the human constraints stay: tired analysts make weaker decisions, which is why the Atlas lists burnout among its core topics.
The four-way decision
For each alert the team writes down how it ended, and the Atlas lists four endings: a genuine threat, a genuine but harmless event, a wrongly raised alarm, or a verdict not yet reachable. Think of them as a measurement outcome with an extra category for “we could not tell”. The benign category deserves emphasis: it covers events that really happened and were legitimate, and a study cited by the Atlas found these were the majority of what analysts called false positives. In measurement language, the detector was not wrong about the event, only about whether it mattered.
That framing also explains a subtle point about reporting. Dashboards often display a “false-positive rate”, but a true rate would need a count of harmless events that never triggered an alert, and a monitoring team cannot see those. What the dashboard really shows is the proportion of alerts that were not real, the false discovery rate, equal to one minus precision. Naming it correctly stops teams chasing the wrong target.
Clocks that mean something
Two timers matter. One measures how long until a person first acts on an alert, and the other measures how long until a decision or hand-off. Automation can tag an alert instantly, but that is not a human looking, so it should not stop the first clock. For summaries, the Atlas prefers the median and the ninetieth percentile per shift and severity. One stale item can wreck an average, which the source illustrates with MITRE’s remark about a two-day-old email.
Protecting the observers
A monitoring centre holds a map of everything an organisation watches, so it is a target. The reference architecture in the Atlas keeps the team’s tools in a protected enclave, and it cites UK guidance warning that if monitoring data escapes, intruders can cover their tracks more easily. For readers in emerging fields, where lab infrastructure is often assembled quickly, that is a useful reminder to secure the watchers as carefully as the systems.
What to remember
- Treat monitoring as a queue: arrivals, handling time and staffing decide waiting.
- Record a four-way disposition for every alert.
- Run two clocks and report medians with a high percentile.
- Keep at least two analysts on shift and plan roughly five people per always-on seat.
- Protect the monitoring tools themselves.
The takeaway is that SOC Ops rewards the same virtues as good experimental work: define the measurement, keep honest records and distrust a single number. Learn the queue model first and the rest of the discipline reads naturally. If a figure in a vendor slide cannot be traced to an arrival rate, a handling time or a staffing level, treat it with the caution you would give an unreplicated result.