Upgrade & Secure Your Future with DevOps, SRE, DevSecOps, MLOps!
We spend hours scrolling social media and waste money on things we forget, but won’t spend 30 minutes a day earning certifications that can change our lives.
Master in DevOps, SRE, DevSecOps & MLOps by DevOps School!
Learn from Guru Rajesh Kumar and double your salary in just one year.
It was a quiet Thursday when an on-call engineer, three coffees in, decided that the team’s alert-correlation project had gone far enough without a drawing. The repository already held a few hundred lines of glue code, a half-trained model, and a dashboard nobody had asked for. Nothing was wrong, exactly. But nobody could say where an event actually went after it left the monitoring tool. So she grabbed a marker and a stack of paper, and what follows is the method she used to sketch an AIOps pipeline before writing anything else.
The idea is simple. A pencil sketch exposes bad assumptions in minutes, and at that stage changing your mind costs almost nothing. The same discovery after the build costs a refactor, a missed deadline, or an outage that teaches the lesson the hard way.
Step 1: Draw the boxes you already believe in
Start with a rectangle for every source of signals you can name. Metrics, logs, traces, deployment events, tickets, the change calendar. Do not worry about neatness. The engineer’s first page had seven boxes and four of them had question marks on them, which was already useful, because a question mark is a to-do item that costs nothing yet.
Then draw one arrow from each source toward the thing you hope to build. Label every arrow with how the data travels: pushed, pulled, batched, streamed. If you cannot label an arrow, you have found your first gap. In her case, the change calendar lived in a spreadsheet that one person updated by hand, and no arrow could honestly describe that.
Step 2: Mark where a human enters the picture
Most AIOps sketches fail quietly because people are left off the page. Add a stick figure wherever a person reads, approves, silences or overrides something. Ask who gets paged, who gets the summary, and who can say the model is wrong.
This step turned her drawing from a data diagram into an operating diagram. The correlated incident card, it turned out, would land in a channel that three teams muted by default. No amount of clever modelling fixes a message that nobody reads, and finding that out with a pencil took four minutes.
One habit helps here: give every stick figure a job title rather than a name, and write the exact moment they would need to act. “Night-shift responder, wakes at the first page” tells you far more about the design than a generic user box ever could. If two figures share a single arrow, ask whether they really need the same message, because they rarely do.
Step 3: Trace one real incident through the sketch
Pick a past incident the team remembers well and walk it along your arrows with a different colour of ink. Where does the first signal appear? Which box would have grouped it with its neighbours? Where does the output reach a person, and how long would that have taken?
She chose a disk-latency incident from the previous quarter. Following it through the drawing showed that the key signal was a log line, yet the sketch had no log parsing step at all. On paper, the repair was a new box and an arrow. In code it would have been a schema change in the middle of the pipeline.
If you want a deeper look at one stage she kept returning to, the grouping stage has its own treatment in an earlier piece on quieting a flood of alerts with clever automation, worth a read next.
Step 4: Ask what happens when a box fails
Take a different pen and cross out one box at a time. What does the rest of the pipeline do? Does it stall, fall back, or quietly produce wrong answers? For each crossed-out box, write one sentence on the page about what the downstream person would see.
The most uncomfortable result was the enrichment step. If the inventory lookup went down, correlation would continue but every incident would lose its service owner, and nobody would be told. The fix she scribbled in the margin was a visible banner on each card that read “owner unknown”, a small design change that was free on paper.
Step 5: Decide what the first slice will be
A sketch is not a plan, but it makes a plan possible. Circle the smallest connected path that delivers something a person can use: perhaps one signal source, one grouping rule, and one notification. Everything outside the circle is explicitly later.
She circled three boxes and wrote a date next to them. The model, the dashboard and the ticket integration all stayed outside the circle. Writing the word “later” next to a box is oddly freeing; it turns a vague obligation into an honest deferral.
- Keep the circled path small enough to explain in a minute.
- Make sure it touches a real person at the end, not just a log.
- Write down one thing that would convince you the idea is wrong.
Step 6: Photograph it, then redraw it once
Take a picture of the first sketch and keep it. Then spend ten minutes redrawing from memory on a clean page, simplifying as you go. The second version usually removes a box or two, and the boxes that survive the redraw are the ones that matter.
Hers dropped from nine boxes to six. The three that vanished had been there out of habit, borrowed from a vendor diagram she had seen in a slide deck. Pencil made the difference between copying a picture and understanding one.
A small detour into outside reading
When a sketch session leaves you with questions that are really about the gear around you, such as laptops, phones and the small devices engineers carry into an on-call shift, a reference outside your own field can help: the independent site Osnivo has its own forum and checkers, and a quick look there keeps the design review moving.
What the sketch bought her team
When the markers dried the team held a pipeline they could describe out loud, a list of five risks, and a smaller first release. The engineer pinned the page above her desk. When someone later proposed adding a predictive component, the sketch gave the conversation a place to start, and the group could weigh the idea against the drawing instead of against a hunch. Her earlier thinking on predictive maintenance in IT systems helped frame that discussion, and the new box went onto the page in pencil.
Try it this week
No special software or workshop is required; paper, two pen colours and an incident from your recent past are enough. Draw the sources, add the people, trace the incident, cross out each box, and circle the first slice. If the drawing feels embarrassing, good: it means you are still at the stage where changing it is cheap.
Then do one more thing: show the drawing to a colleague before you show them any code. Their first confused question is worth more than the next three days of implementation, and it arrives for free.