Anyone who has timed a Marina Beach sunset around a sudden shower, or caught the right Metro line with two minutes to spare, knows that Chennai rewards people who check the signals before they leave the house. The same habit keeps software teams out of trouble, and it is close to what the Ops Atlas describes as ObservabilityOps: looking after the instruments that tell you how your systems are doing. Here is a friendly list of nine habits, each paired with a street-level picture, so you can borrow what works.
If you would like the source definition first, the Atlas has its own page about the discipline. The entry itself concedes that the name is the Atlas’s coinage, so think of these nine items as a checklist borrowed from several real engineering teams rather than a certified standard.
A note on how to read the list. The habits are ordered from cheapest to most involved, so a small team can stop after the fourth and still be far ahead of where it started. A larger group can treat the rest as a roadmap. None of them needs a particular product, and every one can be tried on paper first. Chennai itself works like this: the basics of transport, water and power come first, the glamorous extras later.
1. Know who holds the keys
Every good neighbourhood has the person who knows which shop has the spare key and which tap leaks. Your telemetry system needs a person like that, and probably a deputy. The Atlas treats a named owner as the first real step up from chaos. Without one, nobody retires old dashboards, nobody questions the invoice and nobody notices when a rule has been broken for weeks.
Try this: put the owner’s name and a backup at the top of the inventory sheet, and review the sheet on the first Monday of every month.
2. Keep a list of everything that reports in
You cannot find a bus stop you have never heard of. Draw up an inventory: which services send data, to which stores, under what names. The Atlas suggests watching for data arriving with no service name at all; the share of such “unknown” sources should hover near zero. If it does not, somebody has deployed something that nobody can trace back to a team.
Try this: for each service on the list, record which team gets the call when its data goes missing. Blank cells show you where the map has holes.
3. Fit a heartbeat, like the all-clear siren
After a storm warning, the quiet that follows should mean safety, not a dead loudspeaker. The Atlas describes a heartbeat alert that fires all the time by design, with a separate watcher that raises a flag when the beat stops. The same source stresses testing the entire path with one external check, from raised alarm to the message reaching a human, instead of trusting each part separately.
Try this: break the alert path on purpose in a quiet hour, with the team’s knowledge, and note how long anyone takes to spot it. The answer is usually humbling and always instructive.
4. Treat labels like platform numbers, not passenger names
Metro boards show line and direction, not the name of every passenger. When software attaches identifiers that differ for every event to a measurement, the number of stored series explodes. The Atlas rule of thumb keeps variety on a metric label very low, urges a closer look past a hundred, and says loudly that identifiers like user numbers do not belong there. A trap it highlights in modern setups is a per-instance identifier that changes every time a collector restarts; multiplied across replicas it churns out series forever. The suggested remedy is to aggregate that label away in queries, not to drop labels blindly or lean on a blunt cap.
Try this: list every label on your busiest measurement and write beside each one how many different values it can take. Circle anything open-ended and plan its removal.
5. Make every service speak the same language
Tamil, English and a little Telugu on the same bus is charming; in a log file it is a headache. Agree on field names, a common request identifier, and structured log lines that carry severity and trace identifiers in dedicated fields. The Atlas recommends open, vendor-neutral conventions for this, and notes that the main standard for passing trace context between services is a W3C recommendation which is the default in the OpenTelemetry toolkit.
Try this: write one page, no longer, showing a sample log line with every required field filled in, and pin it where newcomers will find it on day one.
6. Mind the handoffs
The messy moment in any journey is the change of vehicle. Traces, which follow one request across services, break at the same places: message queues, proxies that drop the header, and mixed formats. A related warning in the Atlas is about context riding along with requests. If you attach extra details, they may travel to outside services, so allow only listed keys, strip them on the way out and let a brand-new trace start wherever traffic enters from the open internet.
Try this: pick one request that crosses at least three services and follow it by hand. The place where the trail goes cold is your first fix.
7. Decide how long to remember
Nobody keeps every bus ticket for a decade. Telemetry needs the same honesty. Keep recent detail close, move older data to cheaper storage, and archive what you rarely touch. The Atlas points out that one popular open-source metrics store keeps about two weeks by default, and it gives a simple formula for estimating disk that multiplies retention time, ingest rate and the byte size per sample. Run the sum before you promise a client a year of history.
Try this: ask each team how far back they actually looked last quarter. Many discover the answer is a few days, which makes long, costly retention hard to justify.
8. Review alerts the way you review a timetable
Timetables get revised when routes change. Alerts should be revised too. Page people for symptoms a customer would notice, drop signals nobody reads, and measure how many pages led to action. A number worth watching, according to the Atlas, is the share of alerts that were genuinely actionable; in its worked example only fifteen of forty-two qualified.
Try this: export last quarter’s pages, mark each as acted on or ignored, and retire the ignored ones before the next review.
9. Count the cost per unit
Auto fares make sense because they are metered. Give your telemetry a meter as well. Divide monthly spend by the volume served, and report the result per million requests. A falling number alongside steady coverage means you are getting better at this. If the total rises and nobody can say why, return to habits two and four.
Try this: write the unit cost on the front page of the platform’s own dashboard, where everyone can watch it move.
A bonus habit: Read the claims sceptically
Chennai travellers learn to treat a promise of “five minutes away” with affection and a pinch of salt. Treat vendor claims about intelligent observability or magical data pipelines the same way. The Atlas says to regard them as marketing until the product has been tried on your own data.
Try this: choose one claim from a sales page and test it against a known problem in your own data before any contract is signed.
Where these habits meet a real trip
If you are planning to visit, start with how first-time guests can plan a trip to Chennai, then look at a sensible number of days to spend in the city. Both rely on the same skill: collecting a few reliable signals and ignoring the rest.
Practice the nine habits in order, and you will find the first three or four cost almost nothing yet remove most of the pain. The Atlas outline puts owner, inventory, heartbeat, an outside probe and the three worst labels inside the first thirty days. Start this week, and the next time the monsoon of alerts rolls in, you will already know where the dry ground is.