When Every Report Sounds Urgent: AIOps Diagnosis for Travel Desks

When Every Report Sounds Urgent: AIOps Diagnosis for Travel Desks
Technology

Picture the information desk at a busy railway junction in Bihar during Chhath week. By noon the staff have heard the same complaint forty times: the approach road is jammed. Each caller sounds urgent, each message arrives in a different wording, and the one report that matters, a washed-out culvert on a different road entirely, sits somewhere in the pile. Nobody on that desk is lazy. The desk is simply drowning in repeats. This is exactly the condition that AIOps was invented to treat, and it is a surprisingly good lens for anyone who runs a travel or local-news operation in the region.

Put simply, the field means that software helps a team cope with the flood of alerts, logs and signals that modern systems produce. If you publish a travel portal, run bookings for homestays, or keep a news site alive on festival days, you already live with that flood. This article diagnoses why the flood becomes a problem and what a calm, step-by-step cure looks like, without assuming you have a data science team.

The symptom: forty alarms, one real fault

Let us name the symptoms plainly. First, repeats: one broken thing raises many alarms. Second, ambiguity: the alarm says something is wrong but not whose job it is. Third, collisions: two different problems get lumped together because they happened close in time. Fourth, silence where you need noise: a rule written months ago quietly stops matching anything.

The Ops Atlas, a reference library run by xOps School, describes the AIOps entry in its catalogue as a bundle of unglamorous disciplines first and machine learning second. I found that framing useful because it matches how a good desk supervisor already thinks. You do not buy a clever phone system before you have agreed who answers the Gaya line and who answers the Muzaffarpur line.

Diagnosis step one: agree on what a message must carry

In the Atlas ordering, the opening four practices, sorted by payoff, involve no learning algorithm whatsoever. Number one is a labelling contract: each alert names its service, its owning team, a severity, and the page holding the instructions. Translate that to a travel desk. Every incoming report should carry a route, a district, a severity and a link to the standard reply. A caller who says “road bad” is a message without labels, and no software on earth can sort it reliably.

Think about how many documents travellers are asked to show before collecting a rental scooter. A local guide on this very site, the explainer about paperwork needed at vehicle pickup counters, shows why a fixed checklist saves everyone time. The counter does not improvise per customer; it asks for the same five things. A label contract is the same idea applied to alarms. According to the Atlas, a first maturity level is considered finished when 95 percent of alerts comply with the contract, which is a refreshingly concrete bar.

Diagnosis step two: collapse the repeats

Once messages carry labels, the second practice is deduplication: give each distinct problem a fingerprint and fold repeats into it. The Atlas warns about a subtle trap here. If a label changes every time something restarts, such as the name of a single running container, the fingerprint changes too, and the fold fails. Back at our junction desk, imagine fingerprinting a complaint by the caller’s name. Forty callers, forty fingerprints, no folding. Fingerprint it by road and direction instead and forty complaints become one ticket.

That is why the first thing to inspect in your own tooling is which fields are stable and which are volatile. Stable fields describe the thing that is broken. Volatile fields describe the circumstances of one observation. Build your sorting on the stable ones.

Diagnosis step three: group, but do not over-merge

Grouping takes folding one step further: related tickets are shown together so one person looks at them once. Here the Atlas gives a sobering lab result. Merely bundling alerts that landed inside one five-minute window compressed the pile 5.5-fold, a result that sounds splendid until you notice it also fused unrelated incidents. In desk terms, a flooded culvert and a jammed bridge happened to be reported in the same five minutes, and somebody filed them as one. The count looked great and the culvert stayed unfixed.

The Atlas treats shrinkage as a vanity figure, because a bad merge hurts roughly the way a missed alarm does. So measure your grouping the way a careful editor measures a correction log: of the groups you made, how many were right, and of the real problems, how many did you catch. Those two questions are precision and recall, and the Atlas suggests judging any correlation approach against incidents that humans have already labelled.

Diagnosis step four: use what changed

The fourth practice looks at topology and change. Topology is simply the map of what depends on what. Change is the log of what someone altered recently. Here the Atlas borrows a figure from Google’s SRE book: around 70% of outages, it says, trace back to someone altering a running system. Any travel operator will recognise the pattern: the booking page was fine on Monday, someone updated the payment plug-in on Tuesday, and Wednesday the festival rush revealed the damage.

So when a flood of alerts arrives, the cheapest useful question is “what changed in the last few hours, and what depends on it?” Two plain spreadsheets, one listing dependencies and one logging changes, are what the Atlas names among the prerequisites for the whole discipline. When the lab on that page deleted a lone dependency edge, recall slid from 0.93 to 0.53. Missing maps cost you real findings.

A short worked picture

Suppose your local-news site slows down during a state announcement. Alarms fire from the web servers, the database, the image cache and the comment system. A labelled, fingerprinted, topology-aware desk would fold these into one story: the database is the common dependency, a configuration change landed forty minutes ago, and everything else is a symptom. The editor on duty sees one ticket, with the suspected cause ranked first and the change log attached. Without those steps the same editor sees four screens of red.

Where learning algorithms earn a place, and where they do not

Learning-based anomaly detection enters the Atlas ordering only afterwards, for corners that hand-written rules cannot cover. Its lab run used a seasonal baseline that spotted every one of four planted faults, yet raised 34 false alarms, 32 of them on an undeclared public holiday. Adding one calendar entry silenced all 32. For a Bihar publisher the moral is almost comic: tell the system when Chhath, Holi or an election day falls, and half its apparent intelligence stops being needed.

The same page cautions against expecting magic. It points to Google SRE guidance on steering clear of systems that learn their own thresholds or claim to find causes automatically, and it describes a separate myth that this field can replace good alert rules. Treat any tool that promises certainty with the suspicion you would give a tout promising a confirmed ticket on a sold-out train.

A short course of treatment for this week

  1. Write down the five fields every alert or report must carry, then audit a week of real traffic against that list.
  2. Pick the volatile fields that break your folding and remove them from the fingerprint.
  3. Hold a short weekly noise review: which messages woke someone up and led to nothing?
  4. Keep a plain change log and a dependency list where the duty person can see them.
  5. Add a holiday and festival calendar so predictable surges stop looking like faults.

The Atlas proposes a cadence of weekly, monthly and quarterly reviews, and the rhythm suits a small team since every meeting is brief and asks just one question.

Guarding the cure

One last diagnosis deserves a mention. Automation that acts on its own, such as restarting something or silencing a rule, needs fences. The Atlas sketches fixes restricted to an approved list, reversible, and capped at a handful of runs per day. Think of a station announcer who may repeat a platform change twice but may never reroute a train. Give your automation that same narrow remit, and keep a person on call for everything else.

Bringing it home

The junction desk will always be busy in Chhath week; roads will always flood. What changes is whether the staff meet that busyness with a sorting system or with raw nerves. Should a single idea survive from this diagnosis, let it be the sequence: labels, folding, grouping, change awareness, and only then learning. Travellers planning an overland trip can see the same principle in our walkthrough of choosing between scooter, motorbike and rental car in Colombia: decide the criteria before you look at the vehicles.