This story is made up, though anyone who has run a regional travel desk will recognise its shape. A small team in Patna keeps an enquiry-and-booking page alive for visitors planning trips around Bihar. Every year the traffic spikes before the big festival weeks, and every year somebody ends up apologising on the phone. This time the team tried something different: they borrowed site reliability engineering, usually shortened to SRE, and used it to decide when to change the page and when to leave it alone.
The problem was not the outage, it was the argument
Nobody on the desk disagreed that the booking page had to work. The disagreement was about everything else. The editor wanted to publish the new homestay listings the same afternoon. The developer wanted to freeze all changes after a slow Tuesday. The owner asked a question nobody could answer: how unreliable was the page, really? Each person had a memory of a bad day, and memories made poor evidence.
The team lead had read a summary posted by the Atlas on the SRE discipline, which frames the work as engineering applied to running services, with targets drawn from what users experience deciding the tempo of releases. That framing was exactly what the desk lacked: a shared number that ended the argument before it began.
Step one: define what a visitor needs
The team wrote down the one journey that mattered, a traveller opening the enquiry form, filling it in and receiving a confirmation screen. They then counted two things from the hosting panel: how many form submissions were valid, and how many of those completed. A good result was a confirmation within a few seconds. Anything slower or failed counted against them.
They resisted the temptation to track a dozen dashboards. The Atlas is clear that raw server graphs are a starting point, not a measure of whether people are happy. A page can show a healthy processor while the confirmation screen quietly fails for everyone on a mobile connection in a district town.
Step two: pick a target below perfection
The owner’s first instinct was “it should never go down.” The team explained why that goal collapses under scrutiny. Perfect availability cannot be bought, the Atlas warns that squeezing out one more notch of uptime tends to get steeply pricier each time, and the traveller’s own phone network is rarely flawless anyway. They agreed on 99.5 percent of valid submissions completing within the agreed time, measured over a rolling four weeks.
That number brought something useful with it. A target of 99.5 percent means half a percent of requests are allowed to go wrong. Over the previous month the page had handled roughly 8,000 valid submissions, so the allowance was about forty failures. The team called it the allowance, then learned the proper name: the error budget.
Step three: spend the budget on purpose
Once the budget existed, the arguments changed character. In the second week, a plugin update broke the date picker for about an hour and consumed a noticeable slice of the allowance. Under the old regime this would have started a week of finger-pointing. Now the lead opened the sheet, saw that a bit more than a third was gone, and said so out loud.
The team had earlier agreed, in writing, on what would happen at certain thresholds. If a single incident used more than a fifth of the budget, there would be a short review with one concrete follow-up task. If the whole four-week budget ran out, only security fixes and urgent repairs would be published until it recovered. The editor could still add stories on the separate news section, because those never touched the booking path.
- Under a fifth consumed by one event: carry on, note it in the log.
- Over a fifth by one event: hold a review and assign an owner to a follow-up task.
- Budget exhausted: pause everything except urgent repairs until it recovers.
The festival week test
Three days before the festival, traffic doubled. The team watched a simple signal rather than a wall of charts: how quickly the budget was being used compared with the speed that would use up the allowance just as the four weeks finish. The Atlas calls this a burn rate. At a rate of one, the allowance lasts the whole window; at six, it vanishes in a small fraction of it.
On the second evening the rate climbed past six for a short stretch, because an image gallery from a new article was hogging bandwidth. Because the team had agreed beforehand that a fast burn deserved an immediate message to whoever was on call, the developer was alerted within minutes, compressed the gallery and moved on. A slower leak, such as one broken link on a rarely used form, would have produced a ticket for the next working day instead of a night-time phone call.
This separation did more for morale than any tool. People stopped being woken for small annoyances, and they stopped ignoring the alerts that mattered.
Cutting the repetitive chores
The desk also borrowed the SRE idea of toil: manual, repetitive work that adds nothing lasting. Each week someone copied booking enquiries from the inbox into a spreadsheet by hand. The team counted it honestly: about seven hours between two people. The Atlas suggests keeping such work under half of anyone’s time, and the lead decided that a small automation was justified once the weekly total passed a full working day across the team. Within a month the copy-paste chore was gone, and the hours moved to improving the listings that travellers actually read, such as a piece on village hospitality and cultural homestays across Bihar.
What changed in the way they talked
By the end of the quarter, three habits had stuck. First, every release conversation started with the budget rather than with opinions. Second, every incident ended with an action that had a name and a date beside it; the Atlas treats the follow-through as the whole point of a postmortem, and the desk adopted a blameless style that was still strict about owners. Third, the owner stopped asking for perfection and began asking how much of the allowance remained, which is a far better question.
The desk applied the same thinking to its practical content. When writing guidance for visitors who hire a vehicle, the team used the same checklist mindset, and the pieces about the paperwork needed to collect a rental vehicle and about fuel planning on long remote routes follow the same principle: decide what failure looks like before the journey, then watch the few numbers that tell you early.
Lessons worth taking to your own desk
You do not need a large engineering department to try this. A travel blog with a booking form, a ticketing page for a local event or a small guide service can all start with the same three moves: describe the visitor’s journey, set a target slightly below perfect, and agree in advance what happens when the allowance runs low. The Bihar desk in this story is made up, yet the method is real, and the numbers it relied on came from a spreadsheet and an afternoon of honest counting.
The strongest result was not a statistic. It was that a Tuesday slowdown, which once cost a week of resentment, now cost twenty minutes and a line in a shared log. That is what the discipline is for: turning reliability from a feeling into a decision the whole team can make together.