Challenges

Six months after launch, your app has 10,000 users. This morning a popular newsletter featured you, and traffic doubled before lunch.

A creator with two million followers posted about your app last night. You crossed a million users before breakfast, and almost every one of them is opening the same few trending pages.

You launched photo posts on Friday. By Sunday, eight out of every ten requests are for pictures: profile photos, food shots, sunsets.

A national holiday, and everyone is home scrolling. Feed loads are four times a normal Sunday, and nearly half of all requests are for photos.

Your app's search box now suggests as you type. Every keystroke is a request, and at lunchtime that's 5,000 a second.

The season finale of the country's favourite show drops at 9 PM. For a few minutes, viewers arrive three times faster than any normal evening.

A festival sale, and every shopper keeps opening their cart: adding, checking, comparing. Ninety-five in a hundred requests just read a cart.

Tickets for the biggest concert of the year go on sale in your app at noon. At 12:00:01, scalper bots arrive: thousands of scripted requests a second, on top of every real fan.

Midnight sale. When the clock strikes twelve, orders arrive three times faster than usual.

Your chat app launches in five cities tonight. Half of all requests are new messages, and every one is saved to the message store before it's delivered. The store can save about 750 messages a second, and the chat servers wait for each save before answering, so at launch traffic everything backs up.

Marketing scheduled a 2-crore offer blast for 8 PM, the same minute dinner orders peak. At the spike, alert requests arrive three times faster than usual.

Six o'clock on a weekday. Every office is emptying, every driver is online, and half of all requests are drivers' location pings, saved straight to the trips database. The other half are riders asking for nearby drivers, read from the same database.

Your railway booking app opens Tatkal at 10 AM sharp. For a few minutes, requests arrive three times faster than the busiest normal moment, and two booking servers can't take it: everyone's requests queue, slow down, and time out together.

Your new search feature needs the whole recipe web indexed by Monday. The URL frontier hands out 600 pages a second to fetch and save.

At 3:07 a.m. your phone buzzes. Sam, your SRE, is already on the call: "App server 1 just died. Hardware fault. The whole site is down."

App server 2 has a bad memory chip, and the replacement only arrives on Monday. Until then it crashes every so often and comes back a few seconds later.

Tonight at 9 PM your cloud provider will move the database to new hardware. They promise it'll be unreachable for "about twenty seconds". Twenty seconds at your traffic is thousands of failed requests.

It's the first of the month: salaries land, bills get paid, and your payment app's traffic doubles. Most of it is people tapping "check status" on payments that are still pending, again and again, and every one of those taps reads the payments database.

Your cloud provider has warned that its Mumbai region will have a network problem tonight: for about twenty seconds, nothing in it will be reachable. Today your whole app lives there.

Your app's new festive feature turns a selfie into an AI-made Diwali card. It went live an hour ago, and about four people a second are asking for one, in bursts every time an influencer posts about it.

The AI Diwali cards are a hit. People are sharing them on WhatsApp and Instagram, and every share opens the card's picture from your servers: 3,000 requests a second, and 85 in 100 are just someone looking at a picture.

Your electronics store's AI shopping assistant is busy on sale day: 28 questions a second. Most of them are the same few: "What's the return policy?", "Is there no-cost EMI?", "When will my order arrive?"

Monday, 9:30 AM. Your company's new AI assistant answers questions from 40,000 employees using the HR, IT and travel policies, and the launch email has just gone out.

Your short-video app is having its biggest night: 4,000 requests a second. Most are people watching, and 8 in 10 requests are for videos and thumbnails that never change once posted. About 1 in 25 is someone posting.

Your travel app's AI trip planner sends every question to one model provider through your AI gateway. About 12 questions a second, all day.

Your coding assistant answers about 25 questions a second from developers. On Friday evening someone posts a "how to make it ignore its rules" thread, and within minutes a third of all messages are prompt-injection attempts.

Your app's new AI assistant launches on a Monday morning, and the press coverage brings a rush of people trying it at once.

Your AI assistant now answers questions for a whole bank's customers and staff, about a crore a fortnight, with 600 a second at the morning peak. Each question is searched against the bank's policy index, and the ten closest passages go to the model with it.

Your support assistant answers 40 questions a second, every one on the large model. Answers are excellent, and so is the bill: about $6.50 per 1,000 questions.

Your travel app's planner takes requests like "Three days in Jaipur in March, under ₹20,000, with a day trip". About 10 a second, and most need several steps: check dates, search trains and hotels, fit the budget, then write the plan.