5 rules for EdTech apps that survive exam week in India (2026)

Summary. An EdTech product in India does not have traffic. It has weather. The National Testing Agency held NEET (UG) 2026 on 3 May 2026 and ran a re-examination on 21 June 2026, and every learning app touching those candidates saw the same shape: months of flat usage, then a wall. Two facts decide whether you survive it. The first is that direct database connections do not scale with money: on Supabase, a Micro instance allows 60 direct connections and 200 pooled, and the largest 16XL instance at $3,730 a month allows 500 direct connections and 12,000 pooled. Direct connections plateau at 500 while the pooler goes 24 times further, so a spike is a pooler problem, not a compute problem. The second is that the Digital Personal Data Protection Act, 2023 prohibits tracking, behavioural monitoring and targeted advertising directed at children, defined as anyone under eighteen, and makes parental consent mandatory for them. That deletes most of the standard growth analytics stack from an Indian EdTech app. Here are the five rules that follow.

Rule 1: the connection pool is the wall, not the CPU

The instinct when a spike is coming is to buy a bigger machine. Look at what a bigger machine actually buys you, from the Supabase pricing page.

Compute size	Price per month	Direct connections / pooler connections
Micro (2-core ARM, 1 GB)	$10	60 / 200
Small (2-core ARM, 2 GB)	$15	90 / 400
Medium (2-core ARM, 4 GB)	$60	120 / 600
Large (2-core ARM, 8 GB)	$110	160 / 800
XL (4-core ARM, 16 GB)	$210	240 / 1,000
4XL (16-core ARM, 64 GB)	$960	480 / 3,000
16XL (64-core ARM, 256 GB)	$3,730	500 / 12,000

Read the right-hand column downward. Going from Micro to 16XL multiplies your bill by 373 and your direct connections by 8.3. They stop moving entirely past 4XL: 480, then 490, then 500, then 500. Pooler connections keep climbing to 12,000.

The lesson is not subtle. If your app opens a direct connection per request, no amount of money fixes exam week. You will exhaust 500 connections and start refusing students while the CPU sits idle. Route through the pooler, keep transactions short, and treat a held-open connection as the scarce resource it is.

This is also the cheapest scaling decision available. A pooled Medium instance at $60 a month allows 600 pooler connections, more than the 500 direct connections you get on a 16XL costing $3,730. You can pay 62 times more and still hit the wall sooner.

Rule 2: exam week is a read problem, so cache like it

The traffic on exam day is not balanced. Thousands of students are reading the same syllabus page, the same mock paper, the same result. Almost nobody is writing.

That asymmetry is a gift, and most teams squander it by treating every read as a fresh database query.

The numbers make the case. Supabase Pro includes 250 GB of egress at $25 a month, then charges $0.09 per GB. Cached egress on Pro also includes 250 GB, then $0.03 per GB, which is a third of the uncached rate. Vercel Pro includes 1 TB of Fast Data Transfer and 10,000,000 Edge Requests against its $20 monthly platform fee, with $20 of usage credit on top. On Amazon API Gateway, a 1.6 GB cache runs $0.038 an hour, which is $0.912 a day, or roughly $28 a month, against REST API calls at $3.50 per million.

Put those together for exam week. A syllabus page served from cache costs a third of the same page served from origin, and an API response served from a $28 monthly cache displaces calls billed at $3.50 per million. The static content is the majority of exam-day bytes and it is the part that is trivially cacheable, because a mock paper does not change while ten thousand people read it.

The corollary is that your write path deserves the connections you saved. Submissions, answers and results are the operations that must not fail, and they are a small fraction of the volume.

Rule 3: build offline-first, because Indian exam week is a network event

A student in a coaching hostel at 9pm on results day is on congested mobile data. The app that assumes connectivity is the app that loses the session.

Offline-first is not a feature here, it is the baseline. Content downloads ahead of time. Answers queue locally and sync when the network returns. The UI never blocks on a request that may never complete.

That design has a consequence people miss: every queued write will be retried, sometimes more than once, from a client you do not control. Without an idempotency key, a student who submits an answer on a flaky connection submits it three times, and your database now holds three attempts at one question. We covered the mechanics in idempotency keys and safe retries for REST APIs. For an EdTech app the rule is simple: any client-queued write carries a key generated on the device, and the server deduplicates on it.

Both major cross-platform stacks handle this well, and the choice rarely turns on the sync layer. We have written up the trade-offs in Flutter app development and React Native cross-platform development.

Rule 4: DPDP rewrites your analytics, not just your privacy policy

This is the rule that surprises founders, and it is not a footnote.

The Digital Personal Data Protection Act, 2023 was passed by the Indian Parliament in August 2023. In Anirudh Burman's analysis for the Carnegie Endowment for International Peace, the provisions covering children are unusually direct: consent of the parent or guardian is mandatory in the case of children and minors, defined as those under eighteen years of age; any processing that is likely to have a detrimental effect on a child is not permitted; and the law prohibits tracking, behavioural monitoring, and targeted advertising directed at children. The government can prescribe exemptions from these requirements for specified purposes.

Now hold that against a normal growth stack. Behavioural event tracking to build engagement funnels. Retargeting ads to lapsed users. Cohort analysis on individual usage patterns. Recommendation driven by monitored behaviour. On a product whose users are school students, several of those are not a compliance risk to be managed, they are prohibited as directed at children.

The design responses that work are unglamorous and cheap if you make them early.

Establish age at signup and branch the whole data model on it, rather than bolting on an age gate later. Route under-18 accounts to a parental consent flow that produces a record you could actually show someone. Keep aggregate product analytics, which tell you that a chapter has a high drop-off, and drop per-child behavioural profiling, which tells you what one named student did on Tuesday. Do not run targeted advertising at minors at all, which for most EdTech products means the ad-supported free tier needs a different business model rather than a different ad network.

The wider build-side detail is in our note on DPDP-ready application development. We design applications aligned with DPDP requirements; the compliance determination itself sits with the operator and their counsel.

The engineering judgement here is that age is the highest-use field in an Indian EdTech schema. Get it wrong in week one and every analytics decision afterwards inherits the mistake.

Rule 5: capacity-plan against the calendar, then set a spend cap

Exam dates are published. NTA held NEET (UG) 2026 on 3 May 2026, and when it ran a re-examination on 21 June 2026 the entire load pattern repeated with weeks of notice. This is the rare capacity planning problem where the spike is on a public calendar.

So plan against the calendar rather than a rolling average. Scale compute up before the date and down after it, because Supabase bills compute hourly and you can move between instance sizes as needed. Load-test at the peak you expect, not at last month's mean.

Then protect yourself from your own success. Supabase enables a spend cap by default on the Pro plan, which you must deliberately switch off to allow usage beyond plan limits. Vercel turns on spend management notifications for new customers at $200 per billing cycle. Both defaults exist because usage-based billing on a viral spike is how a good day becomes a bad invoice. The judgement call is which failure you prefer on exam morning: a hard stop at the cap, or an uncapped bill. For a paid product the answer is usually uncapped with alerting; for a free tier it is usually the cap.

Watch the metered lines that move with students rather than with traffic. Supabase Pro includes 100,000 monthly active users and then charges $0.00325 per additional MAU, includes 2 million Edge Function invocations then $2 per million, and includes 500 concurrent Realtime peak connections then $10 per 1,000, with 5 million Realtime messages then $2.50 per million. A live class or a leaderboard on results day is a Realtime concurrency line item, and 500 included connections is a classroom, not a cohort.

Metered line	Included on Supabase Pro	Overage
Monthly active users	100,000	$0.00325 per MAU
Realtime peak connections	500	$10 per 1,000
Realtime messages	5 million	$2.50 per million
Edge Function invocations	2 million	$2 per million
Egress	250 GB	$0.09 per GB, or $0.03 cached

Treat this table as the exam-week budget, because these are the five lines that move when 50,000 students arrive at once. Cost discipline on usage-based infrastructure is the same discipline as cloud FinOps, which we wrote about in cloud FinOps as a managed service.

What we would cut from a first EdTech build

Live video for every class, when recorded content plus a scheduled doubt session covers the need at a fraction of the cost and none of the concurrency risk. Gamification built on individual behavioural tracking, which collides with rule 4. A recommendation engine before you have the content library to recommend from. Native apps on both platforms before one cross-platform build has proven the product.

What we would not cut: the parental consent flow, the idempotency keys, and the connection pooler. Those three are cheap now and structural later.

India-specific considerations

Three things separate an Indian EdTech build from a global one.

Device reality comes first. The median student device is not the phone in your pocket. Test on a low-memory Android device on a throttled connection, because that is the deployment target, and an app that only feels fast on a flagship is an app that only works for the students who least need it.

Language is second. A product serving NEET or state board candidates is serving students who study in more than one language, and retrofitting localisation into a UI built for English is more expensive than designing for it from the start.

Payment timing is third. EdTech revenue in India clusters around admission and exam cycles rather than spreading evenly, which means the billing system sees the same spike shape as the app. The subscription state machine gets exercised hardest exactly when the infrastructure is under most load, which is an argument for testing them together rather than separately.

FAQ

Why does a bigger database instance not fix exam-week traffic?

Because direct connections plateau. On Supabase a Micro instance allows 60 direct connections and the largest 16XL allows 500, despite costing $3,730 a month against $10. Pooler connections scale to 12,000 on the same instance. A traffic spike is therefore a connection pooling problem rather than a compute problem.

What does DPDP prohibit for children's data?

The Digital Personal Data Protection Act, 2023 makes parental or guardian consent mandatory for anyone under eighteen, bars processing likely to have a detrimental effect on a child, and prohibits tracking, behavioural monitoring and targeted advertising directed at children. Government may prescribe exemptions for specified purposes.

Can an EdTech app run behavioural analytics on students?

Not on children as normally practised. DPDP prohibits tracking and behavioural monitoring directed at children under eighteen. Aggregate product analytics showing that a chapter has high drop-off remain useful. Per-child behavioural profiling and retargeted advertising do not survive the provision, which usually means rethinking an ad-supported free tier.

How should EdTech apps handle offline students?

Build offline-first. Download content ahead of time, queue answers locally, and sync when connectivity returns rather than blocking the interface on a request. Every queued write must carry an idempotency key generated on the device so a retry over a flaky connection does not record the same submission three times.

What does exam-week infrastructure actually cost?

The metered lines move, not the base. Supabase Pro at $25 monthly includes 100,000 monthly active users, then $0.00325 each; 500 Realtime peak connections, then $10 per 1,000; and 250 GB egress, then $0.09 per GB or $0.03 cached. Vercel Pro is a $20 platform fee including $20 credit.

Should we turn off the spend cap before a traffic spike?

It depends which failure you prefer. Supabase enables a spend cap by default on Pro, and Vercel sets spend notifications at $200 per cycle for new customers. For a paid product, uncapped with alerting usually beats a hard stop on exam morning. For a free tier, the cap is usually right.

How far ahead can EdTech capacity be planned in India?

Unusually far, because the spike is on a public calendar. NTA held NEET (UG) 2026 on 3 May 2026 and a re-examination on 21 June 2026, both announced in advance. Supabase bills compute hourly, so scaling up before a known date and down afterwards is straightforward.

Is caching worth it for a learning app?

Yes, because exam-day traffic is overwhelmingly reads of identical content. Cached egress on Supabase Pro is $0.03 per GB against $0.09 uncached. An Amazon API Gateway cache of 1.6 GB costs $0.038 hourly, roughly $28 monthly, against REST calls at $3.50 per million.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, CMMI Level 5 appraised and MSME certified, with senior-led engineering teams and partnerships including AWS, Microsoft and Google. We build learning products for the load pattern India actually has: pooled connections rather than bigger machines, cached reads, offline-first sync with idempotent writes, and a data model that branches on age from the first migration rather than the first legal review. We will also tell you which features to cut before exam week rather than during it. If you are building an EdTech product for Indian students, talk to our team.

References

1. Supabase, Pricing and fees
2. Supabase, Compute and disk
3. Supabase, Organization-based billing
4. Supabase, Manage compute usage
5. Vercel, Vercel Pro plan
6. Vercel, Pricing on Vercel
7. Vercel, Spend management
8. Amazon Web Services, Amazon API Gateway pricing
9. Anirudh Burman, Carnegie Endowment for International Peace, Understanding India's New Data Protection Law
10. National Testing Agency, NEET (UG) official site
11. National Testing Agency, Regarding refund of examination fee of NEET (UG) 2026, examination held on 03 May 2026
12. National Testing Agency, Challenge of provisional answer key for NEET (UG) 2026 held on 21.06.2026
13. National Testing Agency, Information Bulletin for NEET (UG) 2026
14. Stripe, Pricing and fees

Last updated: 16 July 2026.