Skip to content
Keshav Raj

Over-the-air updates for ~660 white-label apps

Keshav Raj · · 4 min read
Signed update packages flying from a server rack to many phones, with a rollback dial.
TL;DR — THE SHORT ANSWER

I built self-hosted over-the-air updates for about 660 white-label apps. Manifests are signed through a key service, so nobody can sign one by hand. Rollout goes 10% then 50% then 100%, and each school can opt out. It went from first release to default-on in four days, at a cost of +0.23 s on cold start.

At an ed-tech company I was the architect and sole implementer of self-hosted over-the-air updates for about 660 white-label apps. Before that, every fix needed a store release for every school app, and that took days.

This post covers what I built, what it cost, and the parts I would reuse.

Why do white-label apps make OTA harder?

A normal app has one binary and one update channel. A white-label fleet has one app per customer. Here, every school had its own app, so a single fix meant a store release for each of them.

Over-the-air updates remove the store from the loop for JavaScript and assets. In a fleet, though, they add a new question. It is no longer "does the update work?" It is "does it work for every school, and what happens when it does not?"

For me that meant per-school native rewrites, so each app was set up correctly for updates, and a device matrix to test against.

The four building blocks of the system I built.

How does the update stay trusted?

In general, OTA systems sign an update's manifest, and the app refuses what it cannot verify. Expo describes this in its code signing guide, and its updates protocol spec says a client must verify the signature before it uses a manifest.

My version used KMS-signed manifests, so nobody can sign one by hand.

Why does that matter? In general, an update channel is a way to change code on thousands of phones at once. Whoever controls it controls those phones, so the signing step is the part to protect first. The same logic applies to a fleet of any size, and it is cheap to add early and painful to add late.

A manifest is a small description of an update. This one is illustrative only, with placeholder values, and it is not code from the project:

Illustrative manifest (placeholder values)
{"bundle": "<bundle-hash>","assets": ["<asset-hash>"],"signature": "<kms-signature>"}

The system also uses two buckets. The full build, including the numbers, is in the Self-hosted OTA case study. App Tower is another fleet project of mine, and you can read about it in its own write-up.

What does a staged rollout protect you from?

A signature tells you an update is genuine. It does not tell you the update is good. Staged rollouts exist for that second problem: if a bad update goes out, only part of the fleet sees it, and you get time to notice before the rest of it does.

I rolled out 10%, then 50%, then 100%. A rollback reaches devices in about 2 minutes. That number matters to me more than raw rollout speed, because it sets how much a mistake can cost.

Then there is the opt-out. Each school can switch updates off with a flag, while the default is on. A flag is cheaper than an argument, and the default means most schools never need to think about it.

Is self-hosting OTA worth the cost?

It was for me, with a cost I chose to accept.

Cold start is +0.23 s slower. Every launch pays that. I took the trade because it made fixes same-day instead of waiting on store releases, and I made updates default-on, with the per-school opt-out as the safety valve. The first release to default-on took 4 days.

The other cost is policy. Apple's App Review Guidelines ask apps to be self-contained, and downloaded code should not change what the app is for. Updates that fix and refine existing features are the safer side of that line. If yours change the product itself, ask someone qualified, not a blog.

My next step is to report adoption within 24 h, along with the store releases the system avoided. Without that, "same-day fixes" stays a claim instead of a number.

What to take from this

  • Sign manifests with a key service, so nobody can sign one by hand.
  • Roll out in steps, and know your rollback time before you need it.
  • Make the default safe for most customers, and give each one an opt-out.
  • Write down the cost you accepted, in numbers.

If you are weighing this for your own fleet, start small. Sign one manifest, ship it to a test app, and see what breaks.

FAQ

Sources

Keshav Raj — Lead engineer building white-label SaaS platforms. Open to Staff, AI and platform roles.

Hire me

Written with AI assistance, reviewed by Keshav Raj.