A bad release on the web is a deploy you roll back. A bad release in an app is out there until people update, and the first warning is usually a review.

The version has to be on everything

Not on a separate report, but on every screen view and every error.

That is what lets you ask the only question that matters after a release: is this worse than the last one, for the people who took it? Without the version on each row you can see that crashes went up and not whether the new build caused them or the old build was always doing it and more people are simply awake.

What a crash-free rate is

The share of sessions that finished without a fatal error.

Two things about the number itself. It never exceeds 100 per cent, and where a crash happened before the app's first screen it errs slightly pessimistic: the session is in the numerator and not the denominator. A stability figure that flattered itself would be worth nothing, so the arithmetic leans the safe way.

What it will not tell you

A crash-free rate is a blunt instrument, and three things hide inside a healthy one.

A feature that fails without crashing. The app stays up and the screen does nothing. Screens per session and where people stop are the signal, not stability.

A slow rollout. If ten per cent have the new build, its problems are diluted to a tenth. Read the rate per version, never in total.

Anything before launch. A session that never started is not in your numbers at all.

Date the release on the chart

An annotation on the day you shipped is the difference between "engagement fell in week 12" and "engagement fell when we shipped 4.2". It takes ten seconds and it is the single most useful thing you can add to an analytics chart.

Old builds keep reporting

Skomi's app wire format is frozen at version 1, so a build published two years ago goes on sending data with the SDK it was compiled against.

That is deliberate, because you can change a site's script this afternoon and you cannot make somebody update an app. It has a consequence worth planning for: your numbers include people on releases you have forgotten about. Filtering by version is not an advanced feature here, it is how you read the reports at all.

Related: one place for app and site analytics, measuring an app, and App Analytics.

The number this rests on is the crash-free rate, read against app version, because one is meaningless without the other.