Schema eines Canary-Rollouts: Nutzeranfragen laufen in ein Verteilsystem, das 95 Prozent auf die alte Version und 5 Prozent auf die neue Canary-Version leitet; ein Messblock vergleicht Fehlerrate und Antwortzeit beider Gruppen und führt entweder zur stufenweisen Erhöhung auf 100 Prozent oder zum sofortigen Rollback.

Canary Release

A canary release is a new software version that is initially rolled out to only a small portion of users. If errors occur there, the rollout is stopped before the majority is affected.

When a company finishes a new version of its app or website, it can distribute it in two ways. Either all users get it immediately, or only a small portion at first. The second approach is called a canary release. Typically, one to five percent of users receive the new version, while the rest keep working with the old one. If everything goes well, the share is increased step by step until everyone has been switched over in the end. The name comes from the canary that miners used to bring underground with them: the small bird reacted to toxic gases earlier than humans did, making it an early warning system.

Why companies no longer switch all users over at once

Software is tested before release, but tests can never cover everything. In reality, people use a program on thousands of different devices, with different settings and in a wide variety of situations. Some errors only appear when millions of requests arrive simultaneously. A test lab can only replicate this to a limited extent.

The canary release therefore limits the damage. A severe bug then affects one percent of users instead of a hundred percent. For an online shop, that’s the difference between a few annoyed customers and hours of lost revenue. The company’s reputation also suffers less, because most users never notice the problem at all.

A second reason is speed. Because the risk stays small, development teams feel confident releasing new versions more often. Large providers roll out changes several times a day this way. Without this safety net, they would have to ship less frequently and in larger batches, which would make each individual step more dangerous.

From the one-percent test to full rollout

Technically, the old and the new version run in parallel on the servers for a while. An upstream distribution system decides for each request which of the two versions responds. The selection is usually made randomly or based on fixed criteria, such as region or device type. Users cannot tell which group they ended up in.

During this phase, both groups are measured. Observed metrics include, for example, the error rate, server response time, and the number of abandoned orders. What matters is the comparison: it’s not the absolute value that counts, but whether the new group performs worse than the old one. If the error rate rises from 0.1 to 0.8 percent, for instance, that’s a clear warning sign.

There are two possible responses. If the values look unremarkable, the team increases the share, often in stages such as 5, 25, 50, and finally 100 percent. If they look off, traffic is immediately switched back entirely to the old version. This reversal is called a rollback and takes seconds, because the old version is still running. This is precisely what sets it apart from a classic update, where reverting would first have to be laboriously restored.

Canary releases in browsers, chatbots, and company announcements

The term is most visible with Google Chrome. Alongside the regular version, there is a publicly downloadable variant called Chrome Canary, which is updated daily. Anyone who installs it gets new features early, but has to expect crashes. Similar early versions carry names like Beta, Nightly, or Preview at other providers.

With AI services, it usually happens invisibly. When a chatbot is switched to a new language model, often a small portion of users receive the new version first. What’s measured then is, for example, how often responses get cut off or how long the output takes. Related but not identical is the A/B test: there, two variants are compared to find out which one users prefer. With a canary release, on the other hand, the primary concern is technical stability.

In business news, the term comes up when an outage is being explained. Sentences like “the bug was discovered during the canary rollout and rolled back” mean that the safety net worked. A common misconception, by the way, is that a canary release is simply an unfinished test version. The key point isn’t the software’s maturity, but the gradual rollout to real users during live operation.

Related Products

Subscribe free. Unsubscribe the second it sucks.

High-signal news across AI, business, UX, and tech. Every morning.