Test in production?

This post is part of the Pride & Paradev series.

With continuous deployment, it is common to release new software into production multiple times a day. A regression test suite, no matter how well designed, may still take over 10 minutes to run, which can lead to bottlenecks in releasing changes to production.

So, do you even need to test before going live? Why not just test changes in production?

Test changes in production

The website for The Guardian, the UK’s third largest newspaper, deploys on average 11 times a day, of which all changes are tested in production.

“Once the code is in production, QA can really start.”

“Sometimes deployments go wrong. We expect that; and we accept it, because people (and machines) go wrong. But the key to dealing with these kind of mistakes is not to lock down the process or extend the breadth, depth and length of regression tests. The solution is to enable people to fix their mistakes quickly, learn, and get back to creating value as soon as possible.”

~ Andy Hume on Real-time QA [The Guardian Developer Blog]

The key to testing changes as soon as they hit production is to have real time, continuous real user experience monitoring. This includes metrics like page views and page load time, which directly correlate to advertising revenue, an incentive to keep these healthy.

More comprehensive automated acceptance tests can be written in a non-destructive style that means they can be run in production. This means that these can be run immediately following a fresh production deployment, and as feedback about the tests is received, any issues can be remedied immediately into production and tested again.

Test changes before production

There are not many businesses that are able to release software without any form of testing into production: whether there be legislative requirements requiring testing, or the risk of introducing errors is too high for its target market.

Whilst automated regression tests do take longer to run than unit or integration tests, there are ways to manage these to ensure the quickest path into production. These strategies include running tests in parallel, only running business critical tests, only running against the single most popular browser, or only running tests that are directly related to your changes.

You can set up a deployment pipeline that runs a selected subset of tests before deploying into production then running the remaining tests (in a test environment). Any of the issues found in subsequent tests are judged to see whether they warrant another immediate release or whether they can be included in the next set of changes being deployed into production.

Whilst you definitely should run tests before deploying to production, it doesn’t mean that this has to drastically hinder your ability to continuously deploy.

Author: Alister Scott

Alister is an Excellence Wrangler for Automattic.