Yes it's aspirational, and of course you have to do batching when the rate of commits gets too high. Many companies don't actually end up at the end of the spectrum with full continuous delivery; most of the value is captured by building what's needed _to be able to_ if you chose to, and doing say daily or hourly automated deploys.

If you're not familiar with the general arguments for why CD might be desirable, a good place to start would be the books Continuous Delivery [https://smile.amazon.com/Continuous-Delivery-Deployment-Auto...] or Accelerate [https://smile.amazon.com/Accelerate-Software-Performing-Tech...]. The TLDR would be that most deploy problems in practice turn out to be because the process is manual, infrequently-used, and/or unreliable, and by practicing "if it hurts, do it more often" you are forced to iron out those kinks. For example you can't deploy automatically without a good acceptance/smoke test suite, and without solid DB migration testing/linting/practices, to name just two of many. You could just build all these things without trying to do CD, but in practice it seems that it's quite hard for organizations to prioritize this work in isolation, and the goal of CD seems to provide a fairly clear map to a much higher-quality deploy process. The practice of deploying more frequently to flush out issues, is also quite unintuitive as you note.

Here's a (probably out-of-date) article on how FB did deploys in 2017: https://www.infoq.com/news/2017/09/facebook-release-scale/ -- batched commits and staged canary rollouts automatically cut and released every few hours.

Yes I do get the general argument for CD (though I have issues with how universally it can be applied based on the sort of BVT I've had to do for SaaS platforms I worked on in the past, where it just wasn't possible to safely release new versions of key components without needing several days' worth of testing, most of which was waiting for verification that interactions with banks etc. had worked as expected).

Agreed, like almost everything it’s not universally applicable.

If you have external contracts that are onerous to validate then it can be quite expensive to build high-fidelity internal representations of those contracts. Could still be ROI positive to do so!

In most cases the contracts are either internal or simple (non-temporal) external contracts like a REST API, so most orgs benefit from moving in this direction.

> customer-facing production sites sounds bonkers

We came from very different lands I guess. Batching 50 commits with no easy way of knowing which commit caused the issue seems scarier to me, in case of an issue. And if you code correctly and have warmup requests, your deployment could be without any downtime. Obviously there will be some extra resource usage because of unready resources.

But if you're doing 50+ deploys a day, how can there possibly enough time to know which commit caused whatever issue that gets noticed anyway? If those 50 deploys are done during regular business hours that's 9 minutes per deploy on average!

Reply to wizofaus: 9 minutes is more than much more than enough to revert, at least in our setup. Specially if some automated alert triggers, then we could revert well within a minute.

That’s what staging environments are for : deploying every commit to check everything is fine, doing the QA and get the alerts.

Doing this in production is being disrespectful of your users. Users are not waiting for your PRs to be approved and merged, they can be as delighted by your new button’s shade of lightgrey if it is released some days (even weeks !) after.

That's my feeling, that you get all the benefits of rapid/frequent deployment by doing so to non-production environments, and deploy to production when you have a good reason to do so. But either way I don't really understand being driven by a desire to ensure every commit results in a deploy. It boggles my mind that a complex product, especially involving mobile platforms, could be fully built, ci-tested, deployed then e2e tested within 10 minutes, even with a massive array of distributed computing power. Mostly due to 3rd party tools that give you no choice but to wait 20 minutes or more for certain steps to complete. I'm intrigued to know how it's done.

I can't even begin to think how to get there, at least using CloudFormation.

More advanced/modern tool chains like k8s can help; rollout can be done in minutes. Also makes it easier to do blue/green deploys if you’re serving high throughput services where you can get an error signal from upgrading a small fraction of your service workloads.