What do you mean? Rolling deployments are a very basic feature of container orchestrators.
Erlang also provides "fix in minutes, in production" stories he refers. For most of languages, it's very hard to remotely get access to an object, let alone fix things directly.
With Erlang, you can inspect every object in the cluster to understand what's going on for any user, and fix the plane in the mid-air instead of rebuilding a plane with the same issue, then go through the process to fix it.
My favorite "load balancing" for web-pages is using "DNS round robin" where you add several A records, and the browser automatically try another IP if the request fails.
The browser tries another IP if it cannot connect to the first one.
If a server throws 500s left and right DSN RR won't help, whereas a load balancer would remove the offending server from the pool.
DNS round robin is not failover. If one of your four A records points to a dead box, 25% of your customers are SOL.
Browsers even try other IPs and alternate between IPv4 and IPv6 before the connection is established.
The whole point of it is to connect to multiple A and AAAA records with a very short delay (250ms in chrome). The browsers use poll to get notified about the first established connection and then use it. So if one A/AAAA records IP is unreachable or slow it will use a another one. There will be no fallout (except maybe for clients that don't use something like happy eyeballs, but any browser does that) if you have multiple A/AAAA records and one of them is down.
We deploy to production several times a day. We have an in-house product that our own emplyees use and they absolutely love that they can ask a feature in the morning and get it in a few hours (in best case scenario). That wouldn't be at all feasible if all those deployments caused errors to users. Luckily that's not a problem, because these days it's really easy to get zero downtime deployments with all this containers and "serverless" functions and whatnot.
With what Dark is promising, I believe testing would be easier. Deploy to a test environment in 50ms. Then run a suite of tests, then deploy to production in 50ms.
There could be some complementary tech that figures out what tests need to be run based on what you change to get the time down.
You only need pip or any package install if your testing image doesn't pre-install the required packages before running the tests. Which I don't myself since I always want to see if my packages work with the latest versions of other packages - and Dark wouldn't be able to handle that anyway.
So my test runs actually complete in about 50 ms if I'm not testing any external API functionality. They will also never finish earlier than 5 minutes if I am, even on Dark - since they're based on external services.
Zero downtime deployments are very rare, and ads a lot of complexity, so it's a big trade-off, compared to just uploading the new code and restarting the service. And having the front-end deal with minor hiccups like service restarts, because users will have such issues all the time due to being on mobile networks, trains going through tunnels etc.
You can do as much testing as you want, both manually and automatically, but you still can't detect all issues as efficient as thousands of users in production. So just accept that there will be issues, and instead design your pipeline so that those issues can be fixed within minutes.