Off the top of my head we use AWS ECS which provides a rolling upgrade method. Push up the new container into ECR (from github actions after they pass tests) run the upgrade command, and new containers will start booting. Once they pass their health checks the load balancer starts serving traffic to them. Once they’re live, the old containers are removed.

We also use a blue/green deployment method so we don’t have to worry about breaking the production database with database changes.

Junior-ish DevOps with some blue/green experience.

It’s a very thorny problem, and I think your willingness to put up with the trade offs really would drive what patttern of architectecture.

Most of our blue/green deployment types use a unitary database behind the backend infra. There’s a lot ways to implement changes to the database (mostly done through scripting in the pipeline, we don’t typically use hibernate or other functionality that wants to control the schema more directly), and it avoids the pain of trying to manage consistency with multiple db instances. It helps that most of our databases are document types, so a lot of db changes can be implemented via flag. But I’ve seen some SQL implementations for table changes that lend themselves to blue/green - you just have to be very mindful to not Bork the current live app with what you’re doing in the background. It requires some planning - not just “shove the script into source control and fire the pipeline.”

If we were using SQL with a tightly integrated schema and/or we couldn’t feature flag, I think we’d have to monkey around with blue/greening the database as well. But consistency is non trivial, especially depending on what kind of app it is. And at least one time when a colleague set up a database stream between AWS accounts, he managed a circular dependency, which….well it wasn’t prod so it wasn’t a big deal, but it easily could’ve been. The data transfer fees are really what kills you. We managed to triple our Dev AWS bill prototyping some database streams at one point. Some of it undoubtedly was inefficient code, but point stands. With most blue/green infra, your actual costs are a lot less than 2x what a ‘unitary’ infra would cost, because most infra is pay for use and isn’t necessary except when you go to deploy new code anyway. But database consistency, at least when we tried it, was way MORE expensive than just 2x the cost of a unitary db, because of the compute and transfer fees.

For me(programmer) it usually depends on the upgrade.

The system I am working on contains like 5-10 containers (depending on customer needs).

In generel I try to have as few connections between services as possible, Low Coupling, High Cohesion. No 2 containers share a database scheme, because that makea database updates way more difficult. The 2 services have to talk to each other through a welldefined API which can be versioned.

1. Application bugfix Sending out a bugfix usually only effects the actions of a single container. Then we upload the new docker image to the server, shuts down the old instance and start the new container, but reuses all the settings of the old container. This does result in like 10-20 seconds downtime, which is completly acceptable in my case.

2. Application feature When we implement a big new feature we usually also needs to update the database (new tables/columns). First take a backup of the database before the upgrade. Then like with the bugfix upload and start the new container. In our case we have the application push out the database update (C# dotnet EF core). If it is an even bigger update we may have multiple docker containers that is upgraded at the same time.

3. Database version upgrade We check the release note for the specific database on how to upgrade it, and follow those steps.

4. Other utility containers (Nginx, Elastic search, …) Same as database, check the release note for breaking changes to make sure the upgrade goes well.

Always always always make sure to test the update in a controlled environment that resembles the production environemnt as closely as possible. (maybe even a copy of production if I am lucky)

Though we are moving to kubernetes & helm soon, currently we use migration scripting tools (like alembic) for schema and data migration on app start, and our infrastructure/devops team uses ansible for deployment. Currently, we don’t have CI/CD straight to production—it’s still a manual process—but I hope to change that as our organization starts using k8s.