>I'm specifically asking for that use case.

That's what I answered for.

>I'm not building binaries

If you were, I would have added CPU to the list.

>my "build" steps are actually deployment steps (npm build, composer install, etc)

No, those are build steps. If you weren't using Docker, you would either run all those and shove in a zip/tarball or package into a deb/rpm, etc

>The image I'm deploying by definition also contains my source code

It doesn't contain .git or need credentials to your git/SCM

>I'm not seeing the benefit of the whole "build image, pull on server" pipeline when I can just ditch the registry and added layers by doing those steps on the server as I would normally in other kinds of scenarios

You don't need a registry--you can Docker save/load to push images directly to the server. Images buy you a versioned artifact with all the code-level dependencies baked in. Some maintainer yanks their package from npm? Who cares--you have a copy in your Docker image. Your new app version doesn't work? Edit 1 line to point back to the old image tag and rollback.

>> The build process can exhaust resources on the host

>Maybe, but I've yet to have a host where that's the case for usual CRUD fare.

When the build process completes, it tears down the overlayfs which causes everything to sync which leads to a big I/O spike. Depending on the server and amount of files, it might have no impact. However, I've seen build servers become completely unresponsive for 5+ minutes due to the I/O load when this happens. One place I worked, we had to switch our build servers to NVMe--the Docker container teardown caused spikes over 100k IOPs. Can't remember the exact details--it was React either React web front end or React Native mobile app.

>There's more layers involved there than something like provisioning with Ansible and just having a deploy script to run the usual suspects.

`docker save myimage:tag | gzip | ssh user@server 'gunzip | docker load'`

Not saying creating distributable artifacts is the de-facto answer, but I'd strongly consider whether it's really that much more complicated.