I think the author mainly refers to the situations where terraform can become a bit complex to manage application deployment.

For example, at some point we experimented with building a dataflow template, uploading it and then deploying it, all from terraform. The terraform code contained quite a lot of local bash and gcloud executions, making it depend on the environment itself. Moreover, when something went wrong, it was difficult to debug the code.

We stepped away from doing it all in terraform and are now combining it more with different parts of our CI/CD, as you mention as well. :)

There is less of a boundary between infra and application deployments, especially when you look at the PaaS services like Cloud Run, Cloud Functions, Cloud Dataflow, etc. As setting one up is almost no effort, one will quickly resort to terraform as there seems to be no point to have yet another tool, just for e.g. launching your function.

In the end I think it - yet again - comes down to finding the right balance, which depends on the specific case. :)