Great article! Sounds very familiar.

We experienced quite the same issues in Natural Intelligence.

I want to emphasize The Warm-Up Script.

We chose to use Pod Lifecycle, like:

lifecycle:

postStart:

exec:

command:

- '/bin/sh'

- '-c'

- './warm_up.sh || true'

The lifecycle mechanism is more robust in terms of an init script such as the warm-up.

The true flag seen above is meant to determine that the pod will always be launched in order to avoid a potential production issue. The warm-up script can fail due to a lack of connection to download some dependencies. This is also an issue in itself, however we still want to have the deployment independently.