The only missing piece appears to be environ-config.
This **fails to start** with
```
ImportError: cannot import name '_app_ctx_stack' from 'quart' (/usr/lib/python3/dist-packages/quart/__init__.py)
```
`nc` (from netcat) is required for the healthcheck. In
c1cf6ab1e5, the installation was
erroneously moved to the builder image, instead of the final image, so
it was missing since then from the actual application image, causing it
to always show as unhealthy.
Fixes#126.
Dependencies are not necessarily packaged for all architectures. In some cases
(such as aiohttp, and others, on ARM) pip will attempt to compile the
dependency from scratch. Since switching to multi-stage builds, we have been
installing these without a compiler present which caused the build to fail on
ARM architectures.
This commit temporarily installs build-essential packages while running pip,
then removes them again afterwards.
Previously, the multi-stage build increased the image size by about 30
MiB (163MiB -> 191MiB). Dropping the caches reduces the image size down
to 159MiB, leading to a net improvement of 4 MiB.
Currently the Dockerfile has a single RUN directive with all the needed
commands in it. This optimizes for image size by not creating too many
"layers" (which are only additive). However it means the result that gets
cached can basically never be reused, because any change to the source code
will need to execute the whole RUN block again.
This commit switches to a docker "multi-stage" build, where we have a build
image that is separate from the final one that gets published. The build
image can be cached locally, and size is no longer a significant concern.
This approach allows the single RUN command to be split up into multiple RUN
commands that only execute when strictly needed (i.e. when their result
is not cached locally).
This drastically improves the build time when rebuilding the image after
a simple code change, because the build image doesn't have to install all
the apt packages, for example. This leads to a nicer developer experience
when using docker locally for development and testing.
This reduces the overall image size, but more importantly,
deduplicates nicely with the other Snikket images which all use
buster-slim as base.
Fixes#63.
