Support for container pebble layer defined externally

[dimaqq]

I have this setup:

rockcraft.yaml

services:
  gubernator:
    override: replace
    startup: enabled
    command: /bin/gubernator
    environment:
      GUBER_TRACING_LEVEL: ...
    on-success: shutdown
    on-failure: shutdown
    on-check-failure:
      online: shutdown

When juju starts a container with the image built from this rock, pebble gets this plan, and is started on hold, then charm does does:

    def _on_gubernator_pebble_ready(self, event: ops.PebbleReadyEvent):
        event.workload.replan()
        self.unit.status = ops.ActiveStatus()

Meanwhile in unit tests, scenario.mocking._MockPebbleClient has a concept of containers, but is not instantiated with the layer data.

Which makes it kinda hard to test things, apart from happy path.

I wonder how to best represent this.

[tonyandrewmeyer]

Can't you provide the plan as a layer in the container? Like:

container = Container("gubernator", layers={"rock": pebble.Layer(...)})

The rockcraft approach is perhaps an argument that _base_plan shouldn't be private. It would be nice if we had some documentation on using the rock's Pebble with charms (I can't find anything at all on juju.is) - that would help be clear on what's expected, and whether a public base_plan makes sense.

[dimaqq]

That may just work, let me wrap up a prototype to share here.

[dimaqq]

Thank you, I can surely instantiate container with pebble layers. \ The next question is though, how to observe the (internal?) mock pebble status update in scenario?

Here's what I have so far:

https://github.com/dimaqq/hexanator/commit/8a4ed2781a7fb799def91253c2d129402cb71333

TL;DR:

• at charm code breakpoint, I can inspect event.workload._pebble._service_status which gets updated by replan
• in the test code, neither the output state, nor the scenario event that triggers the charm code carries this service status

[tonyandrewmeyer]

Yeah, that's where #34 comes in. Right now, the Scenario testing Pebble client doesn't initialise the service status dict so even though the methods are inherited they won't work, if I recall correctly.

So the most straightforward solution is to just mock the restart and assert that it's called. If you really want to avoid mocks then you could do it all with the ops class, I think - before running the event do mgr.charm.unit.containers['...'].stop('...') then run it and then do mgr.charm.unit.containers['...'].get_service('...') and assert on the result. However, I'm not sure that will actually run with Scenario not initialising the internal dict. So it may be that it's Harness or mocking.

[dimaqq]

Ugh, I had a silly bug...

It turns out that I can in fact read back the container status:

https://github.com/dimaqq/hexanator/commit/9358ef82cad95e53c557a0ebc53e25ba0a451266

[tonyandrewmeyer]

The rockcraft approach is perhaps an argument that _base_plan shouldn't be private. It would be nice if we had some documentation on using the rock's Pebble with charms (I can't find anything at all on juju.is) - that would help be clear on what's expected, and whether a public base_plan makes sense.

For what it's worth, if this "use the rock plan" approach is documented as a way people should make charms, then I do think _base_plan (maybe under another name?) maybe should be public.