Related to T177 with way more stuff in it now (thanks ;)

Hi @zack Is this task still valid? If so how do I reproduce the error. I tried running pytest in swh-fuse and swh-core but that doesn't throw any error.

Build is green

Minor phrasing fixes

Build is green

• config.spec: Submit first draft
• config.spec: Adress remarks from @ardumont, @vlorentz
• config.spec: Address remarks from @douardda
• config.spec: Rename "singleton" to "record"
• config.spec: Address @ardumont remarks, strip down some sections

Build is green

Rename "singleton" to "record"

Build is green

• Take into account feedback from ardumont and vlorentz
• Take into account feedback from douardda
• still TODO: rename "singletons"

@douardda: taking into account your suggestions.

not finished, but first burst of comments

Closing now since this has been superseded:

1. T1410 tries to summarize our current use in the goal to design a better configuration system.
2. SWHMixin is currently being removed through T1532
3. pytest fixtures helps a lot in bootstraping configuration in testing now. It's now mostly used everywhere.
4. What's described in the description in details is no longer true though the gist of it is (but like i said, see 1.).

In regards to cleaning up deprecated code (cf. T1532#48568):

• D3095: swh.core
• D3953 D3957: swh.deposit
• D4125 D4126 D4127 D4128: swh.loader
• D4140: swh.indexer
• D4141: swh.lister

fallback to default config when the current necessary SWH_CONFIG_FILENAME is not set (instead of current failure)
a service without a configuration should be able to run by itself from a REPL (default-config should target swh services that runs in memory)

I'm not sure we really need these two to happen if the "configuration" of our components ends up being a set of explicit arguments to their initialization function.

In T1410#50054, @ardumont wrote:

Suggestions from irc discussion summary:

• Move the configuration file loading out of the services themselves (loader, lister, indexer, ...) and moves within the scope of the entrypoints instead (swh cli, celery task, gunicorn wsgi, etc...)

Maybe starting a pad/hackmd document would be easier at this point?

Suggestions from irc discussion summary:

Should have said as much here instead...
Nonethess, a way forward for simplifying loaders is described [1]
Implementation proposal is summarized [2] (with implementation diffs reference in summary).

rpc servers, loaders, listers, ... (swh services really): They need a simple
configuration file to parse and load, with eventual checks on missing keys.
In such a case, it should fail early with a clear message about what's
missing (so we can fix fast and restart the service). [1]

For the configuration part, I see 2 use cases: