Log NotEntitledExceptions using logger with <component>.<package> suffix

[mosche]

Log NotEntitledExceptions using logger with <component>.<package> suffix rather than <component>.<module> for more fine-grained mutes.

Relates to ES-12231

[elasticsearchmachine]

Pinging @elastic/es-core-infra (Team:Core/Infra)

[mosche]

this fixes a potential NPE: if callerClass != null, requestingClass is callerClass

[ldematte]

LGTM!
Just double check please if any of the existing mutes need updating to go with this change.

[mosche]

Ugh, you're right @ldematte ... and the current mutes have no comment as for which class failed previously, particularly hard with ALL-UNNAMED :/

Changed it to <component>.<module>.<package> so old mutes will continue working.
Unfortunately logger names might get really long that way, but this should happen only rarely

[prdoyle]

It's a bit unfortunate to be burdened with legacy already, so early in the lifecycle of this project, but I can live with this if Lorenzo likes it.

<component>.<module>.<package> is clever, though it is somewhat ambiguous because a logger called component.a.b.c could be module a and package b.c or it could be module a.b and package c.

If we wanted to go with <component>.<package>...

The initial set of muted warnings probably occur during the CI tests, so we could probably delete the mutes and look for warnings in the CI output to determine the packages for those. For mutes added later, a git blame could take us to the PR that added them, and hopefully the PR would record enough info that we could determine the package name.

[ldematte]

If we want, could we use a different separator? And/or make <module> optional?
E.g. <component>.<module>#<package> and/or allow <component>#<package> (no module)?
Just an idea, I think I'm fine with <component>.<module>.<package> too :)

[mosche]

The initial set of muted warnings probably occur during the CI tests, so we could probably delete the mutes and look for warnings in the CI output to determine the packages for those. For mutes added later, a git blame could take us to the PR that added them, and hopefully the PR would record enough info that we could determine the package name.

I checked old PRs that added mutes, there's not much context documented there :/

If we want, could we use a different separator? And/or make optional?

Making <module> optional would also break previous mutes.
The separator has to be a dot I think, otherwise logger inheritance won't work

[mosche]

.. is clever, though it is somewhat ambiguous because a logger called component.a.b.c could be module a and package b.c or it could be module a.b and package c.

An alternative could be to have both <component>.<module> and <component>.<package> and prior to returning a logger for <component>.<package> check if <component>.<module> is muted, and if so return a no-op logger.
That way we'd support both module and package based mutes without overcomplicating the logger names for package based mutes.

It's just not vanilla log4j anymore that way and one has to remember there's two different ways of doing mutes.

[ldematte]

The separator has to be a dot I think, otherwise logger inheritance won't work

Let's go with the current scheme then (<component>.<module>.<package>)

[elasticsearchmachine]

💚 Backport successful

Status	Branch	Result
✅	9.1
✅	8.19