[pkg/stanza] Make Stanza adapter more synchronous by removing channels and workers

[andrzej-stencel]

Component(s)

pkg/stanza

Is your feature request related to a problem? Please describe.

This issue is created as a result of discussion in #31074. From #31074 (comment):

In my view, the current implementation of the Stanza adapter with multiple channels and asynchronous workers if overly complex without delivering a lot of value, instead creating the possibility of data loss.

and:

The conversion [from Stanza entries to OTLP log records] is currently pretty complex, involving multiple asynchronous workers and a bunch of channels.

Describe the solution you'd like

Remove the channels and workers by having LogEmitter convert Stanza entries to log records and call ConsumeLogs synchronously. Leave the 100-log buffering in place for this change, to minimize the impact. Compare benchmarks.

Describe alternatives you've considered

Leaving the 100-log buffering in place leaves the issue of losing logs still not completely resolved. Ideally we want to get rid of this buffering, but we need to step cautiously, as this may have serious performance impact. We should measure this impact and possibly explore possibilities for Stanza receivers like Filelog receiver / File consumer to emit entries in batches, not one by one. Implementing this would alleviate the possible performance impact that removing of batching in LogEmitter may introduce.

Additional context

See #31074

[github-actions]

Pinging code owners:

• pkg/stanza: @djaglowski

See Adding Labels via Comments if you do not have permissions to add labels yourself.

[andrzej-stencel]

Removing needs triage label as this was discussed with the code owner.

[bacherfl]

Hi! I would like to take this on if this is not taken yet

[andrzej-stencel]

It's yours @bacherfl, thanks for picking it up!

[bacherfl]

Thanks @andrzej-stencel! One quick question to clarify: Currently the LogEmitter and its logChannel are also used by two other components:

• pkg/stanza/operator/parser.go
• logstransformprocessor

Should we remove the async channels from the LogEmitter altogether and also adapt those two other components that use it, or do you think we should provide both options (async and sync) in the LogEmitter and leave those two other components unchanged for now?
Both of these other components would be fairly simple to adapt to also work synchronously though.

[andrzej-stencel]

Good point @bacherfl. I wasn't aware that the container parser uses LogEmitter internally. Yes, ideally the logs channel should be removed from the LogEmitter altogether and both the container parser and the Logs Transform processor should use it the same way as the Stanza adapter, having the LogEmitter emit logs (almost) synchronously.

I put "almost" in the sentence above, because we want to keep using the 100-entry buffer in the LogEmitter, which makes it not completely synchronous after removing the channels and workers.