Switch from "printing" to "logging"

[pmcurtin]

Should we use Python's built-in logging? we could also use some third-party logging library - not sure if you have any favorites

[apoorvkh]

Yes, we should use Python's built-in logging.

I saw this video a while ago, but I am no expert and there may be a better resource. https://www.youtube.com/watch?v=9L77QExPmI0

We should certainly keep it as simple as possible.

[apoorvkh]

Perhaps we should stream all logs up from the workers -> agents -> launcher and then we can choose how to filter and save logs at the launcher-level, such that the user can customize this. We can include some default behaviors and utilities.

[pmcurtin]

I think I'll adapt this solution: https://docs.python.org/2.7/howto/logging-cookbook.html#sending-and-receiving-logging-events-across-a-network

[apoorvkh]

Sounds good -- just be aware those are Python 2 docs :)

[apoorvkh]

also fyi, I keep seeing that print statements appear at the end of the agent log / launcher terminal, whereas logger.info statements appear in the "realtime" order

we can see about that after the current logging adjustments

[pmcurtin]

That's pretty strange. Well, yeah we'll see what it's like once we get logging working.

[pmcurtin]

@apoorvkh Should we still provide a way for worker output to be captured? Else I can just remove all the existing log stuff and we can completely depend on the logging library.

[apoorvkh]

I was thinking that all the workers (e.g. in entrypoint) and agents (e.g. in main) would stream their logs directly to the launcher. Was there something else you had in mind?

[pmcurtin]

That's what I'm doing. I was just curious whether you thought stdout/stderr should also be captured. idk, I guess code in libraries that people might use could output to those streams rather than log? for now I'll ignore it.

[apoorvkh]

stdout/stderr should also be captured

Oh is there a difference between: logging output and std output?

Yeah, I think we should also be streaming the stderr/stdout.

[pmcurtin]

Oh is there a difference between: logging output and std output?

Well, the way we're streaming the logs is a little more advanced than just sending strings from the workers/agents to the launcher. We're sending LogRecord objects which get processed by the logging library. These objects are only created when people log using logger.debug(msg)...

Yeah, I think we should also be streaming the stderr/stdout.

Ok! Just for single worker, perhaps? The launcher process should print this?

[apoorvkh]

Hmm, I guess I don't understand enough about how you're implementing this. I think we need to talk about it in more detail tomorrow.

logger.debug(msg)

I think we should be streaming all logging events, not just debug. Alternatively, maybe we don't need to stream "logging events", but just stderr / stdout, which should include logging statements (at whichever granularity is specified).

Ok! Just for single worker, perhaps? The launcher process should print this?

No, I think we should stream everything. Later, we can include a filtering mechanism to choose what the launcher process will print.

[apoorvkh]

I guess it's good to stream "logging events", rather than exclusively just an stdout/err byte stream. Because that will offer deeply granular control for filtering outputs from the launcher.

[apoorvkh]

Also, maybe we can just redirect stdout to logging

E.g. https://stackoverflow.com/questions/11124093/redirect-python-print-output-to-logger