Logging -- we've added

log_stdout=True

to all of our tasks, and the log messages are showing up just fine in the agent's stdout, but we'd like them to also be available on the web ui. We've discovered that by getting the logger from the context, we can do this, but this is a pretty invasive change for our codebase (most of our modules are using standard logging and we don't want to change that). Is there any magic we can invoke to get our logger messages to show up on the web ui?

Hi Dolor, You're saying that with

log_stdout=True

print statements from tasks show up as logs locally, but aren't visible in the webui?

correct

Oh wait, it sounds like you're expecting

log_stdout

to properly pick up logs generated via other loggers within your tasks? This won't be the case, the way python's logging `StreamHandler`s work our stdout redirection won't properly catch these. You're only seeing things work locally because if you stream local logs through the local agent you're actually just streaming all process output, which is different.

our _LOGGER is by default logging to stdout

Yes, but python's logging mechanism grabs

sys.stdout

early, meaning that redirecting it in a task won't actually work.

interesting -- ok, will try that, ty

Python loggers are singletons and are hierarchical. Assuming the downstream libraries don't configure their loggers themselves on import (and only create them) you can make use of prefect's

extra_loggers

configuration. Loggers with names specified in this list will be configured to log to prefect. Since loggers are hierarchical (and assuming the libraries are using

__name__

for the logger names) this would be a list of package names:

# in your ~/.prefect/config.toml
[logging]
extra_loggers = ['package1', 'package2']

yeah ... that list of packages is a moving target in my enviorrment, so a priori I don't know them

You could also configure via an environment variable, e.g.

PREFECT__LOGGING__EXTRA_LOGGERS="['package1', 'package2', ...]"

if that's easier. Doing it programmatically is fine, but won't transfer to other processes if you're e.g. using a

DaskExecutor

.

yeah, sadly I don't have the list of packages (and even if I did, the list is too volatile to maintain). Yeah we're using a dask executor ... I guess we'll do it once for each @task

That might get chaotic. You'd need to ensure whatever modifications you make are idempotent and threadsafe so you don't add the same handlers more than once.

yeah 😕

You also don't need to do that for everything in

loggerDict

, just the root parents, e.g.

foo

but not

foo.bar

. Most python libraries I interact with don't use logging, I'm surprised the list you do is so large you can't statically configure it once.

good idea -- I'll capture just the root of the namespace, as you say

Excellent! Glad to hear it. Let us know if you run into any other issues.

@Marvin archive “I don’t see all the logs that I expect when using log_stdout=True”