Should logging really use stderr (and not stdout)?

[alaendle]

Hi, I use the docker image of the opentelemetry-collector and was wondering why the logging information is printed to stderr. I think most tooling expects "docker logs" output to be on stdout. So was it really intended to use stderr here?

https://github.com/open-telemetry/opentelemetry-collector/blob/3f5c7180c51ed67a6f54158ede5e523822e9659e/exporter/loggingexporter/factory.go#L96

Or should a line like

config.OutputPaths = []string{"stdout"}

be added to redirect to stdout.

[morigs]

It's a common (POSIX) convention. STDOUT used for program output (used when piping to redirecting). STDOUT usually contains meaningful generated or transformed data and STDERR used for diagnostics info (such as app logs). See this for more information.

As for docker logs official docs mention:

By default, docker logs shows the command’s STDOUT and STDERR.

[bogdandrutu]

Not sure I understand the comment, do we need to change or keep it as is?

[alaendle]

Thanks for your explanation @morigs ❤️. And I totally agree that this makes sense for a (interactive) console application. However I couldn't see any advantage for a daemonized service like the opentelemetry-collector - simple because normally no one would look at STDOUT; and if analysis is needed you wouldn't analyse STDERR & STDOUT separately.

One could also argue that 'logging' is explicitly configured as an exporter, and that the output of this exporter is real payload in this case (my main argument for a change to STDOUT 😃).

Also at least under my configuration STDOUT was always empty (but maybe other receivers, exporters, processors - that I haven't used - distinguish STDOUT from STDERR). And fore sure docker logs reflects STDOUT and STDERR, but something like docker logs | grep abc will never find logging output that matches abc - which was at least counter-intuitive for me - since this is my first discovery of a container image that uses STDERR this way - maybe because java and .net log to stdout by default (therefore my surprise and the recording of this issue).

Personally I still tend to change the behaviour, but to be honest I haven't the overview of the complete ecosystem of all collector components and therefore couldn't estimate the potential consequences of such a change.

So whatever is correct in the opinion of the maintainers of this project, I agree - and want to thank you for your effort and your hard work.

[rakyll]

Per POSIX convention, the collector's own logs should be in stderr rather than loggingexporter's. loggingexporter's output is the program's output, not its diagnostics log.

[tonglil]

I think the logging exporter should send to stdout.

[rally25rs]

Unfortunately, this causes other tools in the observability chain (OpenObserve, k8s, Loki, DataDog, etc) to report OTEL Collector as continually reporting "errors", so trying to set up alerts on "error count" gets a little more difficult. It's understandable to want to flip this to stdout for hosted deployments.

From the docs it looks like you should be able to set this with:

service:
  telemetry:
    logs:
      level: INFO
      encoding: json
      output_paths: ["stdout"]
      error_output_paths: ["stderr"]

~but with that config and otel/opentelemetry-collector-contrib:0.94.0 image, all output still seems to go to stderr.~ update: Realized my config change wasn't working because I switched docker images from 'normal' to 'contrib' and forgot to change the config path accordingly, so it just wasn't reading my config change.

With the above config, Railway at least starts showing the log entries under their correct color-coded categories for info/warn/error.

[mx-psi]

Kubernetes itself seems to output to stderr https://kubernetes.io/docs/concepts/cluster-administration/system-logs/#klog

Output will always be written to stderr, regardless of the output format

[tonglil]

I don't think following Kubernetes conventions makes sense in this case, especially as they also maintain a wrapper kube-log-runner just to assist with redirecting stderr to stdout. Their design decision to output everything to stderr may mean that all messages are to be treated equally, and that they have different conventions for distinguishing errors from info (ie log verbosity).

Here the collector has an opportunity to distinguish its own (diagnostic) logs from the logs it collects from applications.

In my opinion these two programs, particularly the logging collector, have very different goals in mind. Imagine if Docker captured all output, including its own, into stderr.

[codeboten]

In the long term, the collector should be configurable through the same configuration as the OpenTelemetry Configuration schema provides, would the right way to configure this be to have an option on the stdoutlog exporter to specify which writer to use for the different log levels?

[codeboten]

Here the collector has an opportunity to distinguish its own (diagnostic) logs from the logs it collects from applications.

@tonglil maybe I'm misunderstanding what is being said here, but the collector already distinguishes between its own logs and logs it receives from other applications. Its own logs are emitted as configured by the service::telemetry::logs configuration, whereas logs from other applications are configured by logs pipelines

[relistan]

We could sidestep the debate by making this configurable easily, no? Then people can have the behavior they want, even if the app defaults to stderr.

[mx-psi]

@relistan This is already configurable by setting:

service:
  telemetry:
     logs:
       output_paths: stdout