How to log - the 12 factor application way - node.js

Question

I want to know the best practice behind logging my node application. I was reading the 12 factor app guidelines at https://12factor.net/logs and it states that logs should always be sent to the stdout. Cool, but then how would someone manage logs in production? Is there an application that scoops up whatever is sent to stdout? In addition, is it recommended that I only be logging to stdout and not stderr? I would appreciate a perspective on this matter.

Answer 1

Is there an application that scoops up whatever is sent to stdout?
The page you linked to provides some examples of log management tools, but the simplest version of this would be just redirecting the output of your application to a file. So in bash node app.js > app.out. You could also split your stdout and stderr like node app.js 2> app.err 1> app.out.
You could additionally have some sort of service that collects the logs from this file, and then puts them indexes them for searching somewhere else.
The idea behind the suggestion to only log to stdout is to let the environment control what to do with the logs because the application doesn't necessarily know the environment that it will eventually run within. Furthermore, by treating all logs as an event stream, you leave the choice of what to do with this stream up to the environment. You may want to send the log stream directly to a log aggregation service for instance, or you may want to first preprocess it, and then stream the result somewhere else. If you mandate a specific output such as logging to a file, you reduce the portability of your service.
Two of the primary goals of the 12 factor guidelines are to be "suitable for deployment on modern cloud platforms" and to offer "maximum portability between execution environments". On a cloud platform where you might have ephemeral storage on your instance, or many instances running the same service, you'd want to aggregate your logs into some central store. By providing a log stream, you leave it up to the environment to coordinate how to do this. If you put them directly into a file, then you would have to tailor your environment to wherever each application has decided to put the logs in order to then redirect them to the central store. Using stdout for logs is thus primarily a useful convention.

Answer 2

I think it's a mistake to categorically say "[web] applications should write logs to stdout".
Rather, I would suggest:
a) Professional-quality, robust web apps should HAVE logs
b) The application should treat the "log" as an abstract, "stream" object
c) Ideally, the logger implementation MAY be configured to write to stdout, to stderr, to a file, to a date-stamped file, to a rotating file, filter by severity level, etc. etc. as appropriate.
I would strongly argue that hard-coded writes to stdout, without any intervening "logger" abstraction, is POOR practice.
Here is a good article:
https://blog.risingstack.com/node-js-logging-tutorial/

Answer 3

Cool, but then how would someone manage logs in production?
The log sink is what you're looking for.
Is there an application that scoops up whatever is sent to stdout?
Yes and no. It's the log ship (or log router). It could be an application, but it's really just some process within the execution or runtime environment that your app doesn't really know about.
Another way to look at this is separation of concern. As it was stated in a different answer, it's about letting the environment own what happens to the log and only expecting the application to concern itself with emitting log events at all. I think what's missing from the 12FA documentation is that they don't try to complete the puzzle for you because there will be different opinions on where to go from stdout, so I'll help by adding in those missing pieces based on my personal experience and what I'm seeing all over the cloud space.
Logger sends log event to log stream (aka 'the log')
It goes without saying that your application should have some sort of "logger" abstraction, but that's really just an entry point for emitting a log event to stdout. That abstraction's responsibility is to get your log event onto the log stream (stdout) in the desired format and then your application's responsibility is done. In fact, the 12FA documentation ends here.
12 Factor App is about creating cloud-friendly and portable applications, so you have to assume that you don't know what the executing/runtime environment even is. So we don't know what "the environment" is and that's the whole point. So from here, it is the responsibility of the executing/runtime environment to process the stream and move it to the sink.
Log ship/router realizes log stream to log sink
So the way we solve for this now is to have some sort of listener for the stdout stream that will take the output and send it downstream to the log sink.
The "ship" (also known as the log router or scraper) might be something in the environment or the runtime, or honestly it could be something running the background of your application (a stream listener); it could be some other custom process; it could be even be Kafka -- I think GCP uses fluentd to scoop up logs from various sources and put them in stackdriver. The point is that it should be a separate "class" in your application that your application doesn't really know about. It just listens to the stream and sends it to the sink. In some solutions, this is something you need to build, in other solutions, it's handled by your platform. Put simply "how do I get the stream to the sink?"
The "sink" is the destination. This can be the console (hello it's literally a stream reader), it can be a file, it can be Splunk, Application Insights, Stack Driver, etc. There are simple solutions and there are larger more complex enterprise solutions, but the concept stays the same.
So in short, this is the answer to your question, if we're writing to stdout "how do we manage logs in production." It's the log sink or log aggregator that you're looking for. In 12FA vernacular, something like "splunk" isn't the "log". The log is the stream itself (stdout). In terms of 12FA - Your application doesn't know what the sink is and ideally, it shouldn't because that sink could change, in which case all of your applications would break, or there could be many different sinks and that could bog your application down particularly if you're writing straight to the sinks instead of stdout first. It's just another decoupling exercise if nothing else.
You can send to a single sink, multiple sinks at once, or you can send to a single sink and have some other component 'ship' your logs from that sink to another (e.g. write to a rolling file and have a router scrape that into splunk). Just depends on your needs.
You can actually see this popping up more and more in cloud providers by default. For example, on GCP, all logs to stdout automatically get picked up and sent to stackdriver. In Azure, so long as you add the instrumentation to your .NET application (the application diagnostics package), it will emit events to stdout and it'll get picked up by azure monitor. There are also more and more packages out there that are beginning to implement this pattern, so in .NET you could use Serilog to abstract most of these concepts.
Logger -> Log Event -> Log [stream] (stdout) -> Sink -> Your eyeballs
Logger: The thing you use to emit the log, typically an abstraction (e.g. Serilog, NLog, Log4net)
Log Event: The individual log itself
Log Stream (or 'the log'): stdout it's the unbuffered, time-ordered aggregation of all events and has no beginning or end.
Log Ship/Router: The transport that sends the stream to one or more sinks. (e.g. in process like log4net, out of process like fluentd)
Log Sink: The thing that you're actually looking at like a console, file, or index/search engine, or analytics/monitoring platform (e.g. splunk, datadog, appinsights, stackdriver, etc.)
There are packages and platforms that provide one or more of these pieces, but all of those pieces are always there. It makes 12FA logging make more sense when you're aware of them.