Opentelemetry-specification: Ability to name logger - GetLogger(name) and retain this information on event

[maxgolov]

What are you trying to achieve?

Most prominent frameworks across different programming languages allow to name a Logger. This is typically done at GetLogger(name) API call, where subsequently the knowledge of logger name is retained on Logger object. When SDK flows the data emitted from the Logger to Exporter, it should be possible for the Logger to pass-thru not only the Event name, but Logger name that emitted this event.

Additional context.

While we are designing Logging API in OpenTelemetry C++ SDK, we are facing a challenge that the spec does not prescribe the Logger to have a name per se. Whereas if you build a shim or exporter from other prominent logging SDKs to OpenTelemetry SDK, you'd actually have to 'remember' the Logger name passed down to it.

References

Logging frameworks that provide named loggers and provide ability to search/grep/export based on logger name:

• log4j: https://logging.apache.org/log4j/1.2/manual.html - Cit. : Loggers are named entities
• log4cxx: https://logging.apache.org/log4cxx/latest_stable/usage.html
• Python logging cookbook: https://docs.python.org/3/howto/logging-cookbook.html
• Xamarin / C# Logging API: https://docs.microsoft.com/en-us/dotnet/api/java.util.logging.logger.getlogger?view=xamarin-android-sdk-9

Example showing visually why it is important for an Event to carry information about what logger emitted this event:

Column refers explicitly to what logger name it is.

Similar logic also applies to Android OS / platform code, where there is no logger name concept per se, but it can be implicitly mapped to LOG_TAG: https://stackoverflow.com/questions/8355632/how-do-you-usually-tag-log-entries-android - which allows to group / filter by your "named logger" events from the unified stream of all Logcat messages / events.

[maxgolov]

@arminru @andrewhsu

Note that two most popular 3rd party C# libraries also provide the same API / method - GetLogger(name) :

• log4net: https://logging.apache.org/log4net/release/sdk/html/M_log4net_LogManager_GetLogger_2.htm
• NLog: https://nlog-project.org/documentation/v4.3.0/html/M_NLog_LogManager_GetLogger.htm

I believe this issue is absolutely necessary and should be blocking for the Logging API GA milestone. Without this feature it is impossible to bridge the most popular logging frameworks down to OpenTelemetry API surface, i.e. there will be data loss on export from log4j, log4cxx, log4net, Python, Xamarin, NLog, etc.

Why this is relevant: while we are designing OpenTelemetry C++ API surface, it would be great to clarify the spec position w.r.t. adding GetLogger(name) API.

[arminru]

Hi @maxgolov! Thanks for your input. The label "release:after-ga" is about the overall OpenTelemetry 1.0 GA release which will contain traces and metrics but logs are not part of that yet. I expect that those labels will be removed once the next release is tackled and then all issues regarding logs will be addressed.

[maxgolov]

Discussed in Log SIG meeting today: the ask is either for a _top-level field_ or for an _attribute_ in protocol spec. There is no immediate need (no urgency) to have it in the protocol, as long as we have provisions that the API layer + SDK implementation, up to exporter level retain the knowledge of logger name (log name). Subsequently once the field or attribute is added to protocol spec, the exporter may (optionally) populate this new TBD location with the logger (log) name.

[tigrannajaryan]

One possible approach is to do what tracing does with the name passed to GetTracer(name) - put the name in the InstrumentationLibrary.Name, which is a field in OTLP and has mappings defined for other protocols.

I am not quite sure though that this is a very good approach. It seems like historically the Logger was often used as the name of the classes (especially in Java) as opposed to the names of libraries. If Otel-based applications using Otel GetLogger continued the practice of using the class names for logger names it would result in the class name in the InstrumentationLibrary.Name which doesn't sound right.

Perhaps we do need a separate top-level field or a semantic convention for a logger name attribute.

[tigrannajaryan]

Also, note that GetTracer accepts optional library version number. If we decide that GetLogger follows GetTracer approach for consistency then it will have to also accept the version number (which makes it even less appropriate to be used with a class name).

[tigrannajaryan]

We have had a discussion with @zenmoto. It looks like logger name does not fit well any of the existing log data model fields or the InstrumentationLibrary.name. For exampe Java Logger says:

A name for the logger. This should be a dot-separated name and should normally be based on the package name or class name of the subsystem, such as java.net or javax.swing

The package name fits somewhat our definition of InstrumentationLibrary.name (however not fully, a library may be composed of multiple packages). The class name doesn’t fit at all. Java’s recommendations are not specific enough to attribute to any particular existing concept in Otel. We can in theory have separate fields for “library name” (which already exists), “package name”, “class name”, etc. I think this alternate approach won’t work well because it is very language specific (what do you do for languages that have no classes or have no notion of packages but instead have compilation units or namespaces?).

It appears that logger name is sufficiently different to warrant a new concept in Otel logs. The following definition is a possible candidate for a new field in Otel log data model:

LoggerName - name of the logger that is used for producing the log record. The name is typically chosen by the developer to group related log records. Often the class name, package name or other logical unit of the application is selected as the logger name.

One open question is around the values of the LoggerName. In Java for example the value of the logger name can be the fully qualified class name that includes the package name as a prefix e.g. com.myapp.commerce.backend.data.MyClass. This allows for filtering of logs by different hierarchical scopes (e.g. show all `com.myapp.commerce.backend.data.*). It is not clear if we want to prescribe or recommend this behavior for the values in any way, but I am leaning on saying, no, we shouldn't.

Thoughts on this are welcome.

@kumoroku what do you think?

cc @open-telemetry/specs-approvers

[maxgolov]

LoggerName - name of the logger that is used for producing the log record. The name is typically chosen by the developer to group related log records. Often the class name, package name or other logical unit of the application is selected as the logger name.

I like your definition.

One open question is around the values of the LoggerName

My personal opinion we can give freedom of choice to developer. In some systems / exporters (e.g. Java), it may be of benefit to arrange the loggers in tree, like log4j does (but it does not enforce this - the loggers might be all top-level, just a flat list, not a tree). Whereas in other systems (e.g. Android-Java) - the concept of named logger could be associated with LOG_TAG, which could be either a fully qualified name or a short unique tag, such as Storage or NetworkThread or alike. Real-world example:

    private static final String LOG_TAG = "ImsPhoneCall"

I would like to get a confirmation that the concept is valuable and will be supported by Logging spec. So at this point - what we can do in OpenTelemetry C++ SDK, for example, we reserve a storage location in the Logger object and identify some ways of passing the logger name to exporter. Then it will be up to concrete exporter to decide what attribute, top-level field or custom (or maybe even drop as unnecessary?), this value should be assigned to. This will give us ability to experiment with the field before the final OTLP spec is approved.

[kumoroku]

I like the definition as well, and I agree that we should not prescribe a format for the string––even if there is a hierarchy, it might use different separators, for example.

incidentally, I happened to need to do some logging in Rust this week. I am not a Rust pro but I looked at both log4rs and fern. Log4rs, predictably, has a notion of loggers, and the docs suggest a hierarchical model there as well. Fern doesn't call out the "logger" concept but there is a way to set different levels based on Rust module, which are hierarchical from what I understand and are using "::" as a separator... close enough in my book, conceptually. long story short, @tigrannajaryan you might want to pick up on the term "module" for the definition.