starlark

Starlark Processor

The starlark processor calls a Starlark function for each matched metric, allowing for custom programmatic metric processing.

The Starlark language is a dialect of Python, and will be familiar to those who have experience with the Python language. However, there are major differences. Existing Python code is unlikely to work unmodified. The execution environment is sandboxed, and it is not possible to do I/O operations such as reading from files or sockets.

The Starlark specification has details about the syntax and available functions.

Telegraf minimum version: Telegraf 1.15.0

Configuration

[[processors.starlark]]
  ## The Starlark source can be set as a string in this configuration file, or
  ## by referencing a file containing the script.  Only one source or script
  ## should be set at once.

  ## Source of the Starlark script.
  source = '''
def apply(metric):
	return metric
'''

  ## File containing a Starlark script.
  # script = "/usr/local/bin/myscript.star"

Usage

The Starlark code should contain a function called apply that takes a metric as its single argument. The function will be called with each metric, and can return None, a single metric, or a list of metrics.

def apply(metric):
	return metric

For a list of available types and functions that can be used in the code, see the Starlark specification.

In addition to these, the following InfluxDB-specific types and functions are exposed to the script.

• Metric(name): Create a new metric with the given measurement name. The metric will have no tags or fields and defaults to the current time.

• name: The name is a string containing the metric measurement name.

• tags: A dict-like object containing the metric's tags.

• fields: A dict-like object containing the metric's fields. The values may be of type int, float, string, or bool.

• time: The timestamp of the metric as an integer in nanoseconds since the Unix epoch.

• deepcopy(metric): Make a copy of an existing metric.

Python Differences

While Starlark is similar to Python, there are important differences to note:

• Starlark has limited support for error handling and no exceptions. If an error occurs the script will immediately end and Telegraf will drop the metric. Check the Telegraf logfile for details about the error.

• It is not possible to import other packages and the Python standard library is not available.

• It is not possible to open files or sockets.

• These common keywords are not supported in the Starlark grammar:

  as             finally        nonlocal
  assert         from           raise
  class          global         try
  del            import         with
  except         is             yield
  

Libraries available

The ability to load external scripts other than your own is pretty limited. The following libraries are available for loading:

• json: load("json.star", "json") provides the following functions: json.encode(), json.decode(), json.indent(). See json.star for an example.
• log: load("logging.star", "log") provides the following functions: log.debug(), log.info(), log.warn(), log.error(). See logging.star for an example.

If you would like to see support for something else here, please open an issue.

Common Questions

What's the performance cost to using Starlark?

In local tests, it takes about 1µs (1 microsecond) to run a modest script to process one metric. This is going to vary with the size of your script, but the total impact is minimal. At this pace, it's likely not going to be the bottleneck in your Telegraf setup.

How can I drop/delete a metric?

If you don't return the metric it will be deleted. Usually this means the function should return None.

How should I make a copy of a metric?

Use deepcopy(metric) to create a copy of the metric.

How can I return multiple metrics?

You can return a list of metrics:

def apply(metric):
    m2 = deepcopy(metric)
    return [metric, m2]

What happens to a tracking metric if an error occurs in the script?

The metric is marked as undelivered.

How do I create a new metric?

Use the Metric(name) function and set at least one field.

What is the fastest way to iterate over tags/fields?

The fastest way to iterate is to use a for-loop on the tags or fields attribute:

def apply(metric):
    for k in metric.tags:
        pass
    return metric

When you use this form, it is not possible to modify the tags inside the loop, if this is needed you should use one of the .keys(), .values(), or .items() methods:

def apply(metric):
    for k, v in metric.tags.items():
        pass
    return metric

How can I save values across multiple calls to the script?

Telegraf freezes the global scope, which prevents it from being modified. Attempting to modify the global scope will fail with an error.

How to manage errors that occur in the apply function?

In case you need to call some code that may return an error, you can delegate the call to the built-in function catch which takes as argument a Callable and returns the error that occured if any, None otherwise.

So for example:

load("json.star", "json")

def apply(metric):
    error = catch(lambda: failing(metric))
    if error != None:
        # Some code to execute in case of an error
        metric.fields["error"] = error
    return metric

def failing(metric):
    json.decode("non-json-content")

Examples

• drop string fields - Drop fields containing string values.
• drop fields with unexpected type - Drop fields containing unexpected value types.
• json - an example of processing JSON from a field in a metric
• number logic - transform a numerical value to another numerical value
• pivot - Pivots a key's value to be the key for another key.
• ratio - Compute the ratio of two integer fields
• rename - Rename tags or fields using a name mapping.
• scale - Multiply any field by a number
• value filter - remove a metric based on a field value.
• logging - Log messages with the logger of Telegraf
• multiple metrics - Return multiple metrics by using a list of metrics.
• multiple metrics from json array - Builds a new metric from each element of a json array then returns all the created metrics.
• custom error - Return a custom error with fail.
• compare with previous metric - Compare the current metric with the previous one using the shared state.

All examples are in the testdata folder.

Open a Pull Request to add any other useful Starlark examples.