stakion-logger 0.0.8

Logger for stakion.

Stakion monitors ML models in realtime by just logging the input and output features.

Using stakion-logger, you can easily send JSON logs to Stakion for analysis.

Requirements

• Python 2.7 or 3.4+

• requests_futures

• socketio

Installation

The easiest way to install the stakion logger is with pip: ` $ pip install stakion-logger `

Quickstart

Stakion has been developed to integrate as a logging library and is therefore very simple to use.

The first step is to initialise the logger:

from stakion import logger

sender = logger.StakionSender(
    project_id=#Your project ID,
    api_key=#Your api key
)

You can then send log messages asynchronously using:

sender.publish(
    correlation_id=correlation_id,
    event_type=event_type,
    model=model_name,
    data=log_message
)

Usage

In order for Stakion to compute monitoring metrics, the log message need to follow a certain format:

correlation_id

The correlation ID is used to identify one single prediction. This is useful to group input and output features together for example. It is also used when searching through the logs

event_type

Stakion supports 5 events types:

• prediction_requested

  Used to compute the total number of predictions made in a certain period. Can be omited if the number of predictions corresponds to the number of unique correlation ID in the input_features and output_features events.

• input_features

  The input features for a prediction.

• output_features

  The input features for a prediction.

• crash_logger

  Used internally when a message can’t be validated.

• error

  Used to track the number of errors.

data

For Stakion to compute accurate metrics, the data object must be correctly formated based on the event_type:

• prediction_requested

  Data should be None

• input_features

  Should be a dictionary where each key is a feature. Example:

  data = {
      "feature_1": feature_1,
      "feature_2": feature_2
  }

• output_features

  Should be a dictionary where each key is a feature. Example:

  data = {
      "outcome": outcome
  }

• error

  Should be a dictionary with the keys:

  data = {
      "level': "CRITICAL",
      "message': "Can't output to log",
      "traceback": traceback.format_exc()
  }