Algorithmia

Overview

Algorithmia is an MLOps platform that includes capabilities for data scientists, application developers, and IT operators to deploy, manage, govern, and secure machine learning and other probabilistic models in production.

Algorithmia Insights is a feature of Algorithmia Enterprise and provides a metrics pipeline that can be used to instrument, measure, and monitor your machine learning models. Use cases for monitoring inference-related metrics from machine learning models include detecting model drift, data drift, model bias, etc.

This integration allows you to stream operational metrics as well as user-defined, inference-related metrics from Algorithmia to Kafka to the metrics API in Datadog.

Setup

1. From your Algorithmia instance, configure Algorithmia Insights to connect to a Kafka broker (external to Algorithmia).

2. See the Algorithmia Integrations repository to install, configure, and start the Datadog message forwarding service used in this integration, which forwards metrics from a Kafka topic to the metrics API in Datadog.

Validation

1. From Algorithmia, query an algorithm that has Insights enabled.
2. In the Datadog interface, navigate to the Metrics summary page.
3. Verify that the metrics from Insights are present in Datadog by filtering for algorithmia.

Streaming metrics

This integration streams metrics from Algorithmia when a model that has Insights enabled is queried. Each log entry includes operational metrics and inference-related metrics.

The duration_milliseconds metric is one of the operational metrics that is included in the default payload from Algorithmia. To help you get started, this integration also includes a dashboard and monitor for this default metric.

Additional metrics can include any user-defined, inference-related metrics that are specified in Algorithmia by the algorithm developer. User-defined metrics depend on your specific machine learning framework and use case, but might include values such as prediction probabilities from a regression model in scikit-learn, confidence scores from an image classifier in TensorFlow, or input data from incoming API requests. Note: The message forwarding script provided in this integration prefixes user-defined metrics with algorithmia. in Datadog.

Data Collected

Metrics

algorithmia.duration_milliseconds (gauge)

Duration of algorithm run Shown as millisecond

Service Checks

The Algorithmia check does not include any service checks.

Events

The Algorithmia check does not include any events.

Troubleshooting

Need help? Contact Algorithmia support.