NannyML vs Pulsar

NannyML monitors deployed machine-learning models and estimates their performance when ground truth is delayed or unavailable. It detects concept drift as well as multivariate and univariate data drift, and connects alerts to changes in model performance. Data-quality checks and ranked drift alerts can help teams investigate features associated with performance issues. Users can relate model performance to monetary or other business outcomes. NannyML Cloud supports webhooks for retraining actions and an SDK for automating monitoring-data ingestion. Cloud is available through Azure and AWS marketplaces as SaaS or a managed application. With the managed-application option, the application and required infrastructure are provisioned within the customer's cloud subscription. For its Azure managed application, NannyML documents encrypted traffic and disks, but says automated backups were still being implemented and advises customers to safeguard model inputs and outputs elsewhere. The open-source library installs with pip or conda. Direct local dataset uploads through the Cloud UI must be smaller than 100 MB. Plans include a free self-managed Open Source option, Starter at $399 a month, and Scale at $999 a month.

Pulsar is a free, open-source, self-hosted system for monitoring machine learning models in production. It tracks model drift, concept drift, and performance by comparing incoming behavior with training data used as a baseline. An SDK and containerised ingestion service collect model inputs and outputs into a time series database, supporting historical and real-time data. The current database integration is InfluxDB, and the project is designed for large ingestion volumes without degraded performance. A scheduler checks drift and can trigger standard or customised alerts when metrics reach thresholds. Grafana displays drift measures across selected time ranges and user-defined features. Available drift tests include Kullback-Leibler divergence, Kolmogorov-Smirnov, Levene, and Wasserstein; performance measures include accuracy, recall, F1 score, Brier score, mean squared error, and R2. Deployment logos cover Flask, FastAPI, TensorFlow Extended, Seldon, AWS, Azure, Google Cloud, and Grafana.