DataDog/toto

What it solves

Toto is a foundation model designed for multivariate time series forecasting, specifically optimized for observability metrics (system monitoring data). It addresses the challenge of predicting future values of multiple related variables across diverse software stacks and services without requiring extensive domain-specific training for every new dataset.

How it works

Toto uses a decoder-only transformer architecture. Toto 2.0 employs a u-μP-scaled transformer with alternating time and variate attention to efficiently process high-dimensional time series. It provides probabilistic forecasting, meaning it generates both point forecasts and uncertainty estimates (quantiles) rather than just a single predicted value.

Toto 1.0 uses a Proportional Factorized Space-Time Attention mechanism and a Student-T mixture model for probabilistic predictions. The project also includes the BOOM (Benchmark of Observability Metrics) dataset for evaluating forecasting performance on real-world observability data.

Who it’s for

• SREs and DevOps Engineers: Those monitoring system metrics to predict potential failures or capacity issues.
• Data Scientists: Researchers working on time series forecasting and foundation models.
• ML Engineers: Developers implementing zero-shot or fine-tuned forecasting models in production observability pipelines.

Highlights

• Zero-Shot Forecasting: Ability to predict future values without fine-tuning on specific datasets.
• Probabilistic Predictions: Generates uncertainty estimates via a quantile head (2.0) or Student-T mixture model (1.0).
• Massive Pre-training: Toto 1.0 was trained on over 2 trillion time series data points, including 1 trillion points from Datadog metrics.
• Scalable Model Family: Toto 2.0 offers models ranging from 4 million to 2.5 billion parameters.
• Exogenous Variable Support: Toto 1.0 supports the use of known future covariates to improve accuracy.