Summary
Datadog has released Toto 2, an open-source family of time series foundation models purpose-built for observability and DevOps telemetry. Standing for Time Series Optimized Transformer for Observability, Toto 2 is designed to forecast metrics like CPU usage, memory consumption, request latency, and network traffic — the kind of data streams that power tools like Datadog, Dynatrace, and Sumo Logic. In this hands-on walkthrough, Fahd Mirza installs and runs the model on an NVIDIA RTX A6000 GPU with 48GB of VRAM, though he notes it can also run on consumer CPUs given its sub-2GB footprint.
The video explains Toto 2’s three core architectural innovations: contiguous patch masking, which predicts entire future windows in a single forward pass; multivariate forecasting, which models relationships between multiple metrics simultaneously; and a quantile output head that produces uncertainty bands (e.g., 10th–90th percentile ranges) rather than single-point predictions. Mirza walks through two live experiments — one using 512 steps of synthetic data with trend and seasonality to forecast the next 96 steps, and a second collecting real system telemetry (CPU, memory, GPU, network) from his own machine for genuine multivariate inference.
Benchmark comparisons shown in the video place larger Toto 2 variants (up to 2.5 billion parameters) among top performers on both the Boom observability benchmark and Gift Eval general time series benchmark. The full setup, from pip install to dashboard rendering, is demonstrated end-to-end, making this a practical reference for engineers considering time series AI for monitoring pipelines.
📺 Source: Fahd Mirza · Published May 15, 2026
🏷️ Format: Hands On Build
