-
Background Datadog introduces "Toto," a novel foundational model for time series prediction optimized for observability data. Toto is trained on extensive datasets to master real-time analysis and scale to large volumes of data efficiently. The observability data includes metrics from the Datadog platform such as memory usage, CPU load, disk I/O, and network throughput, alongside application performance indicators like hit counts and error rates. The complexity and diversity of these data types pose significant challenges for time series prediction.
-
Existing Work Traditional time series forecasting methods, like ARIMA and exponential smoothing, require training a separate model for each metric and fail to generalize across different metric types. Manual retraining and tuning to accommodate data evolution further add to the operational burden. Deep learning-based methods, while more accurate, have seen limited application in time series analysis due to their scalability challenges.
- Introduced a novel model tackling specifically the complexities of observability data
- Challenge 1: Complexity of observability data time series prediction Toto addresses the specific problems of observability data through its architecture and pre-training. It employs an advanced attention mechanism for effective grouping of multivariate time series features, which reduces calculations yet maintains accuracy, and incorporates probabilistic model heads (like the Student-T mixture model head) for capturing the complex dynamics of time series data more accurately.
- Challenge 2: Enhancing observability data forecasting performance The model enhances performance in observability metric forecasting through targeted pre-training on a large-scale dataset of Datadog observability metrics. Such characteristics are absent in open-source datasets, and this focused training ensures improved performance in observability metrics prediction.
Toto outperforms existing time series foundational models on observability metrics while also excelling at general-purpose forecasting tasks, achieving state-of-the-art zero-shot performance on multiple open benchmark datasets. Its innovations include proportional factorized space-time attention mechanisms and the Student-T mixture model head. These novel features have facilitated Toto’s success in handling high-frequency and high-dimensional time series data with good accuracy and efficiency. Toto represents the first general-purpose model for time series forecasting tuned specifically for observability metrics.
The Toto model, developed by Datadog, is a foundation model for time series prediction, specially designed to handle observability data. Its groundbreaking attention mechanism and pre-training strategy significantly improve the performance and efficiency in tackling observability data.