This technical report describes the Time Series Optimized Transformer for Observability (Toto), a new state of the art foundation model for time series forecasting developed by Datadog. In addition to advancing the state of the art on generalized time series benchmarks in domains such as electricity and weather, this model is the first general-purpose time series forecasting foundation model to be specifically tuned for observability metrics. Toto was trained on a dataset of one trillion time series data points, the largest among all currently published time series foundation models. Alongside publicly available time series datasets, 75% of the data used to train Toto consists of fully anonymous numerical metric data points from the Datadog platform. In our experiments, Toto outperforms existing time series foundation models on observability data. It does this while also excelling at general-purpose forecasting tasks, achieving state-of-the-art zero-shot performance on multiple open benchmark datasets.

Large language model (LLM) agents show promise in automating machine learning (ML) engineering. However, existing agents typically operate in isolation on a given research problem, without engaging with the broader research community, where human researchers often gain insights and contribute by sharing knowledge. To bridge this gap, we introduce MLE-Live, a live evaluation framework designed to assess an agent's ability to communicate with and leverage collective knowledge from a simulated Kaggle research community. Building on this framework, we propose CoMind, an multi-agent system designed to actively integrate external knowledge. CoMind employs an iterative parallel exploration mechanism, developing multiple solutions simultaneously to balance exploratory breadth with implementation depth. On 75 past Kaggle competitions within our MLE-Live framework, CoMind achieves a 36% medal rate, establishing a new state of the art. Critically, when deployed in eight live, ongoing competitions, CoMind outperforms 92.6% of human competitors on average, placing in the top 5% on three official leaderboards and the top 1% on one.

Consistent range-hashing is a technique used in distributed systems, either directly or as a subroutine for consistent hashing, commonly to realize an even and stable data distribution over a variable number of resources. We introduce FlipHash, a consistent range-hashing algorithm with constant time complexity and low memory requirements. Like Jump Consistent Hash, FlipHash is intended for applications where resources can be indexed sequentially. Under this condition, it ensures that keys are hashed evenly across resources and that changing the number of resources only causes keys to be remapped from a removed resource or to an added one, but never shuffled across persisted ones. FlipHash differentiates itself with its low computational cost, achieving constant-time complexity. We show that FlipHash beats Jump Consistent Hash's cost, which is logarithmic in the number of resources, both theoretically and in experiments over practical settings.

Large, distributed data streams are now ubiquitous. High-accuracy sketches with low memory overhead have become the de facto method for analyzing this data. For instance, if we wish to group data by some label and report the largest counts using fixed memory, we need to turn to mergeable heavy hitter sketches that can provide highly accurate approximate counts. Similarly, if we wish to keep track of the number of distinct items in a single set spread across several streams using fixed memory, we can turn to mergeable count distinct sketches that can provide highly accurate set cardinalities. If we were to try to keep track of the cardinality of multiple sets and report only on the largest ones, maintaining individual count distinct sketches for each set can grow unwieldy, especially if the number of sets is not known in advance. We consider the natural combination of the heavy hitters problem with the count distinct problem, the heavy distinct hitters problem: given a stream of $(\ell, x)$ pairs, find all the labels $\ell$ that are paired with a large number of distinct items $x$ using only constant memory. No previous work on heavy distinct hitters has managed to be of practical use in the large, distributed data stream setting. We propose a new algorithm, the Sampling Space-Saving Set Sketch, which combines sketching and sampling techniques and has all the desired properties for size, speed, accuracy, mergeability, and invertibility. We compare our algorithm to several existing solutions to the heavy distinct hitters problem, and provide experimental results across several data sets showing the superiority of the new sketch.