Skip to main content
Uber logo

Schedule rides in advance

Reserve a rideReserve a ride

Schedule rides in advance

Reserve a rideReserve a ride
Data / ML

Implementing Model-Agnosticism in Uber’s Real-Time Anomaly Detection Platform

February 14, 2018 / Global
Featured image for Implementing Model-Agnosticism in Uber’s Real-Time Anomaly Detection Platform
Figure 1: Powered by our F3 forecasting layer, Uber’s anomaly detection platform tracks dynamic thresholds for seasonal metrics.
Figure 2: Under the pre-existing data model, integrating a new model required growing the threshold keyspace linearly. Consumers’ threshold selection becomes implicitly coupled to model selection, and only one model’s expiration time can be recorded at a time.
Image
Figure 3: Under the new data model, integration of a new model only requires adding a new “model identifier” symbol. The threshold keyspace remains constant, and the data model itself directly reflects the separation of distinct concepts.
Figure 4: An illustration of the “impute-and-inject” pattern used as part of the storage layer’s backwards-compatibility logic. Also shown is one of the “workarounds” that initial new model development employed to circumvent the data model’s limitations, as well as this pattern’s ability to take that workaround into account as part of the pre-existing API.
Jonathan Jin

Jonathan Jin

Jonathan Jin is a software engineer on Uber’s Observability Applications team in New York. He likes dry, minerally white wines, folk music, and light-roasted pour-over coffee.

Posted by Jonathan Jin

Category: