Dataflow in Self-Driving: A Look at Waymo's Stateful Time-Series Processing

As the Waymo fleet scales in size and across geos, low-latency observability is increasingly critical for providing our riders a safe and reliable experience. Stateful time-series processing built on Dataflow is a fundamental component of this observability stack, granting us a bird’s eye view of the fleet as our vehicles move through the world to serve our customers.

In this talk, we’ll investigate how Waymo leverages Dataflow to power our near-real-time use cases such as Fleet Monitoring, Importance Sampling, and more. We’ll also take a look at how Waymo is taking advantage of Google’s TPU fleet and LLMs to introduce streaming inference into the observability equation.