Skip to main content
Engineering

Using Machine Learning to Ensure the Capacity Safety of Individual Microservices

March 7, 2019 / Global
Featured image for Using Machine Learning to Ensure the Capacity Safety of Individual Microservices
Figure 1: The system architecture and design of the Maps Reliability Engineering forecasting system incorporates both data collection and storage and data forecasting, facilitating greater microservice reliability.
Figure 2: The forecasting horizon is roughly 25 percent of the training dataset size. Therefore, if there are more points included in the training dataset, more points can be forecasted.
Figure 3: In this hypothetical scenario, the traffic in data center DC 2 during a failover gradually approaches zero, while DC 1 ingests the traffic that used to be in DC 2. After a couple of hours, both datacenters return to normal operation.
Figure 4: A support vector machine model identifies intervals that correspond to datacenter failovers.
Figure 5: When we compare linear and seasonality-based interpolation techniques for preprocessing data, seasonality represents missing time series data more realistically.
Figure 6: We consider the error margin of this example forecast for an individual service RPS, which comes to 12 percent wMAPE, as acceptable.
Figure 7: Using our prediction system, developers can simulate hourly loads from later days of the week.
Ranjib Dey

Ranjib Dey

Ranjib Dey is a Staff Software Engineer on Uber’s Maps Production Engineering team in San Francisco, CA. Ranjib works on end-to-end resiliency engineering practices across change, incident, and capacity management. Outside Uber, Ranjib is enthusiastic about Open Source and The Internet of Things.

Shrey Desai

Shrey Desai

Shrey Desai is a software engineering intern on Uber’s Maps Reliability Engineering team in San Francisco in 2018.

Ruogu Du

Ruogu Du

Ruogu Du was a software engineering intern on Uber’s Maps Reliability Engineering team in San Francisco in 2018.

Posted by Ranjib Dey, Shrey Desai, Ruogu Du

Category: