Problem
trivago sends users to several hundred partner websites (for example agoda.com and expedia.com) and earns revenue from that traffic. Partners differ in size, audience, business model and data quality, and trivago changes its product and business model every day, and its input parameters change dynamically. Traffic therefore jumps all the time, and most jumps are expected: users move from one partner to another while the total stays flat. Analysts need to catch real incidents, such as a partner's data feed going dark, without checking a flood of false alarms.
Example from the article: a partner data outage shows up as daily traffic dropping to zero, and the responsible team must hear about it quickly.
Solution in brief
Forecast expected traffic per partner from its past data, build a confidence interval, and report any actual value outside it. The model is the easy part. The harder work is tailoring monitoring to each partner type (real-time 24/7 alerts for large partners, periodic reports for small ones) and in defining what happens after an alert.
Questions this case answers
- How to monitor thousands of business time series with different data quality and criticality when one alerting policy gives either too many false alarms or missed incidents
- How to decide what response an anomaly alert should trigger when the series is too short or too volatile for a forecast-based detector
- How to detect anomalies in partner-level traffic without drowning the team in false positives?
- How to monitor both large partners with long histories and small partners with short, noisy series?
Tools and methods
- Prophet — forecasting model used in production to predict expected traffic.
- ARIMA+ (Google) — alternative forecaster that gave similar results.
- Elastic ML — tested; not adopted because it is hard to customize.
- Email and Slack messages — delivery of alerts, with charts of the series and the flagged points.
Google or a one-time email code. No password.