Delivery Hero moved ad measurement from hourly batches to real time with Amazon Managed Service for Apache Flink

Delivery Hero operates an advertising platform across roughly 65 countries, connecting more than 1.5 million restaurant partners and local vendors to millions of consumers. The platform handles tens of thousands of messages per second and processes billions of ad events per day. Advertising contributed nearly EUR 1.5 billion in revenue in 2025. That scale forced a re-evaluation of how ad impressions, clicks, and orders were measured and billed.

The problem with the legacy pipeline

• No event-time semantics: Events were bucketed by processing time because most lacked usable event timestamps. When ingestion lagged or events arrived out of order, time-based metrics (for example ROAS) skewed.
• Slow processing: Hourly batching produced an average latency of 61 minutes between when an event occurred and when it was recorded. That delay was too long for budget pacing and real-time ad serving.
• Synchronous enrichment: Each event triggered blocking external API calls to fetch campaign and product metadata. Spikes exhausted connection pools and caused cascading failures across billing, ad serving, and reporting.

Incomplete event context compounded data quality problems. Enrichment was best-effort and synchronous, so some events were written with blank fields. Missing session IDs affected roughly 30–40% of events, leaving a significant share unattached to user sessions.

Practical takeaways for teams considering a similar move

• Event-time processing matters: If events arrive out of order or without timestamps, your metrics will drift when you rely on processing time.
• Avoid synchronous per-event enrichment calls in high-throughput systems: they create back-pressure and hidden coupling between ingestion and external services.
• Reassess database choices as workloads change: storage and access patterns that worked for occasional document reads don't hold up for continuous per-event mutations and deduplication.
• Build reprocessing paths into the streaming architecture so recovery doesn't require ad-hoc tooling or manual coordination.

Delivery Hero's migration to a managed Flink service shows how a streaming-first architecture can reduce latency and costs while improving the reliability of billing-grade ad measurement.