Thunder
The ad-tracking event pipeline behind an ad platform, now ~1M events a day. I moved it from RabbitMQ to GCP Pub/Sub and added retries and dead-letter queues so events are not lost.
Category
Back-End
Duration
Oct 2025 – present
Events / day
~1M
Broker
RabbitMQ → Pub/Sub
Overview
- I joined Thunder in October 2025 as Senior Backend Engineer. I work on the core Express.js and PostgreSQL services behind its ad platform, and on the messaging layer that carries ad-tracking events between them.
- Ad-tracking events are the part of the system that cannot be dropped, so most of my work there is about delivery guarantees and throughput rather than features.
The problem
- Traffic was growing, and the event pipeline still ran on RabbitMQ.
- Without retries and a dead-letter path, an event that failed processing had nowhere safe to go and could be lost.
- The core services had to stay reliable and fast under peak ad traffic.
The solution
- Migrated the event pipeline from RabbitMQ to GCP Pub/Sub, after weighing Kafka against the operational overhead it would add for a team our size.
- Added retry logic and dead-letter queues to the messaging layer, so failed events are retried and then kept instead of dropped.
- Designed and improved core Express.js and PostgreSQL services for API reliability and performance under peak traffic.
Decisions & trade-offs
GCP Pub/Sub over Kafka
The pipeline had outgrown RabbitMQ. Kafka was the obvious candidate, but a Kafka cluster needs someone to run it, and we are a small team. Pub/Sub is managed, scales without capacity planning, and carries ~1M events a day for us. The cost: less control than Kafka over retention, replay and partition-level ordering, and a bill that grows with message volume.
Dead-letter queues instead of retrying forever
An event that can never succeed would otherwise be retried indefinitely or dropped. Failed events are retried, and after the maximum attempts they move to a dead-letter topic, where they are kept for inspection instead of lost. The cost: a dead-letter queue is only useful if someone watches it and can replay from it.
Outcomes
- The pipeline now carries ~1M events a day on GCP Pub/Sub.
- Failed events are retried, then kept in a dead-letter topic instead of being lost.
- No Kafka cluster for a small team to operate.
My role
Senior Backend Engineer
RabbitMQ, Express.js, Node.js, PostgreSQL