Skip to content

3x orders

BYJU'S, monolith to events

Order and student domains on Kafka, plus $110K a month cut from AWS.

Organisation
BYJU'S
When
2020 - 2023
Areas
Event-driven architecture, Kafka, AWS ECS

Problem

The order and student domains lived in one monolith.

TODO(shubham): What hurt: incidents, scaling limits, release speed.

Constraints

  • The student service had to be low-latency and highly available for 2M+ active users.
  • A team of 4 engineers, which I led and mentored.

TODO(shubham): Other constraints: migration without downtime, deadlines.

What I built

I split the monolith into order and student microservices that talk through Kafka, running on AWS ECS.

From monolith to event-driven services
  1. Monolith
  2. Kafka events
  3. Order service
  4. Student service

TODO(shubham): How the migration was sequenced, and the event design.

Outcome

  • 3x concurrent orders.
  • 40% better performance and 60% less processing time.
  • $110K+ a month saved on AWS.
  • 70% fewer production incidents, with APM monitoring and alerting on Datadog, Prometheus and Kibana.
  • 50% fewer production errors through TDD.

What I'd do differently

TODO(shubham): What you'd change with hindsight.