Multi-Tenant messaging systems
summary
Compares concrete isolation patterns and trade-offs for multi-tenant queues: one global MTQ is simplest but risks starvation; per-tenant STQs eliminate interference but blow up cost and ops; cell sharding limits blast radius (four cells → ~25% affected); shuffling assigns tenants to multiple shards (example: 100 shards, 5 per tenant → collision ≈ 1 in 75 million) but forces producers to poll shard depths. Describes dynamic overflow queues and the new SQS Fair Queues feature that auto-detects noisy tenants and creates subqueues, plus the cleanup and consumer-priority trade-offs.
AI-written summary · about
chapters (12)
- 0:00Multi-tenant messaging intro
- 1:10Ride booking scenario
- 3:21Noisy neighbors in queues
- 4:46Mitigating noisy neighbors
- 6:10Load shedding and back pressure
- 7:56Single vs multi-tenant queues
- 9:00Cell sharding blast radius
- 10:44Shuffle sharding isolation
- 14:03Hybrid tenant isolation
- 16:53Detecting noisy neighbors
- 20:28Amazon SQS fair queues
- 22:39Trade-offs and resources
related talks
02h 42m!!! Do not publish or share !!!!
25mAvoiding Overload in Distributed Systems
16mTo queue or not to queue, that is the question.
40mEverything Will Flow
12mCeph QoS: How to support QoS in distributed storage system
40mSystem Performance and Queuing Theory - Concepts and Application
52mWhy today, more then ever you need Dependency Isolation in Distributed Systems
40mQueues and Locks: Theory and Practice