Last Minute Notes
  • Github
  • Docs
  • Learn
    @sayanstwt
Lessons
0/37 done
  • Timeouts, Retries & Breakers
  • Multi-Region Design
  • Cells & Shuffle Sharding
  • SLIs, SLOs & Error Budgets
  • Observability
  • Deployment Strategies
  • Backups & Disaster Recovery
Learn/System Design/Reliability & Operations

Cells & Shuffle Sharding

7 min Advanced 7 steps

Shrink the blast radius so one bad tenant hurts few others.

Fleet
one, shared
Customers hit
-
Blast radius
everyone
Customer AServerCustomer BLoad balancerServerDatabasesharedServerCustomer C
request response data
In this lessonStep 1 of 7
  1. The problem
  2. Cells
  3. Fail small
  4. Shuffle shard
  5. Poison
  6. Static stability
  7. Recap
Step 1 / 7problem

One fleet, one fate

Every customer shares the same servers and database.

So a bad deploy, an overload, or one poisonous customer can take down everyone at the same moment.

Key idea

The blast radius of a failure is everything that shares the broken part.

←→ step·R replay·Space pause

You'll learn

  • Why a shared fleet makes every failure a total failure
  • Split the system into cells behind a thin router
  • Fail small with staged rollouts and health gates
  • Use shuffle sharding to isolate a noisy or poisonous customer
  • Keep the data plane statically stable without the control plane
Read the full notesDistributed Systems, 13