Cache placement

Problem

What becomes confusing, fragile, or impossible without understanding cache placement? This lesson answers that through explanation, a worked example, two runnable exercises, and a reference solution. No teacher-supplied worksheet is required.

Distributed architectures address concrete capacity, reliability, and coordination limits—but often arrive before those limits do.

Concept explanation

Cache placement belongs to “Earn caches, jobs, and queues”. Caches, jobs, queues, and rate limits add failure modes and must be earned by measured constraint.

For cache placement, trace concrete input, state transition, output, and failure through an architecture decision tied to explicit requirements, estimates, state ownership, failure modes, and measured limits.

Start with requirements and estimates, choose the smallest architecture, locate its first bottleneck, then evolve one measured constraint at a time. Apply that model to supplied normal, boundary, and failure cases; each case below names its input and expected evidence.

Example bank

Compare normal, boundary, failure, and cross-layer cases. Predict each observation before revealing the explanation.

Worked example

Start from supplied design.md. Focus: Evolve notification system from synchronous send to durable job queue.

1. Run: Fill design.md, then verify every component traces to one stated requirement or measured constraint.
2. Save baseline evidence. Trace requests, state ownership, capacity, queues, caches, replication, failure domains, recovery, operational load, and cost.
3. Boundary case: Define retries, idempotency, poison message, and per-user rate limit.
4. Failure case: Kill worker after side effect before acknowledgment.

RESULT
Recovery avoids silent loss and documents duplicate policy, backoff, queue depth, and invalidation. Starter-level baseline: Document names one action, target, non-goal, numeric estimate, state owner, request path, bottleneck, and recovery procedure.

Starter material

Create design.md, paste this exact content, then run the command below.

# cache placement

## Requirements
- One primary user action:
- One reliability target:
- One explicit non-goal:

## Estimate
- Requests/second:
- Stored bytes/day:
- Peak concurrent work:

## Smallest design
- State owner:
- Request path:
- First bottleneck:
- Recovery procedure:

STOP / CLEANUPStop any process started by your slice with Ctrl+C; run git status before leaving.

Independent exercise

Find the boundary: cache placement

1. Create second case from blank file: Define retries, idempotency, poison message, and per-user rate limit.
2. Then create controlled failure: Kill worker after side effect before acknowledgment.
3. Use Architecture notebook · load tests · traces · Git history to prove behavior, then repair controlled failure.
4. Compare result against supplied acceptance checks and reference approach before marking complete.

Expected result

• Recovery avoids silent loss and documents duplicate policy, backoff, queue depth, and invalidation.
• Document names one action, target, non-goal, numeric estimate, state owner, request path, bottleneck, and recovery procedure.
• Controlled cache placement failure produces captured evidence; repair restores stated invariant without hiding error.

Hints

Solution

Attempt both exercises before opening reference approach.