Appearance
14.3 Networking & Waiting Room SSE Streams
1. Concept & Plain-Words Overview
DNS routing, edge CDN caching, Layer 7 load balancing, and comparing REST vs WebSockets vs Server-Sent Events.
In 18-flash-sale-inventory, this domain is engineered to eliminate contention, maintain strict transactional invariants, and deliver predictable sub-millisecond latencies.
2. Implementation in This Project
- Key Codebase References:
backend/src/flow/,backend/src/stock/,backend/src/models/ - Guarantees: Conservation invariant, zero oversell, automated 300s TTL hold release.
3. Comparison of Architectural Approaches
| Architectural Choice | Chosen Pattern [This Project] | Industry Alternative A | Industry Alternative B |
|---|---|---|---|
| Throughput Under Contention | High (Metered & Level) | Low (Pessimistic Queues) | Variable (Optimistic Aborts) |
| Consistency Guarantee | Strict Serializability / ACID | Eventual Consistency | Read Committed |
| Failure Modes | Graceful 409 Out-of-Stock | Cascading Timeouts | Negative Stock Overselling |
4. What Breaks at 10x, 100x, and 1,000x Load
- At 10x (5,000 RPS): In-memory waiting room and single PostgreSQL instance perform smoothly without saturation.
- At 100x (50,000 RPS): PostgreSQL row lock contention on a single SKU becomes a bottleneck; requires Redis Lua offload.
- At 1,000x (500,000 RPS): Single Redis instance hits network I/O limits; requires multi-cluster Redis key partitioning by SKU ID.
5. 30-Second Spoken Answer
"In
18-flash-sale-inventory, we design for extreme burst resilience by decoupling high-frequency availability checks from heavy database transactions. By leveling incoming traffic through a virtual waiting room and executing atomic stock deductions, the architecture sustains peak load without database starvation or overselling."