2025
Hyperlinkr
Rust · Axum · DragonflyDB · Docker
A URL shortener treated as a systems problem.
~50K RPS on 8 GB hardware>95% cache hit rate
Hyperlinkr layers four caches between a request and the database: NUMA-aware L1/L2 with TinyLFU admission, a Bloom filter to swallow misses, and DragonflyDB behind them — together holding a >95% hit rate at ~50K RPS on an 8th-gen i3 with 8 GB of RAM.
The performance work is diagnostic, not decorative: flamegraph profiling identified BASE62 code generation as the real bottleneck, and atomic contention in the Bloom filter surfaced under concurrency. IP-based rate limiting, device- and geo-aware analytics, and a circuit breaker on the DragonflyDB layer round out the production posture.
Key decisions
- 4-layer cache: NUMA-aware L1/L2 + TinyLFU + Bloom filter + DragonflyDB, >95% hit rate
- Flamegraph-profiled — BASE62 codegen identified as the true bottleneck
- Circuit-breaker failover on the cache layer, IP-based rate limiting
- Real-time analytics with device detection and geo-location