Twitter Timeline Service
VERIFIED AT SCALE · 2026“I wanted to understand why home timelines become difficult at scale, so I built one from scratch.”
Most tutorials explain fan-out on write and call it a day. In reality, accounts with 10M followers melt Redis if you blindly fan-out. I implemented fan-out-on-write alongside fan-out-on-read, benchmarked them head-to-head under synthetic follower distributions, and built a hybrid cutoff model that merges precomputed caches with live queries at read time.
LAB INSERT // SYNTHETIC LOAD & REDIS SIZING TELEMETRYUnfold log ▾
Synthetic Ingest Peak50,000 req/sKafka queue push burst
Naive Fan-Out Redis Size14.2 GBBlowup on 100k handles
Hybrid Model Footprint1.1 GB (−92.2%)Cutoff at 25k threshold
p99 Read Latency4.8 msMerged timeline assembly
✎ Under synthetic Zipfian distributions, 0.2% of accounts produced 68% of fan-out writes. Isolating the top tier directly protected Redis memory watermarks.
