Perplexity Unveils CobbleDB, Built by 2 Engineers and Hundreds of AI Agents in 2 Months

Perplexity published a research report on CobbleDB, its in-house key-value database. It powers its search engine's fast fetching and reading of pre-prepared web content, replacing the previous AWS DynamoDB. The core infrastructure was built in two months by 2 engineers leading hundreds of proactive, always-on AI coding agents.

Confirmed

• Development model: hundreds of proactive coding agents work under engineer direction, providing continuous checks and follow-ups across sessions; engineers handle architecture design, code review, and release authorization—a "agents do the work, humans keep the gate" division of labor
• Measured performance: production data after replacing DynamoDB shows batch-read median latency dropped from 31.4ms to 5.60ms and p99 latency from 123ms to 24.2ms; internal cost models estimate at least 20% savings versus DynamoDB
• Read architecture: reads prioritize fast batch reads, grouping page keys by partition and reading in parallel via RocksDB MultiGet; the router prefers same-availability-zone replicas and switches to another replica when one responds slowly, preventing a single slow response from stalling the whole batch
• Write path: to optimize the read path, CobbleDB separates persistent document state, update delivery, and read serving into three components—Pillar, Lorry, and CobbleDB. Pillar stores versioned document state in YTsaurus and enqueues exports, Lorry handles update delivery, and replicas each consume updates in order, complementing the parallel batch-read design

Why it matters

CobbleDB demonstrates a new human-AI collaboration paradigm: a tiny number of engineers, through architectural oversight and review authorization, can direct large-scale AI coding agents to build core infrastructure quickly, validating AI agents' feasibility for complex systems engineering. Meanwhile, the significant latency and cost gains from replacing DynamoDB offer an engineering reference for high-throughput read-heavy workloads.

2026-09-16 ~ 2026-09-16 · 8 related posts

• [source] Perplexity: two engineers plus hundreds of AI agents built CobbleDB in two months — perplexity_ai · 2026-09-16
• How Perplexity ran coding agents: continuous cross-session follow-through, engineers gate production — perplexity_ai · 2026-09-16
• Inside CobbleDB: partitioned RocksDB MultiGet reads with same-zone-first routing — perplexity_ai · 2026-09-16
• Why CobbleDB split state, delivery, and serving: optimizing reads required rewiring updates — perplexity_ai · 2026-09-16
• [source] CobbleDB's write path: state, delivery, and serving split across Pillar, Lorry, and CobbleDB — perplexity_ai · 2026-09-16
• CobbleDB vs DynamoDB: median batch-read latency 31.4ms→5.6ms, 20%+ cost savings — perplexity_ai · 2026-09-16
• [source] Perplexity's CobbleDB cuts p99 latency from 123ms to 24.2ms, replacing DynamoDB — perplexity_ai · 2026-09-16
• Perplexity: two engineers and hundreds of AI agents built CobbleDB, replacing DynamoDB to save ~$100M/year — cameronstow · 2026-09-16