Perplexity 开发并开源了 CobbleDB，这是一种专为高效批量读取设计的新型键值数据库，显著提高了其搜索工作负载的性能。这个新系统取代了 Amazon DynamoDB，将中位数批量读取延迟从 31.4 毫秒大幅降低到 5.60 毫秒，并将 p99 延迟从 123 毫秒降低到 24.2 毫秒。CobbleDB 的开发得到了 AI 代理团队的加速，核心基础设施由两名工程师和数百个此类代理在两个月内完成。 AI
影响 加速了 AI 搜索性能，并为开发人员提供了一个新工具，有可能提高 AI 相关应用程序的效率。
排序理由 Perplexity 发布了一项新的数据库技术并详细介绍了其开发过程，这是一个产品发布/发布。
AI 生成摘要 · Google Gemini · 来自 7 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
CobbleDB is purpose-built for repeated batch reads of prepared page records, optimizing performance and cost for Perplexity’s search workload. We plan to open-source CobbleDB for other teams building AI search. Read more: https://t.co/iK1dd8WDhR
After replacing DynamoDB, production before-and-after measurements show median batch-read latency falling from 31.4 to 5.60 ms and p99 from 123 to 24.2 ms. Our internal cost model estimates at least 20% savings relative to DynamoDB. https://t.co/Bs8c9vsRte
That separation lets each replica ingest updates at its own pace. Pillar stores versioned document state in YTsaurus and queues exports. Lorry writes partitioned S3 batches that CobbleDB replicas apply independently and in order, so recovering nodes don’t delay others. https://t…
Optimizing reads also required changing how updates reached the store. We separated durable document state, update delivery, and serving across Pillar, Lorry, and CobbleDB. https://t.co/pL7vCBTyC3
The architecture prioritizes fast batch reads. CobbleDB groups page keys by partition and reads them in parallel through RocksDB MultiGet. The router favors same-zone replicas and queries another copy when a replica is slow, so one delayed response need not hold up the batch. ht…
We accelerated CobbleDB development with a team of proactive coding agents. The agents provided continuous inspection and follow-through across sessions under the direction of engineers responsible for architecture, review, and production authorization.
We’re publishing research on how we built CobbleDB, our key-value database that serves web content for Perplexity search. Two engineers and a team of hundreds of proactive, always-on AI agents built the core infrastructure in two months. https://t.co/OsCtY15UW6