Qdrant(@qdrant_engine)

How do you run vector search at 20B+ vectors across 5 regions for 38 teams? @HubSpot built VAST - V...

8.5内容质量
How do you run vector search at 20B+ vectors across 5 regions for 38 teams? @HubSpot built VAST - V...

TL;DR · AI 摘要

HubSpot通过自研Kubernetes操作符优化Qdrant集群管理，实现200亿向量跨5大区的高效搜索服务。

核心要点

  • Qdrant集群规模达150个，单集合存储95亿向量，写入峰值达10万/秒
  • 自研Kubernetes操作符将集群创建时间从小时级缩短至分钟级
  • 资源分配不均衡降低65%，BM42稀疏向量集合优化效果显著

结构提纲

按章节快速跳转。

  1. §VAST架构概述

    介绍HubSpot基于Qdrant构建的Vector as a Service系统规模与性能指标

  2. ·技术挑战

    Helm无法处理Qdrant集群的动态扩展与状态感知需求

  3. ›解决方案

    自研Kubernetes操作符实现自动分片管理与资源调度

  4. ·优化成果

    集群创建时间缩短至分钟级，资源分配不均衡降低65%

思维导图

用一张图看清主题之间的关系。

查看大纲文本（无障碍 / 无 JS 友好）
  • VAST架构与优化
    • 系统规模
      • 200B+向量存储
      • 5大区部署
    • 技术挑战
      • Helm扩展限制
      • 状态感知需求
    • 优化方案
      • Kubernetes操作符
      • 自动分片管理

金句 / Highlights

值得收藏与分享的关键句。

#Qdrant#Kubernetes#向量搜索#HubSpot#VAST
打开原文

Qdrant on X: "How do you run vector search at 20B+ vectors across 5 regions for 38 teams? @HubSpot built VAST - Vector as a Service, entirely on Qdrant. 150 clusters, 2K+ pods, 9.5B vectors in a single collection, 5K writes/sec with spikes to 100K. the real story: they outgrew Helm fast. Helm can't call the Qdrant API to transfer shards, maintain replication factor, or handle state-aware scaling. cluster creation took hours. so they built a Kubernetes operator specifically for Qdrant. shard management, replication, lifecycle automation, all handled automatically. cluster spin-up: hours → minutes. result: 65% reduction in resource skew on a 3B+ point BM42 sparse vector collection. full talk here: https://t.co/rWDfiq2n1s thanks Oleg Tereshin and Xin Liu from @HubSpot team, for sharing at Vector Space Day SF 🙌" / X

Qdrant

@qdrant_engine

How do you run vector search at 20B+ vectors across 5 regions for 38 teams?

@

HubSpot

built VAST - Vector as a Service, entirely on Qdrant. 150 clusters, 2K+ pods, 9.5B vectors in a single collection, 5K writes/sec with spikes to 100K. the real story: they outgrew Helm fast. Helm can't call the Qdrant API to transfer shards, maintain replication factor, or handle state-aware scaling. cluster creation took hours. so they built a Kubernetes operator specifically for Qdrant. shard management, replication, lifecycle automation, all handled automatically. cluster spin-up: hours → minutes. result: 65% reduction in resource skew on a 3B+ point BM42 sparse vector collection. full talk here:

youtube.com/watch?v=46aQff…

thanks Oleg Tereshin and Xin Liu from

team, for sharing at Vector Space Day SF 🙌

4:00 PM · Jul 20, 2026

725

Views

3

6

17