Multi Tenancy — Apache Hbase
summary
Flipkart runs a consolidated multi‑tenant HBase cluster serving ~200 tenants and peak tested at 6M QPS with latencies under 50 ms. The solution pins region writes with favored nodes stored in region meta to force HDFS blocks and WALs onto tenant‑specific disks, adds new balancer cost functions and candidate generators to prevent skew and spillover, and implements tenant‑scoped replication endpoints for Kafka and Pulsar. An operator automates per‑tenant config, snapshots, and CDC so tenants get stronger isolation without separate clusters.
AI-written summary · about
chapters (10)
- 0:00Introduction to multi-tenancy
- 2:09HBase architectural overview
- 3:48Multi-tenant design goals
- 6:21The data isolation problem
- 8:54Implementing favored nodes
- 14:08Addressing skew and spillover
- 15:36Stochastic load balancing
- 19:43Solving spillover issues
- 23:05WAL and CDC isolation
- 24:10Replication and final summary
related talks
21mCase Study: FOSS (Apache Hbase) at Flipkart
49mHow Flipkart runs HBase across zones and clouds seamlessly
43mMulti-tenant HBase Solutions at Facebook
55mApache HBase: crazy dances on the elephant back
52mMultitenancy: cost-efficient and scalable patterns for Software-as-a-Service
28mDivide and Conquer: Multi-tenancy in Postgres
32mContaining an Elephant
59mHadoop Essentials for the SQL Server Professional