Multi-tenant HBase Solutions at Facebook
summary
Facebook runs HBase for multiple production systems including Messages which handles 180 billion messages per month. Infrastructure details: HBase on HDFS with ZooKeeper and BookKeeper experiments, Haystack for large attachments, MapReduce for schema upgrades and rollups, memcache for hot reads, and MD5-based sharding in a controlled multi-tenant layer called hash out. Live footprint claims ~90 billion read/write ops for messages, 4.5 PB compressed storage (≈15 PB replicated), ~400 TB growth per month, and clusters from ~16 to ~200 nodes.
AI-written summary · about
chapters (19)
- 0:00Start
- 0:41Multi-Tenant Hbase Solutions
- 2:50Hbase Overview
- 5:46Horizontal Scalability
- 6:29Summary of the Right Path and Read Path
- 7:25System Overview
- 7:49What Sort of Use Cases Do We Have in Production
- 8:03Facebook Insights
- 8:57Facebook Messages
- 9:55Stats
- 22:47Pre-Production
- 23:21Shadow Testing
- 27:08Do You Need a New Column Family or a New Table
- 28:06Real-Time Facebook Insights
- 32:10An Operational Data Store
- 32:53Schema Design
- 34:09Lessons Learned
- 35:44Multi-Tenant Used Cases
- 39:43Constant Upgrades
related talks
29mMulti Tenancy — Apache Hbase
51mBIG DATA STORAGE & PROCESSING: Introduction to Hadoop and Its Ecosystem
01h 01mPanel: Facebook's Distributed Database Evolution
55mApache HBase: crazy dances on the elephant back
21mCase Study: FOSS (Apache Hbase) at Flipkart
23mScaling Memcache
41m"A Day in the Life of Facebook Operations"
09mHigh Order Bit: The In
