SREcon16 Europe

The mission of the Data Warehouse Storage team at Facebook is to run an HDFS deployment that stores hundreds of petabytes reliably and efficiently.

Due to its stateful nature, there are some unique challenges to operating a storage system. How do we take a machine out of production for repair without compromising data availability? What trade-offs do we make between replication strategy and data availability to make sure we get more bang for the buck? How do we ensure that colocated tasks that run on our storage nodes exploit available resources as much as they can without tipping hosts over?

In this talk, we'll share some of the lessons that we have learnt from running HDFS at Facebook. We will discuss our biggest operational challenges and we'll outline the evolution of the different solutions that we put in have in place over time. We will also introduce Warm Storage, a novel block storage system that we built at Facebook to replace HDFS, and we'll discuss how our learnings from HDFS have affected the design of Warm Storage.