LISA18

Ten years ago Facebook released the first version of Scribe which has formed the basis for our logging infrastructure ever since. Back then logging was mostly about aggregating files full of text but since then “logs” have evolved to include abstract streams of events and blurred the line with messaging and pub/sub systems. Scribe continues to aggregate massive amounts of log data but now also forms the basis for many real time stream processing applications at Facebook. Having a common transport layer allows us to apply new stream processing and data science tools to business or operational data with equal ease.


I'll look back at some of the early design choices that have allowed Scribe to evolve with our needs. This evolution has required us to replace major components and scale up the infrastructure by several orders of magnitude. I'll walk through the steps in this evolution with particular emphasis on our most recent upgrade: replacing the HDFS-based persistence layer with LogDevice and how we accomplished that upgrade with little user impact.