Facebook Configuration Management: Practices, Statistics, and Cultural Insights

This series translates and consolidates material from Facebook Research’s paper "Holistic Configuration Management at Facebook". It links to seven translated articles that cover the challenges of massive service configuration, the toolset used, authoring and error‑prevention techniques, reliable distribution, the GateKeeper rollout tool, the MobileConfig mobile‑side tool, engineering culture, and operational experiences.

Facebook’s engineering culture heavily shapes its configuration management. Early practices relied on selective diff review and manual testing; over time they evolved to mandatory diff reviews, automated canary testing, continuous integration, whitelist‑based access control, and automated verification tools such as VAR‑IOS.

The "configuration as code" hypothesis is examined through Configerator, which stores Python and Thrift source files, compiled JSON configurations, and raw configuration files. Approximately 89% of raw‑configuration changes are performed by automation, confirming that engineers prefer generating configuration code over manual edits.

As of 2015 Configerator holds hundreds of thousands of configuration items, with 75% being compiled configurations. The rapid growth is illustrated in the following chart:

Configuration size distribution shows that raw configurations have a median (P50) of 400 bytes and compiled configurations 1 KB, while the 95th percentiles are 25 KB and 45 KB respectively; the largest raw and compiled configurations reach 8.4 MB and 14.8 MB. This distribution is shown below:

During their lifecycles, raw, compiled, and source configurations are updated on average 44, 16, and 10 times respectively. Raw configurations are updated 175% more often than compiled ones because most (89%) raw changes are automated.

Change size analysis reveals that roughly 50% of configuration updates modify only one or two lines of code, while about 8.7% involve changes larger than 100 lines. The distribution of change sizes is illustrated here:

Collaboration statistics show that most configurations are touched by few authors: 79.6% of compiled configurations and 91.5% of raw configurations are modified by one or two contributors. A few outliers have many contributors—for example, a single site‑var was edited by 727 authors over two years. This author‑distribution is visualized below:

The data demonstrates Facebook’s mature, data‑driven approach to configuration management, providing detailed insight into scale, automation, and engineering practices. The article notes that a forthcoming piece will discuss performance statistics of Configurator and GateKeeper and share operational experiences.