Improving Facebook's performance on Android with FlatBuffers

Facebook describes how it stores social graph data on Android as a tree structure rather than normalizing it into tables. JSON was too slow to parse, too costly to initialize, and too allocation-heavy, so the app needed a better storage format.

Reading notes

#

• FlatBuffers lets the app read individual subcomponents of a tree directly, without deserializing the whole object first.
• Data can be stored directly in a file and mmap-ed into memory, which lowers memory footprint and avoids intermediate object creation.
• Because FlatBuffers are immutable, the team tracks mutations alongside the base buffer so small updates can be applied without redownloading an entire story.
• The mutation buffer keeps an index from absolute positions in the base buffer to new data, plus the replacement data in FlatBuffer format.
• FlatBuffers are used for storage, networking, and the in-memory app format, which removes transformation steps between server response and UI display.
• The Flat Models approach moves application and networking logic below storage, leaving the UI to read from storage through standard Android cursors.
• The article reports reductions in story load time from 35 ms to 4 ms, transient memory allocations by 75 percent, cold start time by 10 to 15 percent, and storage size by 15 percent.