Skip to content

Matrix performance problems

Noted by on his . Last updated . Changelog

Synapse is incredibly slow, which is why I run the Conduit matrix server. Server performance is the main price paid for Matrix’ history replication.

This also gives Matrix a spam DDoS problem: sometimes people mass-create accounts on hundreds of servers with open registration, and mass-join rooms. The flood of thousands of join-events will OOM most connected servers. My Conduit server (RocksDB backend, 1 vCPU, 1GB RAM) is actually fast and light enough to ride some of these out and go back to normal speeds a couple hours after the floods die down.

You generally don’t want to join a room like Matrix-HQ unless you have good hardware. A new Postgres backend is coming to Conduit soon, which should improve the situation.

Once Conduit implements a few missing features (federated backfill, spaces hierarchy and room discovery), and once better spam controls roll out (e.g. automatically restricting joins to under N per second and/or automatically blocking open-registration servers during high join frequencies), I’d say the worst performance problems will be gone.