aasheeshlikepanner/vase

Vase is a publish/subscribe message broker written in Go. A pub/sub broker is the middle piece in a system where one program (the publisher) wants to send the same message to many other programs (the subscribers) at the same time, for things like live updates or chat fan-out. The README pitches Vase as a faster alternative to Redis Pub/Sub for cases with many subscribers, claiming 62 times the throughput at 100 subscribers in the author's benchmark. The author starts by explaining why Redis slows down. Redis Pub/Sub runs in a single thread that loops over every subscriber on each publish call, so throughput scales as O(n) in the subscriber count. Their measurements show Redis dropping from 38,000 messages per second with one subscriber to 1,800 with one hundred. Vase takes a different approach: the publisher writes a message once into a shared ring buffer, and each subscriber holds its own cursor and reads from the buffer independently. That makes the publish step O(1) regardless of how many readers are listening. The ring buffer is lock-free. A publisher does one atomic increment to claim a slot, writes the message, and marks the slot ready with a sequence number. Each subscriber compares its cursor against that sequence number to know when a new message is available. The buffer uses power-of-two capacity so indexing can use a bitwise AND instead of a modulo, and each cursor is padded to a full 64-byte cache line so that one subscriber's writes do not invalidate another's cached state. For durability, Vase has an optional write-ahead log. Every message is appended to a memory-mapped 128MB segment file with a CRC32 checksum before it is treated as published. On restart, the log is scanned, checksums are verified, and playback resumes from the last valid offset. The broker speaks a small custom binary protocol over TCP, one goroutine per connection, with batching of up to 64 messages per flush. The README is upfront about limits: single node only, no replication, slow subscribers get disconnected, and end-to-end publish rate still touches the network serialization layer.