Siberite is a simple, lightweight, leveldb backed message queue written in Go.

Download .zip Download .tar.gz View on GitHub


License Build Release Go Walker

Siberite is a simple leveldb backed message queue server
(twitter/kestrel, wavii/darner rewritten in Go).

Siberite is a very simple message queue server. Unlike in-memory servers such as redis, Siberite is designed to handle queues much larger than what can be held in RAM. And unlike enterprise queue servers such as RabbitMQ, Siberite keeps all messages out of process, using goleveldb as a persistent storage.

The result is a durable queue server that uses a small amount of in-resident memory regardless of queue size.

Siberite is based on Robey Pointer's Kestrel - simple, distributed message queue. Like Kestrel, Siberite follows the "No talking! Shhh!" approach to distributed queues: A single Siberite server has a set of queues identified by name. Each queue is a strictly-ordered FIFO, and querying from a fleet of Siberite servers provides a loosely-ordered queue. Siberite also supports Kestrel's two-phase reliable fetch: if a client disconnects before confirming it handled a message, the message will be handed to the next client.

Compared to Kestrel and Darner, Siberite is easier to build, maintain and distribute. It uses an order of magnitude less memory compared to Kestrel, and has an ability to consume queue multiple times (using durable cursors feature).


  1. Siberite clients can consume single source queue multiple times using get <queue>.<cursor_name> syntax.

    • Usually, with get <queue> syntax, returned message gets expired and deleted from queue.
    • With cursor syntax get <queue>.<cursor_name>, a durable cursor gets initialized. It shifts forward with every read without deleting any messages in the source queue. Number of cursors per queue is not limited.
    • If you continue reads from the source queue with usual syntax again, siberite will continue deleting already serverd messages from the head of the queue. Any existing cursor that is internally points to an already expired message will start serving messages from the current queue head on the next read.
    • Durable cursors are also support two-phase reliable reads. All failed reliable reads for each cursor get stored in cursor's own small persistent queue and get served to other cursor readers.
  2. Fanout queues

    • Siberite allows you to insert new message into multiple queues at once by using the following syntax set <queue>+<another_queue>+<third_queue> ...


Siberite performance benchmarks


Make sure your GOPATH is correct

go get
cd $GOPATH/src/
go get ./...
go build siberite.go
mkdir ./data
./siberite -listen localhost:22133 -data ./data
2015/09/22 06:29:38 listening on
2015/09/22 06:29:38 initializing...
2015/09/22 06:29:38 data directory:  ./data

or download darwin-x86_64 or linux-x86_64 builds


Siberite follows the same protocol as Kestrel, which is the memcache TCP text protocol.

List of compatible clients

Telnet demo

telnet localhost 22133
Connected to localhost.
Escape character is '^]'.

set work 0 0 10

set work 0 0 2

get work
VALUE work 0 10

get work/open
VALUE work 0 2

get work/close

STAT uptime 47
STAT time 1443308758
STAT version siberite-0.4.1
STAT curr_connections 1
STAT total_connections 1
STAT cmd_get 2
STAT cmd_set 2
STAT queue_work_items 0
STAT queue_work_open_transactions 0

# other commands:
# get work/peek
# get work/open
# get work/close/open
# get work/abort
# get work.cursor_name
# get work.cursor_name/open
# get work.my_cursor/close/open
# set work+fanout_queue
# flush work
# delete work
# flush_all

Not supported

  • Waiting a given time limit for a new item to arrive /t= (allowed by protocol but does nothing)