\Rainbows! is like unicorn, but Different…
While \Rainbows! depends on unicorn
for its process/socket management, HTTP parser and configuration language;
\Rainbows! is more ambitious.
Architectural Diagrams
unicorn uses a 1:1 mapping of processes to clients
unicorn master
\_ unicorn worker[0]
| \_ client[0]
\_ unicorn worker[1]
| \_ client[1]
\_ unicorn worker[2]
| \_ client[2]
...
\_ unicorn worker[M]
\_ client[M]
\Rainbows! uses a M:N mapping of processes to clients
rainbows master
\_ rainbows worker[0]
| \_ client[0,0]
| \_ client[0,1]
| \_ client[0,2]
| ...
| \_ client[0,N]
\_ rainbows worker[1]
| \_ client[1,0]
| \_ client[1,1]
| \_ client[1,2]
| \_ client[1,3]
| ...
| \_ client[1,N]
\_ rainbows worker[2]
| \_ client[2,0]
| \_ client[2,1]
| \_ client[2,2]
| ...
| \_ client[2,N]
...
\_ rainbows worker[M]
\_ client[M,0]
\_ client[M,1]
\_ client[M,2]
...
\_ client[M,N]
In both cases, workers share common listen sockets with the master and pull
connections off the listen queue only if the worker has resources
available.
Differences from unicorn
- log rotation is handled immediately in \Rainbows! whereas unicorn has the
luxury of delaying it until the current request is finished processing to
prevent log entries for one request to be split across files.
- load balancing between workers is imperfect, certain worker processes may
be servicing more requests than others so it is important to not set
worker_connections too high. unicorn worker processes can never be
servicing more than one request at once.
- speculative, non-blocking accept() is not used, this is to help load
balance between multiple worker processes.
- HTTP pipelining and keepalive may be used for GET and HEAD requests.
- Less heavily-tested and inherently more complex.
Similarities with unicorn
While some similarities are obvious (we depend on and subclass off unicorn
code), some things are not:
- Does not attempt to accept() connections when pre-configured limits are hit
(worker_connections). This will first help balance load to
different worker processes, and if your listen() +:backlog+ is overflowing:
to other machines in your cluster.
- Accepts the same signals for process
management, so you can share scripts to manage them (and nginx, too).
- supports per-process listeners, allowing an external load balancer like
haproxy or nginx to be used to balance between multiple worker processes.
- Exposes a streaming "rack.input" to the Rack application that
reads data off the socket as the application reads it (while retaining
rewindable semantics as required by Rack). This allows Rack-compliant
apps/middleware to implement things such as real-time upload progress
monitoring.