Services/MessageQueuing

Planning Questions

What is the simplest MQ system and features to add from there?

• Put a single message on a queue for later retrieval in a FIFO manner, non-persistent, via a RESTful service to push and poll for messages (send XX many at once per client instruction?)
• Transient messages (Semi-durable, higher throughput at the possible loss of some recent messages in a crash)
• Persistent messages (Accessible until consumed)
• Prioritizing of messages? Or allow for multiple queues and have the client prioritize where it pulls messages from?
• Allow messages to go unacknowledged to allow a client to 'time-out' on handling it?
• Message push (persistent MQ connection), reduces latency but increases complexity. Pub/Sub or load-balanced on multiple connected clients?
• Chat and 'live' apps using the MQ for comet?

Trade-offs of implementation / back-ends

• Non-persistent allows use of memcached for higher throughput
• Semi-durable allows backends that do occasional fsync's (Mongo / Redis)
• Fully persistent (i.e., backed by a db, RabbitMQ persist) would be much slower

Or should an App creator get to choose which characteristic they want when creating a queue? i.e. someone making a real-time chat/broadcast system could choose non-persistent for lots of messages of little importance if they're missed... someone sending verification emails would choose fully persistent to ensure it actually gets sent eventually

Crypto

• Payload could be encrypted at the client layer
• If routing keys are used (RabbitMQ style brokering), those could be encrypted as well
• Crypto less useful unless other tasks needed to hide payloads, which should likely be done on client-side

Apps should be able to create both durable and temporary queues on-demand, as well as assign a TTL to a queue for temporary use.

Authentication

• maybe AppKeys would work, depending on if only the App could send messages... might need something else if end-user clients can directly push messages via Javascript
• If used for message passing between end-user's vs. a separate Service app, clients would need permission to be on the queue, likely based on whether they're 'logged in' to the App using the queue

Message Handling

• Should the MQ Service do anything beyond merely holding messages and allowing connections for their retrieval?
• Simplest has no built-in handling, app would need to poll MQ service for new messages
• Allow a message payload to contain a callback to a RESTful service (URL + POST payload)?
• Message TTL?

Queues

• Single default queue per App seems reasonable
• Multiple queues per App? Routing of messages to multiple queues based on route key like in RabbitMQ?
• Temporary queues for return-paths? (i.e., using a queue to farm out something then waiting for it to complete and return data)

Queue Workers?

Simplest: No queue workers, service calls back into special App URL Drawback: Can't determine as easily if the Apps are busy processing 'task' jobs or end-user requests

• Introduces additional challenge of requiring Apps to have the code the worker should call for a message accessible to the MQ Worker Service
• Feels like a separate service from the MQ Service (though highly related)
• Requires worker pools on machines, worker watchers, resource monitoring, worker killers for excessive resource usage, deploying additional machines as App tasks require more task processing power under load
• Provide 'built-in' queue workers for delayed common tasks? i.e. sending email, update caches.
• Handling 'excessive resource usage' requires consistent way to determine CPU usage per task regardless of existing machine utilization, kill workers that use more than XXX RSS memory, etc.

Introspection / Administration

• How much should one be able to see about current queue activity? Total queues per app? Total messages in each queue? Ability to examine payload of individual message?
• Most basic feature-set should include some way for an App to clear/reset one of its queues