distributed-task

Node Distributed Task

distributed-task provides a framework upon which one can easily build distributed task systems in Node. The original reason for building a new distributed task system was to allow 'watching' of tasks, specifically long/indefinitely running tasks. distributed-task consistes of a few features:

• Redis as a global state store
• distributor
  • handle the distribution of tasks
  • read incoming tasks from a Redis queue
  • monitors its own set of tasks
  • monitor all tasks on longer interval
• worker
  • has tasks as Node functions w/JSON config
  • does the actual task work
  • connects to the distributors
  • updates task info in Redis
  • listens for watcher connections
• watcher
  • talks direct to workers and can subscribe/unsubscribe to task live related events
  • gets task/worker info from Redis

Synopsis

Config.json

{
    "share_key": "abc",
    "key": "test/tls/key.pem",
    "cert": "test/tls/certificate.pem",
    "redis": {
        "host": "localhost",
        "port": 6379
    },
    "host": "localhost",
    "port": 5000,
    "distributors": [
        ["localhost", 5000]
    ]
}

distributor.js

#!/usr/bin/env node
 
var distributor = require('../distributed-task/distributor');
distributor.init(require('./config.json'));

worker.js

#!/usr/bin/env node
 
var worker = require('../distributed-task/worker');
worker.init(require('./config.json'));
worker.addTask('name', someFunction);

Full Example

A full example (almost deploy ready) can be found here.

Definitions

Task

{
    clean_end: true,      // allow distributors to clean up the task when it ends (& loose end status)
    state_optional: true; // allow the task to continue running when Redis goes down
                          // shouldn't happen but no guarantee the task won't also be requeued w/ multiple distributors-over-WAN
                          // it basically assumes all distributors share the same connection to Redis
}

Task Status

• RUNNING - task is running, should have updated timestamp
• STOPPED - task stopped intentionally by distributor/worker loosing related worker/distributor connection
• END - task is complete, pending cleanup

Redis Bits

• new-task - simple list based queue
• tasks - set of all task_id's
• task-<uuid> - individual task details

Failover/HA

• A shorter loop is recommended for workers than distributors
• All distributors/workers must have the same timezone set
• Distributors/workers send health checks to each other according to a configured time interval
• Both distributor and worker acknowledge disconnect on stream close + after missed health check
  • distributor flags worker as down and clears after interval, will requeue tasks
  • worker stops running tasks for this distributor (which should requeue them)
• Upon disconnect from Redis: workers, for each connected distributor:
  • notify the distributor, which stops sending new task requests
  • pause running distributor tasks
  • check if we're connected to the distributor
  • if so, assumes distributor has also lost Redis connection & won't requeue
  • continue distributor tasks w/o state_optional = true
  • stop all other tasks from this distributor
• Distributors watch for tasks w/ state STOPPED & RUNNING w/out of date timestamp, requeues
• Same as above for END tasks with clean_end = true
• Distributors watch their own tasks on a short interval, and all tasks on a long interval
• Upon reconnect to Redis, workers check state_optional task timestamps before updating, and stop if newer one