taskmaster is a simple distributed queue designed for handling large numbers of one-off tasks.
We built this at DISQUS to handle frequent, but uncommon tasks like "migrate this data to a new schema".
You might ask, "Why not use Celery?". Well the answer is simply that normal queueing requires (not literally, but it'd be painful without) you to buffer all tasks into a central location. This becomes a problem when you have a large amount of tasks, especially when they contain a large amount of data.
Imagine you have 1 billion tasks, each weighing in at 5k. Thats, uncompressed, at minimum 4 terabytes of storage required just to keep that around, and gains you very little.
Taskmaster on the other hand is designed to take a resumable iterator, and only pull in a maximum number of jobs at a time (using standard Python Queue's). This ensures a consistent memory pattern that can scale linearly.
Create an iterator, and callback:
# We must ensure default timeout **is not set** or random shit will hit the fan.
# last would be sent if state was resumed
# from a previous run
for i in xrange(last, 100000000):
# jobs yielded must be serializeable with pickle
# this **must** be idempotent, as resuming the process may execute a job
# that had already been run
print "Got %r!" % i
Spawn a master:
Spawn a slave:
Or spawn 8 slaves (each containing a threadpool):
tm-spawn taskmaster.example 8
Dont like the magical function discover for master/slave? Specify your own targets:
tm-master taskmaster.example:get_jobs $ tm-slave taskmaster.example:handle_job
Note: All arguments are optional, and will default to localhost with no auth key.