At my work, we use ruby heavily and sidekiq is an essential part of our stack. Sometimes, I long for the concurrency primitives from Elixir, but that's not what today's post is about.
A few days ago, I caused a minor incident by overloading our databases. Having been away from ruby for a bit, I had forgotten that sidekiq runs multiple threads per each worker instance. So, I ended up enqueuing about 10K jobs on Sidekiq, and Sidekiq started executing them immediately. We have 50 worker instances and run Sidekiq with a concurrency of 20. So, essentially we had 400 worker threads ready to start crunching these jobs. Coincidentally we have 400 database connections available and my batch background job ended up consuming all the connections for 5 minutes during which the other parts of the application were connection starved and started throwing errors 😬.
That was a dumb mistake. Whenever you find yourself making a dumb mistake, make sure that no one else can repeat that mistake. To fix that, we could set up our database with multiple users in such a way that the web app would connect with a user which could only open a maximum of 100 connections, the background worker with a user with its own limits and, so on. This would stop these kinds of problems from happening again. However, we'll get there when we get there, as this would require infrastructure changes.
I had another batch job lined up which had to process millions of rows in a similar fashion. And, I started looking for solutions. A few solutions that were suggested were to run these jobs on a single worker or a small set of workers, you can do this by having a custom queue for this job and executing a separate sidekiq instance just for this one queue. However, that would require some infrastructure work. So, I started looking at other options.
I thought that redis might have something to help us here, and it did! So, redis allows you to make blocking pops from a list using the BLPOP
function. So, if you run BLPOP myjob 10
, it will pop the first available element in the list, However, if the list is empty, it will block for 10 seconds during which if an element is inserted, it will pop it and return its value. Using this knowledge, I thought we could control the enqueuing based on the elements in the list. The idea is simple.
- Before the background job starts, I would seed this list with
n
elements wheren
is the desired concurrency. So, if I seed this list with2
elements, Sidekiq would execute only 2 jobs at any point in time, regardless of the number of worker instances/concurrency of sidekiq workers. - The way this is enforced is by the enqueue function using a
BLPOP
before it enqueues, so, as soon as the enqueuer starts, it pops the first 2 elements from the redis list and enqueues 2 jobs. At this point, the enqueuer is stuck till we add more elements to the list. - That's where the background jobs come into play, at the end of each background job, we add one element back to the list using
LPUSH
and as soon as an element is added the enqueuer which is blocked atBLPOP
pops this element and enqueues another job. This goes on till all your background jobs are enqueued, all the while making sure that there are never more than 2 jobs at any given time.
Let's put this into concrete ruby code.
module ControlledConcurrency
# I love module_function
module_function
# The name of our list needs to be constant per worker type, you could
# probably extract this into a Sidekiq middleware with a little effort
LIST_NAME = "migrate"
def setup(concurrency:)
# if our list already has elements before we start, our concurrency will be
# screwed, so, this is a safety check!
slot_count = Redis.current.llen(LIST_NAME)
raise "Key '#{LIST_NAME}' is being used, it already has #{slot_count} slots" if slot_count > 0
# Seed our list with as many items as the concurrency, the contents of this
# list don't matter.
Redis.current.lpush(LIST_NAME, concurrency.times.to_a)
end
# A helper function to bump up concurrency if you need to
def increase_concurrency(n = 1)
Redis.current.lpush(LIST_NAME, n.times.to_a)
end
# A helper function to bump the concurrency down if you need to
def decrease_concurrency(n = 1)
n.times do
puts "> waiting"
Redis.current.blpop(LIST_NAME)
puts "> decrease by 1"
end
end
# This is our core enqueuer, it runs in a loop because our blpop might get a
# timeout and return nil, we keep trying till it returns a value
def nq(&block)
loop do
puts "> waiting to enqueue"
slot = Redis.current.blpop(LIST_NAME)
if slot
puts "> found slot #{slot}"
yield
return
end
end
end
# Function which allow background workers to signal that a job has been
# completed, so that the enqueuer can nq more jobs.
def return_slot
puts "> returning slot"
Redis.current.lpush(LIST_NAME, 1)
end
end
# This is our Sidekiq worker
class HardWorker
include Sidekiq::Worker
# Our set up doesn't enforce concurrency across retries, if you want this,
# you'll probably have to tweak the code a little more :)
sidekiq_options retry: false
# the only custom code here is in the ensure block
def perform(user_id)
puts "> start: #{user_id}"
# mock work
sleep 1
puts "> finish: #{user_id}"
ensure
# make sure that we return this slot at the end of the background job, so
# that the next job can be enqueued. This doesn't handle retries because of
# failures, we disabled retries for our job, but if you have them enabled,
# you might end up having more jobs than the set concurrency because of
# retried jobs.
ControlledConcurrency.return_slot
end
end
# ./concurrency_setter.rb
ControlledConcurrency.setup(concurrency: ARGV.first.to_i)
# ./enqueuer.rb
# Before running the enqueuer, we need to set up the concurrency using the above script
# This our enqueuer and it makes sure that the block passed to
# ControlledConcurrency.nq doesn't enqueue more jobs that our concurrency
# setting.
100.times do |i|
ControlledConcurrency.nq do
puts "> enqueuing user_id: #{i}"
HardWorker.perform_async(i)
end
end
That's all folks! Hope you find this useful!
The full code for this can be found at: https://github.com/minhajuddin/sidekiq-controlled-concurrency
Top comments (0)