Lost messages with sidekiq

January 28, 2018

Sidekiq is a great library for getting asynchronous library for quickly moving work into background jobs in Rails/Ruby apps, but it can drop messages. Understanding of when this can happen will hopefully help prevent you from putting critical ‘cannot-lose-this’ messages somewhere they can be lost or cut down debugging time.

The general API for Sidekiq looks like this, UserMailer.perform_async enqueues a job to Redis and that eventually runs then UserMailer.new.perform method:

class User < ActiveRecord::Base
  after_commit :greet, :on => :create

  def greet

class UserMailer
  include Sidekiq::Worker

  def perform(id)
    contact = Contact.find(id)


  def send_email(contact)
    # email code

Sidekiq ‘best practices’ for rails generally advise firing events in an after_commit hook to prevent race conditions. The race being, if the sidekiq job fires before the database transaction commits, the worker will error when trying to find the not-yet-commited database record. This is an important detail that we will get back to in a second.

Where messages can be dropped:

1) Redis as a datastore

If Redis restarts after enqueueing a job, messages are lost

Redis is incredibly fast and one of the reasons sidekiq itself is so quick, but the drawback to speed in this case is data durability. Using the default options When writing a value to Redis, Redis will confirm the write without it having been persisted to disk. Redis default is to only persist in memory - which is greate for speed. There is an option to turn on an Append Only File, where redis persists data to disk with fsync, but by default the write to disk happens once every second appendfsync everysec, the implications of this is that there is still a window where your message only exists in memory and a Redis restart will lose your message!

It is possible to toggle appendfsync always which writes to disk after every command. The speed and throughputs of writes will drop drastically, so you probably want to use a diffferent Redis instance for speed sensitive operations like caching.

2) After Commit

If your app or Redis restarts while enqueuing a job, messages are lost

The recommendation for using after_commit touches on the issues of distributed transactions. What our after_commit code is doing:

  1. Save record to primary database, then
  2. Save data to Redis

The semantics of what we want is our data saved in two places in an ACID-style transaction, but our code is doing this in two separate operations. After the first operation finishes, the intent to save data in Redis exists only in memory, if our app restarts during that time - the message is lost.

For durability, a potentially implementation would persist the intent to publish to Redis inside the first database transaction (e.g. after_save) and cleanup on job run. With that, should the application restart it can check for in-flight messages and continue where it left off.

There is an post I wrote on the Outbox Pattern for ways to impliment this.

3) Reliable Push

If your app restarts while enqueuing a job, messages are lost

Reliable Push is a pro feature designed to add durability around publishing while a Redis connection is interrupted or unavailable. If the connection is unavailable, Sidekiq will store the message in memory and retry when the connection becomes available again.

The issues here are again around your application restarting while enqueueing a job. When the job is being held in memory an application restart will drops it. This feature reduces the chance of a dropped message, but architecturally there is still a gap in durability guarantees.

4) Worker restarts

If your worker restarts while processing work, messages are lost

This one differs a little depending on which version of Sidekiq you are using.

In the public version: sidekiq workers uses BRPOP to fetch the latest job from Redis:

def retrieve_work
  work = Sidekiq.redis { |conn| conn.brpop(*queues_cmd) }
  UnitOfWork.new(*work) if work

As soon as this happens, the job only exists in memory of that worker, a worker restart will drop the message

In Sidekiq Pro version < 3: Sidekiq workers use the atomic BRPOPLPUSH command to push in-process work into a ‘working’ queue, one for each worker. If you ever wondered what that -i index flag is used for in the command line, this is roughly the code block it gets used in:

def working_queue_name(q)
  if options[:ephemeral_hostname]

This constructs the name of the ‘working’ queue that the sidekiq work lives in. When a worker restarts it checks its ‘working’ queue to see if there were any in-flight jobs is needs to resume before fetching new work. The interesting part here is the Socket.gethostname call. If you are using docker, this will be different on each worker restart. The implication to this is if options[:ephemeral_hostname] returns false, jobs will become permanently orphaned in ‘working’ queues - essentially lost. E.g. placing in-flight jobs into a working queue named app:queue:default_9990eeababc0_0, and on restart incorrectly looking at a new working queue named app:queue:default_285c9f0f9d3d_0 for those flight jobs.

Sidekiq Pro version >= 4 has job recovery feature that periodically looks for orphaned jobs and recovers them.

Final thoughts

One of the very real questions to all of this is, should I care? Depending on your requirements maybe not.

For example, one on end if you are sending welcome emails, dropping a few here or there might have a very small business impact compared to the value of building another feature. But on the other end, if you find yourself using Sidekiq for mission critical jobs, data replication, or some other ‘cannot-lose-this’ function, then this might become an issue.

Sidekiq itself has a price chart for making cost tradeoffs (although it is for upgrading to pro), but the concept is essentially the same tradeoff here. The biggest advantage to Sidekiq is how little developer time it takes to get features out the door.