Recently we helped a client overcome widespread errors in Sidekiq. Between all of the deadlocks, threading errors and the need for widespread rate limiting, they were crippled and unable to scale. Some background jobs could only run one at a time but would take 30 seconds or more each. Other jobs ran quickly, but subsequently heaped more jobs onto the slow queues. The queues kept on growing, but dialing up the concurrency simply caused more problems.
On most days they could only let Sidekiq run for a few hours at a time before the system would break down into constant errors. Some of errors compounded so badly it ground the web server to a halt, taking down the entire site. In an attempt at exerting control, all onboarding jobs were started manually rather than automatically. There was no trust in the system, and no obvious way to scale their way out of the situation.
After brief investigation it was clear that there were fundamental problems with the way jobs were enqueued, distributed, and limited. What follows are the high level changes that were implemented, each presented as an observation and a solution.
Entangle Jobs Between Queues
Various categories of jobs were bucketed into different queues, but the queues weren’t weighted. This forced each queue to drain completely before jobs in the next queue were started. Queues are processed in the order they are listed, so piling thousands of slow running jobs onto a queue in the middle guarantees inefficient processing. A mass of the same slow job running simultaneously just exacerbates resource contention and prevents jobs in the subsequent queues from ever starting.
Declare equal weights for each of the queues so that jobs are plucked randomly between them. That forces fast jobs from the default queue to run alongside slow jobs. Spreading busy jobs between queues becomes even more important when there are limits on how many of each job type can run concurrently.
Force random queue priorities:
sidekiq -q google,1 -q facebook,1 -q linkedin,1 -q twitter,1 -q default,1
Learn more about advanced queuing in the Sidekiq Wiki.
Scale With Processes and Threads
The application’s database drivers for Neo4j used HTTPS as a transport.
Instead of connection pooling, it executed every request across a single
NetHTTPPersistent connection. When concurrent jobs deadlocked or
threw errors the thread stopped waiting for a response but kept holding the
connection. With one or two errors in quick succession the connection would
recover, but with rapid fire errors the connection problems quickly choked out
The ultimate solution would be to use a more efficient transport than HTTP, and to institute connection pooling. However, that would require upstream library changes and would take far longer than the client had. The temporary fix was to lean on processes for concurrency rather than threads. With fewer threads per process there was less of a drain on the single connection and it was easy to keep it healthy.
On a platform like Heroku it is simply a matter of scaling the number of worker
instances. On a host that uses Upstart to manage Sidekiq workers it is as simple
NUM_WORKERS within the workers conf. Even better, with
Enterprise you can make use of managed multi process using swarm.
env COUNT=4 exec bundle exec sidekiqswarm -e production
Use Concurrency Throttling
Many of the same type of job ran concurrently and contended for the same database or network resources. In an attempt to prevent clobbering the output of each job all of the records were being pessimistically locked, leading to database deadlocks. Some deadlocks would resolve themselves, but any that didn’t slowed the queue to a crawl, caused unpredictable errors, and put an extra burden on the throttling constructs.
Impose limits on the number of concurrent jobs that can run at once. With a single process and strictly segregated jobs it is trivial to control the concurrency by limiting a queue. But what happens when there is more work than a single process can handle (or a poorly behaving database driver forces you to parallelize with processes)? Now there are as many queues as there are processes, and any limits that were being enforced has scaled right along with them.
The proper solution is to use a distributed concurrency construct like the throttling available from Sidekiq Enterprise. The throttle enforces job concurrency limits across all threads, processes, and hosts; ensuring that at most N jobs can run at a time. Be aware that throttling operates at the job level, not at the queue level. For example, with a throttle that limits one job at a time and concurrency set to 25 Sidekiq will still start 25 of the same job type, but each job will wait for the lock to release and they will run sequentially. That makes it crucial to balance queues evenly so that multiple job types are enqueued simultaneously.
Configure a concurrent rate limiter that only permits 2 concurrent jobs, with generous timeouts:
LINKEDIN_THROTTLE = Sidekiq::Limiter.concurrent( 'linkedin', 2, wait_timeout: 10, lock_timeout: 60 ) def perform LINKEDIN_THROTTLE.within_limit do # Talk to LinkedIn end end
See the details of concurrent rate limiting at the Wiki.
Nothing is a Silver Bullet
Happily the workers are now plowing through
300,000+ jobs a day without any
downtime or hiccups.
None of these changes by themselves were enough to get the system running smoothly. It required a healthy amount of defensive coding, bug fixes, and configuration tuning to smooth out the platform. It may have been possible without the industrial strength features offered by Enterprise, but it would have required a lot of plugins and some hand rolled throttling. I didn’t even mention using batches, unique jobs or time based rate limiting—used properly, Pro/Enterprise save tremendous amounts of developer time.
No, I don’t receive kickbacks on Enterprise sales!