Skip to content

Latest commit

 

History

History
39 lines (25 loc) · 3.15 KB

iteration-how-it-works.md

File metadata and controls

39 lines (25 loc) · 3.15 KB
Raw

Iteration: how it works

The main idea behind Iteration is to provide an API to describe jobs in an interruptible manner, in contrast with implementing one massive #perform method that is impossible to interrupt safely.

Exposing the enumerator and the action to apply allows us to keep a cursor and interrupt between iterations. Let's see what this looks like with an ActiveRecord relation (and Enumerator).

  1. build_enumerator is called, which constructs ActiveRecordEnumerator from an ActiveRecord relation (Product.all)
  2. The first batch of records is loaded:
SELECT  `products`.* FROM `products` ORDER BY products.id LIMIT 100
  1. The job iterates over two records of the relation and then receives SIGTERM (graceful termination signal) caused by a deploy.
  2. The signal handler sets a flag that makes job_should_exit? return true.
  3. After the last iteration is completed, we will check job_should_exit? which now returns true.
  4. The job stops iterating and pushes itself back to the queue, with the latest cursor_position value.
  5. Next time when the job is taken from the queue, we'll load records starting from the last primary key that was processed:
SELECT  `products`.* FROM `products` WHERE (products.id > 2) ORDER BY products.id LIMIT 100

Signals

It's critical to know UNIX signals in order to understand how interruption works. There are two main signals that Sidekiq and Resque use: SIGTERM and SIGKILL. SIGTERM is the graceful termination signal which means that the process should exit soon, not immediately. For Iteration, it means that we have time to wait for the last iteration to finish and to push job back to the queue with the last cursor position. SIGTERM is what allows Iteration to work. In contrast, SIGKILL means immediate exit. It doesn't let the worker terminate gracefully, instead it will drop the job and exit as soon as possible.

Most of the deploy strategies (Kubernetes, Heroku, Capistrano) send SIGTERM before shutting down a node, then wait for a timeout (usually from 30 seconds to a minute) to send SIGKILL if the process has not terminated yet.

Further reading: Sidekiq signals.

Enumerators

In the early versions of Iteration, build_enumerator used to return ActiveRecord relations directly, and we would infer the Enumerator based on the type of object. We used to support ActiveRecord relations, arrays and CSVs. This made it hard to add support for other types of enumerations, and it was easy for developers to make mistakes and return an array of ActiveRecord objects, and for us starting to treat that as an array instead of as an ActiveRecord relation.

The current version of Iteration supports any Enumerator. We expose helpers to build common enumerators conveniently (enumerator_builder.active_record_on_records), but it's up to a developer to implement a custom Enumerator.

Further reading: ruby-doc, a great post about Enumerators.