Allow separate *_keys for execution/enqueue/rate limits

[pgvsalamander]

What

I think the ability to set separate job-identification keys would be useful, especially if/when ratelimiting is implemented. The separate keys would be optional, defaulting to the value for key if a value is not provided / a lambda evaluates to nil.

good_job_control_currency_with(
  key: "MyJob",
  enqueue_key: -> { "MyJob-#{arguments.first.get_identifier}" },
  # execution_key is not set
  ratelimit_key: "external-service-name"
)

# Evaluates to
enqueue_key == "MyJob-<entity-specific identifier here>"
execution_key == "MyJob"
ratelimit_key == "external-service-name"

Why

My use case here is handling multiple jobs that interact with the same database entities, guarding against duplicate jobs and write-collisions.

The enqueue_key would be set using the job identifier and the entity identifier. This prevents duplicate jobs from being queued (same job + entity), while still allowing multiple jobs that reference the same entity to be queued (job A + entity 1, job B + entity 1).

The execution_key would be set using just the entity identifier, which prevents multiple jobs from trying to modify the same entity at a time (job A + entity 1 and job B + entity 1 can both exist, but cannot run simultaneously).

If/when ratelimiting is implemented it would be another layer on top of this, allowing the previous behavior while ensuring that every job with the ratelimit_key doesn't exceed the set limit.

How

I think the biggest part of this work would be accomplished by switching .where(concurrency_key: ...) to .where(execution_key: ...) in /lib/good_job_active_job_extensions/concurrency.rb:95, and .where(concurrency_key: ...) to .where(enqueue_key: ...) on lines 53, 55 of the same file. The aforementioned columns and properties would also need to be created, but those are more straightforward.

If I'm on the right track with these changes and if you think they're a good fit for the project I'd be happy to try them out and put a MR in.

Tangentially related issues:

• Add an ActiveJob extension for throttling #315
• Feature: dynamic options for good_job_control_concurrency_with #684

[bensheldon]

@pgvsalamander thank you for the proposal! I'll need to think about it a bit more.

I want to throw out a thought seed: instead of moving towards more specific keys, is there a more generic data structure we could use? I'm wondering if jobs had more generic tags (array column of strings) whether that would be enough to do indexing/counting and give flexibility for generating concurrency keys.

[pgvsalamander]

I like the idea, although there might be some performance concerns depending on the implementation.

How would the tags be used? The only way that comes to mind would be applying a type to each tag so they're a type-value pair, and then applying the different type(s) to the different concurrency controls.

Each job would receive a tag representing the job name (job: MyJob) and a tag with the entity identifier (entity: EntityType#1).
The example from above would be implemented as the execution constraint require a lock on the entity tag to run, and the enqueue constraint requiring a unique lock on the [job, entity] pairing.

I think that'd require an additional WHERE search_tag IN job_tags for each type-constraint. An array-intersection would be more ideal (WHERE search_tags IN job_tags) but I'm not sure if that's possible in the query builder, or in all databases. All that said, I'm not sure what the performance difference is between WHERE A = B and WHERE A IN B, so I may be trying to optimize the queries prematurely.

More thought seeds - How are multiple tags of the same type, or missing tags handled? Are they allowed at all, or are you required to define a specific set of tags on all jobs / all jobs of a given type? Are the tags only a thing on the GoodJob backend, and they're displayed as separate pre-defined field types on the user-side?