Add capability to discard duplicate jobs with concurrency configuration #523

joelzwarrington · 2025-02-25T08:50:36Z

Closes #176

Add ability to discard duplicate jobs instead of becoming blocked. When configuring concurrency_limits with at_limit: discard jobs scheduled above the concurrency limit are discarded and not executed.

Implementation Details

adds class_attribute concurrency_at_limit and at_limit option to concurrency_limits
adds methods to Semaphore to detect when a job would be above the concurrency limit
adds logic to discard a job on dispatch when at concurrency limit

…ify control flow

rosa · 2025-03-05T09:58:57Z

app/models/solid_queue/job/executable.rb

@@ -66,6 +66,8 @@ def prepare_for_execution

      def dispatch
        if acquire_concurrency_lock then ready
+        elsif discard_concurrent?
+          discard


I don't think this is the way 🤔 We're in the middle of a transaction here, and the job hasn't even been committed to the DB. It makes no sense to delete a record in the same transaction you're creating it. It'd make sense to roll that transaction back instead.

Perhaps I don't fully understand how we're getting into this code path, but from my investigation, it looks as though there isn't an open transaction at this time.

I tried the following and running the test suite and didn't hit any open transactions.

raise "open transactions" if ApplicationRecord.connection.open_transactions.positive?

I still believe we'd want to discard here. Let me know your thoughts.

rosa · 2025-03-05T10:02:15Z

app/models/solid_queue/semaphore.rb

@@ -39,6 +43,14 @@ def initialize(job)
        @job = job
      end

+      def at_limit?
+        if semaphore = Semaphore.find_by(key: key)
+          semaphore.value.zero?


This is vulnerable to race conditions and the reason we don't check it in this way when blocking jobs. If two concurrent jobs are claiming the semaphore, both of them will see it open, and none will be discarded. Then, both will run together because we won't block them either.

You raise an excellent point. Do you have a suggestion to how we might be able to avoid this race condition?

Are you thinking Pessimistic Locking might help?

Yes, I was thinking that perhaps we could rely on the same check we're already doing to block the job, but instead of blocking the job, we'd rollback the transaction 🤔 What I'm not sure is whether this should actually raise some exception to indicate the job hasn't been enqueued. Maybe that's not necessary, but at least it should set the successfully_enqueued attribute in the active job to false.

joelzwarrington · 2025-03-06T17:10:06Z

Thanks for the review @rosa ❤️, I'll review your comments more thoroughly and make some changes. 😊

…carded jobs

joelzwarrington

Hi @rosa, I've made a few changes and believe to have encountered weird interactions with bulk enqueuing which might be a bug. Let me know if you'd like those changes split out into another PR.

joelzwarrington · 2025-03-08T20:25:23Z

app/models/solid_queue/job.rb

+          active_jobs.each do |active_job|
+            job = enqueued_jobs_by_active_job_id[active_job.job_id]
+            active_job.provider_job_id = job&.id
+            active_job.successfully_enqueued = job.present?


I've reorganized this method so that we can efficiently set successfully_enqueued for both enqueued/blocked and discarded jobs.

🤔 I think the previous implementation did that as well (only setting successfully_enqueued = true for jobs that had been actually enqueued), it should work without doing anything for discarded jobs, as long as discarded jobs don't create a job record. I must be missing something here, why do you think the previous implementation didn't set that correctly?

@rosa previous implementation would only set successfully_enqueued = true, but wouldn't set successfully_enqueued = false if it wasn't enqueued.

If you're okay with that quirk, then I can revert this change.

joelzwarrington · 2025-03-08T20:31:19Z

app/models/solid_queue/job.rb

@@ -49,7 +52,7 @@ def create_from_active_job(active_job)
        def create_all_from_active_jobs(active_jobs)
          job_rows = active_jobs.map { |job| attributes_from_active_job(job) }
          insert_all(job_rows)
-          where(active_job_id: active_jobs.map(&:job_id))
+          where(active_job_id: active_jobs.map(&:job_id)).order(id: :asc)


I think this may have been an unintentional bug, or would be something that's good to do moving forward with more concurrency configuration.

Previous to adding this order, the jobs could be dispatched out of order they were provided in. This isn't really a problem when there isn't concurrency controls on the job, but when using concurrency controls, it might have unintended consequences.

So if you were to do:

jobs = [ job_1 = DiscardedNonOverlappingJob.new(@result, name: "A"), job_2 = DiscardedNonOverlappingJob.new(@result, name: "B") ] enqueued_jobs_count = SolidQueue::Job.enqueue_all(jobs)

There wasn't any guarantee that job_1 would be dispatched before job_2. Now, it should be as these jobs are bulk inserted with auto-incrementing IDs, job_1 will have the lower id.

Concurrency controls don't guarantee anything about the order, though 🤔 I think in the case you're enqueuing two jobs together (in bulk), concurrency-controlled, there shouldn't be any guarantees that one is run before the other.

Intuitively, if I were using bulk enqueue and concurrency controls, I would want it to enqueue the jobs in the order they were created, no?

If you don't want to make that guarantee, I can remove these changes and instead add a more prominent call-out in the bulk enqueue documentation that ordering is not guaranteed.

joelzwarrington · 2025-03-08T20:34:38Z

app/models/solid_queue/job/executable.rb

@@ -66,6 +66,7 @@ def prepare_for_execution

      def dispatch
        if acquire_concurrency_lock then ready
+        elsif discard_concurrent? then discard


I don't believe discard does anything here, since it just deletes the execution. But I believe in this point we won't have an execution?

This isn't deleting the job as I once thought it was. This is effectively doing 'nothing'.

I don't believe discard does anything here, since it just deletes the execution. But I believe in this point we won't have an execution?

That's right, we won't have an execution yet. However, we need to ensure we don't create the job. I think this will result in the job record being created. That's why I think we need to rollback the current transaction. There should be a transaction open since this is called as an after_create callback on the job.

@rosa ah - thanks for pointing out the callback. Because dispatch is also used in bulk enqueueing, should we rollback all jobs in the bulk enqueue if one failed to be enqueued due to concurrency controls?

rajeevriitm · 2025-03-10T12:02:49Z

I am really in need of this feature. As its not yet available , could you help me in implementing this ? I have a ContinuousSearchJob thats run with recurring.yml . Sometimes the job takes longer and I dont want another job to be added to the queue unless the existing one is finished. Could you help me with the best way I can implement this currently?

joelzwarrington · 2025-03-13T15:19:52Z

@rajeevriitm if you have a look at the read me you should find concurrency controls are available for you to use.

The only difference being that jobs will always be blocked, so you may get several jobs that are awaiting execution.

joelzwarrington · 2025-04-28T05:57:05Z

@rosa any updates on your review here?

rynonl · 2025-05-30T14:48:40Z

@rosa Just pinging this. Would love to see this functionality get merged!

joelzwarrington added 2 commits February 25, 2025 01:40

Add capability to discard duplicate jobs with concurrency configuration

f290af1

Remove 'duplicate' verbiage and use concurrency limits instead, simpl…

e561356

…ify control flow

joelzwarrington mentioned this pull request Feb 26, 2025

Discard duplicate jobs #176

Open

rosa reviewed Mar 5, 2025

View reviewed changes

joelzwarrington added 4 commits March 7, 2025 13:33

Fix race condition vulnerability by changing logic to enqueue

90793ad

Add assertions when bulk enqueuing jobs with concurrency controls

b8dae8e

Dispatch jobs in the order they were enqueued

ddb513f

Set ActiveJob successfully_enqueued for both enqueued/blocked and dis…

a3a7049

…carded jobs

joelzwarrington commented Mar 8, 2025

View reviewed changes

toydestroyer mentioned this pull request Mar 18, 2025

Skip recurring jobs if a queue is paused #535

Open

joelzwarrington requested a review from rosa March 19, 2025 21:30

joshleblanc mentioned this pull request Jun 10, 2025

How to check if a solid-queue job is already pending with arguments? #571

Open

Add capability to discard duplicate jobs with concurrency configuration #523

Are you sure you want to change the base?

Add capability to discard duplicate jobs with concurrency configuration #523

Uh oh!

Conversation

joelzwarrington commented Feb 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Implementation Details

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

joelzwarrington commented Mar 6, 2025

Uh oh!

joelzwarrington left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rajeevriitm commented Mar 10, 2025

Uh oh!

joelzwarrington commented Mar 13, 2025

Uh oh!

joelzwarrington commented Apr 28, 2025

Uh oh!

rynonl commented May 30, 2025

Uh oh!

Uh oh!

joelzwarrington commented Feb 25, 2025 •

edited

Loading