-
Notifications
You must be signed in to change notification settings - Fork 143
Description
This is a followup to this issue and my comment here. I want to make a stand-alone feature request. So copying my comment here:
I am a newcomer to River, but explanation why JobStateRunning is required is surprising to me based on the rest of the documentation. UniqueOpts is part of InsertOpts and that to me means that it only controls when the job can be inserted, not how it can transition later on. For my use case (I described in #1178) I care that (transitionally) I can skip adding additional job if one is already pending, but not running. Jobs are designed so that they could run 100 of them in parallel, they are also idempotent, it is just that it is unnecessary to have more then one at any given time being scheduled to run.
About how many jobs are running in parallel, my understanding is that this is what concurrency limits are for, not insert options.
To me this is also surprising because when as a user I insert a job, it is never in any other state than an initial state. I cannot insert a job in the "running" state. That makes no sense. Only River can move job to a running state. So why would "InsertOps" control how can River move jobs between states? To me "InsertOps" is only about the user - can I as a user insert a job if any job is in a pending state but not yet running?
In fact, to me this is a feature: River should be able to transition any available job to running state. But if River does that, then my insertion (where "running" state is not among UniqueOpts) should transactionally succeed. And this is exactly the behavior I would need. If at any given moment job is not yet running or even not being started to be run, skip adding the job. It is unnecessary. The job which will shortly run will do the job. But if the job is already running, then schedule it. Because it might be that the job running will miss a bit of work.