Yes service accounts are fine and dandy but [a] isn't that more hoops borne by choosing the wrong tool for the job (gotta stay mindful of those permissions, you can have an Ansible typo, etc) and [b] as I said, proper queue services would not even allow tampering of enqueued data. For no extra cost other than installing the service. In contrast, your service account falling into a malicious actor is still a threat vector.
> For a more complex — but perhaps useful? — option ...
Now this is just a queue service with extra steps running in a relational DB instead of natively as an OS process. You did cite it as just an option but I don't see why this is an attractive option.
> I didn't see any mention of anything I'd consider an OpEx cost.
It seems to me we have differing definitions of OpEx and, judging by your previous comment, you value this almost exclusively in terms of "the number of ops salaries you have to pay". Even if I play by that definition, I could tell from experience (and also corroborated by other commenters here, in other subthreads of this discussion), that operating a Redis cluster does not need extra warm bodies. The people who monitor that your app responds, your servers aren't running out of memory, could take on the task with little to no additional training.
The experience I want to cite, bearing in mind of course that everyone's mileage varies: in my previous job, customer base of 2M and growing across two countries, we operated and monitored a redundant Redis cluster with an ops team of 3. In my first job I talked about in another subthread here, we managed three high-traffic Redis queues (and a few other smaller-traffic special cases) with a team of 7. In both jobs, we also did active across-the-stack development; we weren't dedicated to monitoring, nor to the components related to the queues.
In fact I would argue running a DB (queue or no queue) is just more complex than running a queue service. Rather telling is the need for "DB Administrators/Specialists" but no such equivalent for Redis or other proper queueing solutions.
> Which is usually also a cost already paid for you in advance when you use an abstract job-queue library. They all know how to expose Prometheus metrics endpoints...
Honestly, this sounds fine but I'm rather wary of the caveats that might be present. How do they do it efficiently? What do I need to keep in mind to keep things efficient? As such, unless you can point me to your preferred job-queue library which does all these wonders, I have to reserve judgment.
Edit: I just remembered mentioning this in another subthread. Even if you have a fantastic library/middleware abstracting all those queue concerns, that ties you in with that library. If someone wants to build integrations into your queue, they have to play by that library. If majority of your app is in Java (and so is the library), and Data Science (who, in this example, uses Python, not improbable to happen) wants to share your data pipeline, if that library isn't available for them, tough luck I guess? More dev time for someone.
And also, whatever underlying DB features or semantics your library might rely on in order to enforce a queue structure, you can't be assured that future versions of your DB would support those features because, you know, your DB isn't really in the business of queues. It opens you up to running an outdated DB version just to keep that queue functionality.