Feature #9937
Updated by Zack Cerza over 9 years ago
Currently, to run scheduled jobs, we have loads of @teuthology-worker@ processes hanging around polling the @beanstalkd@ queue. The number of workers per @machine_type@ is hardcoded and essentially magic. The workers themselves are started by a very basic shell script that lives out-of-tree: "worker_start.sh":https://gist.github.com/zmc/e8c148eb7f43dcd99755 This approach often leads to many more jobs running than can actually get ahold of the resources they need. It is also not at all friendly to other organizations. Worse, when workers die we have to *notice* and start them up again. What we should have is an in-tree @teuthology-supervisor@ process that gets told how many workers to start, and *keeps that many running* - or, alternatively, eschews the worker concept altogether. running*.