Scheduling¶

Overview¶

The scheduling process is described below.

As the above diagram illustrates, scheduling works like so:

1. Scheduler gets a request spec from the "super conductor", containing resource requirements. The "super conductor" operates at the top level of a deployment, as contrasted with the "cell conductor", which operates within a particular cell.
2. Scheduler sends those requirements to placement.
3. Placement runs a query to determine the resource providers (in this case, compute nodes) that can satisfy those requirements.
4. Placement then constructs a data structure for each compute node as documented in the spec. The data structure contains summaries of the matching resource provider information for each compute node, along with the AllocationRequest that will be used to claim the requested resources if that compute node is selected.
5. Placement returns this data structure to the Scheduler.
6. The Scheduler creates HostState objects for each compute node contained in the provider summaries. These HostState objects contain the information about the host that will be used for subsequent filtering and weighing.
7. Since the request spec can specify one or more instances to be scheduled. The Scheduler repeats the next several steps for each requested instance.
8. Scheduler runs these HostState objects through the filters and weighers to further refine and rank the hosts to match the request.
9. Scheduler then selects the HostState at the top of the ranked list, and determines its matching AllocationRequest from the data returned by Placement. It uses that AllocationRequest as the body of the request sent to Placement to claim the resources.
10. If the claim is not successful, that indicates that another process has consumed those resources, and the host is no longer able to satisfy the request. In that event, the Scheduler moves on to the next host in the list, repeating the process until it is able to successfully claim the resources.
11. Once the Scheduler has found a host for which a successful claim has been made, it needs to select a number of "alternate" hosts. These are hosts from the ranked list that are in the same cell as the selected host, which can be used by the cell conductor in the event that the build on the selected host fails for some reason. The number of alternates is determined by the configuration option scheduler.max_attempts.
12. Scheduler creates two list structures for each requested instance: one for the hosts (selected + alternates), and the other for their matching AllocationRequests.
13. To create the alternates, Scheduler determines the cell of the selected host. It then iterates through the ranked list of HostState objects to find a number of additional hosts in that same cell. It adds those hosts to the host list, and their AllocationRequest to the allocation list.
14. Once those lists are created, the Scheduler has completed what it needs to do for a requested instance.
15. Scheduler repeats this process for any additional requested instances. When all instances have been scheduled, it creates a 2-tuple to return to the super conductor, with the first element of the tuple being a list of lists of hosts, and the second being a list of lists of the AllocationRequests.
16. Scheduler returns that 2-tuple to the super conductor.
17. For each requested instance, the super conductor determines the cell of the selected host. It then sends a 2-tuple of ([hosts], [AllocationRequests]) for that instance to the target cell conductor.
18. Target cell conductor tries to build the instance on the selected host. If it fails, it uses the AllocationRequest data for that host to unclaim the resources for the selected host. It then iterates through the list of alternates by first attempting to claim the resources, and if successful, building the instance on that host. Only when all alternates fail does the build request fail.