Handling High Load
When worker capacity is constrained, such as in a fixed pool of hardware, the demand for work can outstrip available capacity. When this occurs, Taskcluster provides a few tools that can help.
Task priorities are useful to ensure that the most important tasks are executed first. When the capacity issues are transient, this can be helpful. However, it can cause starvation of lower-priority tasks if higher-priority tasks consume all of the limited resources.
Sometimes tasks are not useful after a given time has passed. For example, the results of a test run are probably not useful more than 12 hours after the push. Setting appropriate task deadlines can help to remove such useless tasks from the backlog if they are not executed earlier.
In truth, most often tasks become less useful when some later task would provide equivalent results. For example, a test on revision 76 provides useful results about revision 75 as well: if it passes, then likely the changes in revisions 76 and 75 were both OK. In Taskcluster terms, the task for revision 76 has superseded that for revision 75.
Most workers support superseding. It is implemented by calling out to an external service that determines the "supersedes" relationship between tasks. When superseding is possible, the worker marks older tasks as "supserseded" while performing the newest task.
See the queue reference for the technical details of the supserseding convention.