Under specific circumstances and due to the caching of queued tasks ($miniqueue) in the task managers get_next_adhoc_task() function a task that was marked as failed is getting re-executed right after, ignoring the new nextruntime/faildelay.

For the bug to get triggered it is required that the number of "unique tasks" (tasks from different classes) do not change between subsequent calls of manager::get_next_adhoc_task():

        $uniquetasksinqueue = array_map(

            ['\core\task\manager', 'adhoc_task_from_record'],

            $DB->get_records_sql(

                'SELECT classname FROM {task_adhoc} WHERE nextruntime < :timestart GROUP BY classname',

                ['timestart' => $timestart]

            )

        );        if (!isset(self::$numtasks) || self::$numtasks !== count($uniquetasksinqueue)) {

            self::$numtasks = count($uniquetasksinqueue);

            self::$miniqueue = [];

        }

The occurrence of the bug can be seen in the task log showing multiple entries for the same task id with a started timestamp just a few seconds apart. I also verified the bug by reviewing the code and inserting a debug output into the foreach loop of the get_next_adhoc_task() function:

if ($record->nextruntime >= $timestart) {

    mtrace("Running a task with its nextruntime ($record->nextruntime) in the future (now: $timestart). Oups!");

}

This bug was introduced with MDL-67648 and is still present on the master branch.