btw, wouldn't median time be better estimator than average?

SELECT EXTRACT (epoch
  FROM percentile_cont(0.5) WITHIN GROUP (ORDER BY completion_time - start_time))
  FROM build
  JOIN events ON build.create_event = events.id
  WHERE build.pkg_id = %(packageID)i
    AND build.state = %(st_complete)i
    AND build.task_id IS NOT NULL;

More explanation: updateWeight is run only once in the beginning of task. So updating weight according to current (whole builder) memory usage is a) abusing of what weight means b) not too relevant. It just tries to catch, that resources are possibly near to be exhausted.

Other option would be for kojid to monitor memory independently on tasks and throttle its own task_load when memory is going low. So, task_load could be computed as sum(task_weights) * used_memory_coefficient. Coefficient could be something like for 100% free memory 1.0, for 0%, inf, for 20% 2.0 (ad hoc numbers, exponential/linear/whatever). (Coefficient could be further/instead modified with CPU/memory ratio).