Major Description
We are seeing an issue where user limit factor does not cap the amount of resources a user can consume in a queue in a partition. Suppose you have a queue with access to partition X, used resources in default partition is 0, and used resources in partition X is at the partition's user limit. This is the problematic code as far as I can tell: (in LeafQueue.java)
if (Resources
.greaterThan(resourceCalculator, clusterResource,
user.getUsed(label),
limit)) {
// if enabled, check to see if could we potentially use this node instead
// of a reserved node if the application has reserved containers
if (this.reservationsContinueLooking) {
if (Resources.lessThanOrEqual(
resourceCalculator,
clusterResource,
Resources.subtract(user.getUsed(), application.getCurrentReservation()),
limit)) {
if (LOG.isDebugEnabled()) {
LOG.debug("User " + userName + " in queue " + getQueueName()
+ " will exceed limit based on reservations - " + " consumed: "
+ user.getUsed() + " reserved: "
+ application.getCurrentReservation() + " limit: " + limit);
}
Resource amountNeededToUnreserve = Resources.subtract(user.getUsed(label), limit);
// we can only acquire a new container if we unreserve first since we ignored the
// user limit. Choose the max of user limit or what was previously set by max
// capacity.
currentResoureLimits.setAmountNeededUnreserve(Resources.max(resourceCalculator,
clusterResource, currentResoureLimits.getAmountNeededUnreserve(),
amountNeededToUnreserve));
return true;
}
}
if (LOG.isDebugEnabled()) {
LOG.debug("User " + userName + " in queue " + getQueueName()
+ " will exceed limit - " + " consumed: "
+ user.getUsed() + " limit: " + limit);
}
return false;
}
First it sees the used resources in partition X is greater than partition's user limit. Then the reservation check also succeeds because it is checking user.getUsed() - application.getCurrentReservation() <= limit and returns true.
One fix is to just set Resources.subtract(user.getUsed(), application.getCurrentReservation()) to Resources.subtract(user.getUsed(label), application.getCurrentReservation()).
This doesn't seem to be a problem in branch-2.8 and higher since YARN-3356 introduces this check:
if (this.reservationsContinueLooking && checkReservations
&& label.equals(CommonNodeLabelsManager.NO_LABEL)) {
so in this case getting the used resources in default partition seems to be correct.