----- Original Message ----- > Zhou, > > I haven't read the *Design* yet, but I don't think it is solving the same > > question between priorities and quota. For example, assume we only have 10G > > memory reservating for framework A totally, then another urgency framework > > is getting nothing. which is statical partition still. While priority can > > pre-empt that. > > > > I'm not sure what is your concern here. If all you have is 10GB and both > your frameworks A and B may need 10GB each at the same time, you definitely > need to add more RAM : ). Mesos uses fair sharing for distributing > resources among frameworks, quota will not be an exception.
Apologies if these questions are already answered somewhere in the quota design doc, but I couldn't see it. What kind of number is a quota? What units does it come in? If I say "the quota for role A is 1000", 1000 what? What does it mean to distribute quota among frameworks using DRF, "like any other resource" Is quota another kind of resource separate from others? Are there plans to handle sub-quotas (In the vocabulary of HTCondor, Hierarchical Group Quotas). That is, will users be able to configure quotas inside of higher-level quotas? How does resource preemption work, when role A is using surplus quota that suddenly is required by role B to satisfy B's configured quota guarantee? > If all you have > is 10GB and for both A and B you have reserved 10GB, total reserved > resources are 20GB, which means your cluster is under quota. I would say, > if this is happening in production cluster, several devops engineers should > have already been paged : ). > > It's up to allocator implementation to decide what to do in this situation. > An obvious approach is to throttle (i.e. revoke resources) both frameworks > proportionally to their role weights. Quota does not introduce static > partitioning, it rather guarantees, a production framework gets enough > resources regardless of any events happening in the cluster, given these > resources are available. >

