Details
-
Improvement
-
Status: Accepted
-
Major
-
Resolution: Unresolved
-
None
-
None
Description
Status Quo
Currently resources allocated to frameworks in a role with quota (aka quota'ed role) beyond quota guarantee are marked non-revocable. This impacts our flexibility for revoking them if we decide so in the future.
Proposal
Once quota guarantee is satisfied we must not necessarily further allocate resources as non-revocable. Instead we can mark all offers resources beyond guarantee as revocable. When in the future RevocableInfo evolves frameworks will get additional information about "revocability" of the resource (i.e. allocation slack)
Caveats
Though it seems like a simple change, it has several implications.
Fairness
Currently the hierarchical allocator considers revocable resources as regular resources when doing fairness calculations. This may prevent frameworks getting non-revocable resources as part of their role's quota guarantee if they accept some revocable resources as well.
Consider the following scenario. A single framework in a role with quota set to 10 CPUs is allocated 10 CPUs as non-revocable resources as part of its quota and additionally 2 revocable CPUs. Now a task using 2 non-revocable CPUs finishes and its resources are returned. Total allocation for the role is 8 non-revocable + 2 revocable. However, the role may not be offered additional 2 non-revocable since its total allocation satisfies quota.
Resource math
If we allocate non-revocable resources as revocable, we should make sure we do accounting right: either we should update total agent resources and mark them as revocable as well, or bookkeep resources as non-revocable and convert them to revocable when necessary.
Coarse-grained nature of allocation
The hierarchical allocator performs "coarse-grained" allocation, meaning it always allocates the entire remaining agent resources to a single framework. This may lead to over-allocating some resources as non-revocable beyond quota guarantee.
Quotas smaller than fair share
If a quota set for a role is smaller than its fair share, it may reduce the amount of resources offered to this role, if frameworks in it do not accept revocable resources. This is probably the most important consequence of the proposed change. Operators may set quota to get guarantees, but may observe a decrease in amount of resources a role gets, which is not intuitive.
Attachments
Issue Links
- relates to
-
MESOS-4529 Update the allocator to not offer unreserved resources beyond quota.
-
- Resolved
-