-
Type:
Improvement
-
Status: Open
-
Priority:
Major
-
Resolution: Unresolved
-
Affects Version/s: 1.9.0
-
Fix Version/s: 1.13.0
-
Component/s: Runtime / Coordination
-
Labels:None
As we are enabling fine grained resource management in 1.10, we can have various resource scheduling strategies. Such strategies generally should make the following three decisions.
- When to launch new / release existing TMs? (How many TMs)
- What and how many resources should TMs be started with?
- How to allocate between slot requests and TM resources?
We may want to make above decisions differently in different scenarios (active/reactive mode, perjob/session mode, etc.). Therefore, we propose to make the scheduling strategies pluggable.
We propose to make the following changes:
- Make SlotManager an interface, and implements it differently for different strategies strategies.
- Modify ResourceManager-SlotManager interfaces to cover all the three decisions mentioned above in SlotManager. In particular, SlotManager needs to allocate TM resources instead of slot resources from ResourceActions.
- blocks
-
FLINK-14187 FLIP-56 Dynamic Slot Allocation
-
- Open
-
- is related to
-
FLINK-19324 Map requested/allocated containers with priority on YARN
-
- Closed
-
- is required by
-
FLINK-5791 Support an optimal matching based slot manager for flip6 yarn mode
-
- Closed
-
- relates to
-
FLINK-17061 Unset process/flink memory size from configuration once dynamic worker resource is activated.
-
- Closed
-
- links to