[ https://issues.apache.org/jira/browse/FLINK-15959?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17039668#comment-17039668 ]
Xintong Song commented on FLINK-15959: -------------------------------------- [~liuyufei] bq. otherwise scheduler will use slots from SlotPool until available slots are empty, this also can cause load balance issue. Not sure about this. In case of job failover, all the slots in {{SlotPool}} should be free when recovering the job, and the `evenly-spread-out-slots` introduced in FLINK-12122 applies to those slots in {{SlotPool}} as well. It's true that we cannot always guarantee to evenly spread out slots across all the TMs in case of TM lost. We may first spread out slot requests across the existing TMs (whose slots are still in {{SlotPool}}), and the remaining unfulfilled slot requests will be allocated to new TMs once the minimum slots are registered to RM. I think this might provide good enough load balance in most cases, especially in failovers cased by only one or a few TMs are lost. bq. maybe should implement it in scheduler, allocate enough resource before execution scheduling. I think this is a good idea for solving the load balance problem. My concern it that, it might be conflict with some of the community road map plans. - One thing I'm aware of is that, the community is planning for a declarative resource management approach, in which a job declares its all resource requirements at once, instead of currently requesting slots for individual task/execution separately. This effort is still in early design discussions, and may not be finished in the next release. But I would try to avoid making many changes to {{Scheduler}} and {{SlotPool}} at this time considering they might be changed again in short time. - Another thing that I heard of, not completely sure about this, is that people are considering getting rid of {{SlotPool}}, or at least make it as less responsibility as possible. Because currently we do not benefit much from caching slots in {{SlotPool}}, but suffers from the complication that resources are managed at two places, the {{SlotPool}} and the {{ResourceManager}}. That's also why I do not like the idea of adding more responsibility to {{SlotPool}}. I'm not very familiar {{Scheduler}} and {{SlotPool}} though. Maybe [~trohrmann] or [~zhuzh] could chime in. > Add min/max number of slots configuration to limit total number of slots > ------------------------------------------------------------------------ > > Key: FLINK-15959 > URL: https://issues.apache.org/jira/browse/FLINK-15959 > Project: Flink > Issue Type: Improvement > Components: Runtime / Coordination > Affects Versions: 1.11.0 > Reporter: YufeiLiu > Priority: Major > > Flink removed `-n` option after FLIP-6, change to ResourceManager start a new > worker when required. But I think maintain a certain amount of slots is > necessary. These workers will start immediately when ResourceManager starts > and would not release even if all slots are free. > Here are some resons: > # Users actually know how many resources are needed when run a single job, > initialize all workers when cluster starts can speed up startup process. > # Job schedule in topology order, next operator won't schedule until prior > execution slot allocated. The TaskExecutors will start in several batchs in > some cases, it might slow down the startup speed. > # Flink support > [FLINK-12122|https://issues.apache.org/jira/browse/FLINK-12122] [Spread out > tasks evenly across all available registered TaskManagers], but it will only > effect if all TMs are registered. Start all TMs at begining can slove this > problem. > *suggestion:* > * Add config "taskmanager.minimum.numberOfTotalSlots" and > "taskmanager.maximum.numberOfTotalSlots", default behavior is still like > before. > * Start plenty number of workers to satisfy minimum slots when > ResourceManager accept leadership(subtract recovered workers). > * Don't comlete slot request until minimum number of slots are registered, > and throw exeception when exceed maximum. -- This message was sent by Atlassian Jira (v8.3.4#803005)