[ https://issues.apache.org/jira/browse/YARN-1040?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15167794#comment-15167794 ]
Vinod Kumar Vavilapalli commented on YARN-1040: ----------------------------------------------- Still catching up on some of your discussion, but quick comments on a few things that I care We should really call them instead as {{Allocation}} and {{Container}}. That's the nomenclature I used at YARN-4726. Till now, YARN has combined the notion of Allocation and Container, which is the main de-linking that we need to do here. Process has an OS level connotation, and doesn't work well with more things in the picture like process-trees / multiple-processes / Docker (YARN-2466). Taking this further, here's how the overall picture can look like the following *ResourceManager* - RM only does allocations in the scheduling path. ResourceManager does all scheduling based on AllocationRequests and tracks Allocations.. - RM receives AllocationRequests and returns fulfilled Allocations (and AllocationTokens) to AMs. *Applications* - AM can in turn use the Allocations (and AllocationTokens) to launch multiple Containers on the NM. -- Simple case: AM only launches containers one-after-another. It's up to the app to do this. -- General case: AM launches multiple containers at the same time. This is essentially container-groups - we should keep this option open. - AMs can specify *single-use* AllocationRequests, at which point RM can simply return Containers and Container-Tokens (today's code-path). - Each Container exits when the process-tree / linux-container exits. - Each Container has an Identifier. -- For single-use allocation-requests, RM generates ContainerIDs -- For multi-use allocation-requests, apps could optionally specify a container-name that is scoped under the allocation. NM always returns a (generated or app-specified) ContainerID based off the allocation-ID. Essentially, allocationID + containerID is unique *NodeManagers* - NodeManager also understands incoming Allocations and ties them to Container groups: it deals with Allocation activation/deactivation and Container start/stop. but does -- the following *decoupled from both allocations and containers*: localizations / re-localizations. This means local-resources should now have more scopes: container, allocation, application etc. -- *per allocation*: enforcement of resource-limits -- all of the following *per container*: (a) process/OS-container activation / deactivation, (b) process/OS-container auto-restart (YARN-4725) log-aggregation > De-link container life cycle from the process and add ability to execute > multiple processes in the same long-lived container > ---------------------------------------------------------------------------------------------------------------------------- > > Key: YARN-1040 > URL: https://issues.apache.org/jira/browse/YARN-1040 > Project: Hadoop YARN > Issue Type: Sub-task > Components: nodemanager > Affects Versions: 3.0.0 > Reporter: Steve Loughran > > The AM should be able to exec >1 process in a container, rather than have the > NM automatically release the container when the single process exits. > This would let an AM restart a process on the same container repeatedly, > which for HBase would offer locality on a restarted region server. > We may also want the ability to exec multiple processes in parallel, so that > something could be run in the container while a long-lived process was > already running. This can be useful in monitoring and reconfiguring the > long-lived process, as well as shutting it down. -- This message was sent by Atlassian JIRA (v6.3.4#6332)