Did test with streaming job and FileSystemHaService using VoidBlobStore (no HA Blob), looks like job was able to recover from both JM restart and TM restart. Any idea in what use cases HA Blob is needed?
Thanks, Alexey ________________________________ From: Alexey Trenikhun <yen...@msn.com> Sent: Friday, August 28, 2020 11:31 AM To: Khachatryan Roman <khachatryan.ro...@gmail.com> Cc: Flink User Mail List <user@flink.apache.org> Subject: Re: FileSystemHaServices and BlobStore Motivation is to have k8s HA setup without extra component - Zookeeper, see [1] Purpose of BlobStore is vague to me, what kind of BLOBs are stored? Looks like if we start job from savepoint, then persistence of BlobStore is not necessary, but is it needed if we recover from checkpoint? Thanks, Alexey [1]. https://issues.apache.org/jira/browse/FLINK-17598 ________________________________ From: Khachatryan Roman <khachatryan.ro...@gmail.com> Sent: Friday, August 28, 2020 9:24 AM To: Alexey Trenikhun <yen...@msn.com> Cc: Flink User Mail List <user@flink.apache.org> Subject: Re: FileSystemHaServices and BlobStore Hello Alexey, I think you need FileSystemBlobStore as you are implementing HA Services, and BLOBs should be highly available too. However, I'm a bit concerned about the direction in general: it essentially means re-implementing ZK functionality on top of FS. What are the motivation and the use case? Regards, Roman On Fri, Aug 28, 2020 at 5:15 PM Alexey Trenikhun <yen...@msn.com<mailto:yen...@msn.com>> wrote: Hello, I'm thinking about implementing FileSystemHaServices - single leader, but persistent RunningJobRegistry, CheckpointIDCounter, CompletedCheckpointStore and JobGraphStore. I'm not sure do you need FileSystemBlobStore or VoidBlobStore is enough. Can't figure out, should BlobStore survive JobManager crash. I see that ZookeeperHaServices use FileSystemBlobStore, but not clear is to due to having multiple JobManagers (leader + follower) or necessity to preserve BLOBs on restart. Thanks, Alexey
