> b. Observed that many objects in OM are serialized and deserialized multiple times, which add the pressure to GC.
Here is an idea to avoid multiple serialization/deserialization in OM: https://issues.apache.org/jira/browse/HDDS-9290 Tsz-Wo On Wed, Mar 6, 2024 at 12:30 AM Devesh Singh <[email protected]> wrote: > Thank you Sammi for update. Regarding the below point: > > > > When DN is dead or decommissioned, DN info is still held in SCM > > memory. After SCM restarts, those DNs info will be gone. But Recon still > > shows these DNs on the UI. > > - Currently being tracked here and working on it: > https://issues.apache.org/jira/browse/HDDS-10409 > > Regards > Devesh > > > > > On 06-Mar-2024, at 1:55 PM, Sammi Chen <[email protected]> wrote: > > > > Attender:Hao, Mingyu, Yiyang, Xi, Bin, Jianghua, Hualong, Yuanben, > Conway, > > Chungen, Wei-Chiu, Sammi > > 1. DiDi > > a. Bad replicas are not timely detected and reported in the system > so > > that healthy replicas are deleted by RM. Some data is lost due to this. > > Suggest turning on Disk Scanner by turning on > > "hdds.container.scrub.enabled". > > b. Two consecutive tasks are scheduled to replicate the same > > container on the same DN. HDDS-9661 has fixed this issue. > > c. Plan to use high density storage servers as Ozone DN. Wei-Chiu > > shares a document about the last test experience of Ozone on high density > > storage. > > 2. Shopee > > a. When DN is dead or decommissioned, DN info is still held in SCM > > memory. After SCM restarts, those DNs info will be gone. But Recon still > > shows these DNs on the UI. > > b. Want to know the different type buckets and their supportability > > in each Ozone access interface(s3g, ofs, cli). > > > https://blog.cloudera.com/apache-ozone-a-multi-protocol-aware-storage-system/ > > 3. Qihoo > > a. multipart key info serialization and deserialization overhead is > > very high for big objects(100G). Evaluate to add a new table for > multipart > > key info. Will create a JIRA for this later. > > b. Observed that many objects in OM are serialized and deserialized > > multiple times, which add the pressure to GC. > > c. Per user throughput throttle to avoid most of the throughput > being > > used by one user. FairCallQueue is suggested. > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > For additional commands, e-mail: [email protected] > >
