Best Practices of Production-Grade Rook/Ceph Cluster

My presentation slide at Cephalocon 2023
https://sched.co/1JKZf

Transcript

1. Best Practices of Production-Grade Rook/Ceph Cluster Apr. 17th, 2023 Cybozu, Inc. Satoru Takeuchi 1
2. Agenda 2 ▌Cybozu and our storage system ▌Quick introduction to K8s and Rook ▌Advanced Rook/Ceph cluster ▌Efforts and challenges ▌Conclusion
3. Agenda 3 ▌Cybozu and our storage system ▌Quick introduction to K8s and Rook ▌Advanced Rook/Ceph cluster ▌Efforts and challenges ▌Conclusion
4. About Cybozu ▌A leading cloud service provider in Japan ▌Providing Web services that supports teamwork 4
5. Our infrastructure ▌We have an on-premise infrastructure ▌Current system has many problems ⚫Not scalable ⚫A lot of day 2 work ▌Developing a new infrastructure ⚫Kubernetes(K8s) ⚫Ceph and Rook(Ceph orchestration in K8s) 5
6. Why Ceph? ▌Fulfill our requirements ⚫block & object storage ⚫Rack failure tolerance ⚫Bit-rot tolerance ▌Open source ⚫Detailed evaluation investigation of problems ⚫Use our custom container in case of emergency (see later) 6
7. Why Rook? ▌Manage storage system in k8s way as other system components ▌Offload a lot of work to K8s ⚫Lifecycle management of hardware ⚫Mon failover ⚫Restart problematic Ceph daemons 7
8. OSD(data) OSD(data) OSD(data) OSD(data) Our storage system 8 Ceph cluster for RGW Ceph cluster for RBD HDD SSD HDD HDD SSD NVMe SSD LogicalVolume LVM VolumeGroup OSD(data) LogicalVolume OSD(index) OSD Applications bucket RBD Applications Applications LogicalVolume node node • 3 replicas • Rack failure tolerance Rook manage
9. Agenda 9 ▌Cybozu and our storage system ▌Quick introduction to K8s and Rook ▌Advanced Rook/Ceph cluster ▌Efforts and challenges ▌Conclusion
10. What is K8s ▌A container orchestration ▌All services work as “pod” ⚫Pod is a set of containers 10 Kubernetes node Pod Container Container Pod Pod node node … Pod Pod Pod Pod Pod Pod
11. The concept of K8s ▌All configurations are described as resources ▌K8s keeps the desired states 11 Kubernetes Pod node Pod resource PersistenVolumeClaim (PVC) resource PersistenVolume (PV) resource volume 10 GiB volume please! There is A 10 GiB volume! Provisioned by storage Provisioner (driver)
12. ▌All Ceph components are described as resources ▌Rook keeps the desired state of Ceph clusters The concept of Rook 12 CephCluster Resource CephBlockPool resource Pod resource(MON) Pod resource(MGR) Pod resource(OSD) PV resource (for OSD) PVC resource (for OSD) MON pod MGR pod OSD pod disk node node node admin node Rook pod create watch create K8 cluster
13. Agenda 13 ▌Cybozu and our storage system ▌Quick introduction to K8s and Rook ▌Advanced Rook/Ceph cluster ▌Efforts and challenges ▌Conclusion
14. Our Rook/Ceph clusters ▌Requirements ⚫3 replicas ⚫Rack failure tolerance ⚫All OSDs should be spread evenly over all racks/nodes ▌Typical operations ⚫Create and upgrade clusters ⚫Manage OSDs (add a new one, replace a damaged one) ⚫etc. 14
15. ▌Just create the following resource 15 Create a Ceph cluster kind: CephCluster Metadata: name: ceph-ssd … mgr: count: 2 … … mon: count: 3 … storage: … count: 3 • All OSDs are evenly spread over all nodes and racks • Special configurations are necessary (see later) rack rack rack MON pod MGR pod OSD pod node node node node node node OSD pod OSD pod MON pod MON pod MGR pod
16. Create an RBD pool ▌Just create the following resources 16 kind: CephBlockPool metadata: name: block-pool spec: replicated: size: 3 failureDomain: zone kind: StorageClass metadata: name: ceph-block parameters: clusterID: ceph-ssd pool: block-pool csi.storage.k8s.io/fstype: ext4 Ceph cluster A pool for rbd Name:block-pool Replicated: 3 Failure domain: rack “zone” means “rack” in our clusters ceph-block provisioner
17. Consume an RBD image ▌Just create the following resources 17 kind: PersistentVolumeClaim metadata: name: myclaim spec: … storageClassName: ceph-block kind: Pod metadata: name: mypod volumes: - name: myvolume persistentVolumeClaim: claimName: myclaim Ceph cluster A pool for RBD Name: block-pool Replicas: 3 Failure domain: rack RBD volume mypod create consume ceph-block provisioner
18. Configurations ▌ Edit the “rook-config-override” ConfigMap resource corresponding to “/etc/ceph.conf” ⚫ Restarting Ceph pods is necessary after that ▌Some configurations (e.g. “ceph set”) can’t use this way ⚫ Run Ceph commands in the “toolbox” pod 18 kind: ConfigMap metadata: name: rook-config-override data: config: | debug rgw = 5/5
19. ▌Just edit the CephCluster resource 19 Expand a cluster kind: CephCluster … storage: … count: 6 Change from 3 rack rack rack OSD pod node node node node node node OSD pod OSD pod OSD pod OSD pod OSD pod MON pod MGR pod MON pod MON pod MGR pod
20. ▌Just run our home-made script ⚫There isn’t such the official job/script Replace a damaged OSD 20 rack rack rack OSD pod node node node node node node OSD pod OSD pod OSD pod OSD pod OSD pod MON pod MGR pod MON pod MON pod MGR pod OSD pod remove create A home-maid script
21. Upgrade Rook and Ceph ▌Edit the CephCluster resource ▌All Ceph container images will be upgraded after that ▌Other work might be needed ⚫See files under “Documentation/Upgrade” of Rook repository 21 kind: CephCluster … image: ceph/ceph:v17.2.6 Change from “ceph/ceph:v17.2.5”
22. Troubleshooting ▌Same as other Ceph orchestrations ⚫Running Ceph commands ⚫Referring logs and metrics ⚫Report bugs to upstream ▌Rook is convenient but is not a silver bullet ⚫Rook can’t debug and fix bugs 22
23. Advanced configurations ▌Even OSD deployment over all nodes/racks ▌Automatic OSD deployment 23
24. A subject about even OSD deployment ▌K8s deploy Pods to arbitrary nodes by default ⚫ OSDs might be unevenly spread 24 node node kind: CephCluster … storage: … count: 3 One node lost results in data loss node OSD OSD OSD
25. Solution ▌Use the “TopologySpreadConstraints” feature of K8s ⚫ Spread specific pods evenly over all nodes (or racks, and so on) 25 node kind: CephCluster … storage: … count: 3 … topologySpreadConstraints: - labelSelector: matchExpressions: - key: app operator: In values: - rook-ceph-osd - rook-ceph-osd-prepare topologyKey: topology.kubernetes.io/hostname OSD node OSD node OSD See also: https://blog.kintone.io/entry/2020/09/18/175030
26. rack In our clusters ▌Two constraints for both nodes and racks 26 kind: CephCluster … storage: … count: 12 … topologySpreadConstraints: - labelSelector: … topologyKey: topology.kubernetes.io/zone - labelSelector … topologyKey: topology.kubernetes.io/hostname rack rack node OSD OSD node OSD node OSD node OSD node OSD node OSD OSD OSD OSD OSD OSD
27. A subject about automatic OSD deployment ▌OSD creation flow 1. Rook create a PVC resources for an OSD 2. K8s bind a PV resource to this PVC resource 3. Rook creates an OSD on top of the block device corresponding to the PV ▌It’s suspended if the PV resource isn’t available 27 OSD OSD Ceph cluster OSD A Storage provisioner Not provisioned yet kind: CephCluster … storage: … count: 3
28. Solution ▌Use storage providers supporting dynamic provisioning ⚫e.g. many cloud storages, TopoLVM (for local volumes) ▌Provision PV resources beforehand ⚫e.g. local-static-provisioner 28 storage: … count: 3 volumeClaimTemplates: - spec: storageClassName: nice-provisioner OSD OSD Ceph cluster OSD Nice provisioner Provision on demand or pre-provisioned
29. In our cluster ▌NVMe SSD ⚫Use TopoLVM ⚫PV resources and the corresponding LVM logical volumes are created on OSD creation ▌HDD ⚫Use Local persistent volume ⚫Create PVs for all HDDs on deploying nodes ⚫PV resources are already available on OSD creation 29
30. Agenda 30 ▌Cybozu and our storage system ▌Quick introduction to K8s and Rook ▌Advanced Rook/Ceph cluster ▌Efforts and challenges ▌Conclusion
31. Daily check of upstream Rook/Ceph ▌Check every update of both Rook and Ceph projects every day ⚫Watch the important bugfix PRs and backport them if necessary ▌e.g. a data lost bug in RGW (PR#49795) ⚫Some object might be gone on bucket index resharding ⚫Pending resharding operation until the upgrade to >= v17.2.6 31
32. Upstream-first development ▌We’ve shared everything with Rook/Ceph communities ⚫ Reduce the long-term maintenance cost ⚫ Make both communities better ▌Major contributions ⚫ Implemented the Rook’s advanced configurations ⚫ I’ve been working as a Rook maintainer ⚫ Resolved some problems in containerized Ceph clusters ⚫ https://speakerdeck.com/sat/revealing-bluestore-corruption-bugs-in-containerized-ceph-clusters 32
33. Running the custom containers ▌If there is a critical bug and we can’t wait the next release, we want to use our custom containers ⚫ Official release + critical patches ▌Have being tried to run Teuthology in our test environment to verify custom containers ⚫ Succeeded in running all tests, but most tests still fail ⚫ We’ll continue to fix this problem 33
34. Other remaining work ▌Backup/restore ▌Remote Replication ▌More automation 34
35. Agenda 35 ▌Cybozu and our storage system ▌Quick introduction to K8s and Rook ▌Advanced Rook/Ceph cluster ▌Efforts and challenges ▌Conclusion
36. Conclusion ▌Rook is an attractive option of Ceph orchestration ⚫Especially if you are familiar with K8s ▌There are some advanced configurations ▌We’ll continue to provide feedbacks to Rook/Ceph communities 36
37. Thank you! 37