Cloudera Machine Learning requirements (OCP)

To launch the Cloudera Machine Learning service, the OpenShift Container Platform (OCP) host must meet several requirements. Review the following CML-specific software, NFS server, and storage requirements.

Requirements

If necessary, contact your Administrator to make sure the following requirements are satisfied:
  1. If you are using OpenShift, the installed OpenShift Container Platform must be version 4.10 or 4.8 (when upgrading from 1.4.x to 1.5.x, not performing a fresh installation). For ECS, refer to the Hardware and Software Requirements section in CDP Private Cloud Experiences Installation Hardware Requirements and Managing a Private Cloud Experience Cluster 1.5.x.
  2. CML assumes it has cluster-admin privileges on the cluster.
  3. Storage:
    1. Persistent volume block storage per ML Workspace: 600 GB minimum, 4.5 TB recommended.
    2. 1 TB of external NFS space recommended per Workspace (depending on user files). If using embedded NFS, 1 TB per workspace in addition to the 600 GB minimum, or 4.5 TB recommended block storage space.
    3. Access to NFS storage is routable from all pods running in the cluster.
    4. For monitoring, recommended volume size is 60 GB.
  4. On OCP, CephFS is used as the underlying storage provisioner for any new internal workspace on PVC 1.5.x. A storage class named ocs-storagecluster-cephfs with csi driver set to "openshift-storage.cephfs.csi.ceph.com" must exist in the cluster for new internal workspaces to get provisioned.
  5. A block storage class must be marked as default in the cluster. This may be rook-ceph-block, Portworx, or another storage system. Confirm the storage class by listing the storage classes (run oc get sc) in the cluster, and check that one of them is marked default.
  6. If external NFS is used, the NFS directory and assumed permissions must be those of the cdsw user. For details see Using an External NFS Server in the Related information section at the bottom of this page.
  7. If CML needs access to a database on the CDP Private Cloud Base cluster, then the user must be authenticated using Kerberos and must have Ranger policies set up to allow read/write operations to the default (or other specified) database.
  8. Ensure that Kerberos is enabled for all services in the cluster. Custom Kerberos principals are not currently supported. For more information, see Enabling Kerberos for authentication.
  9. Forward and reverse DNS must be working.
  10. DNS lookups to sub-domains and the ML Workspace itself should work.
  11. In DNS, wildcard subdomains (such as *.cml.yourcompany.com) must be set to resolve to the master domain (such as cml.yourcompany.com). The TLS certificate (if TLS is used) must also include the wildcard subdomains. When a session or job is started, an engine is created for it, and the engine is assigned to a random, unique subdomain.
  12. The external load balancer server timeout needs to be set to 5 min. Without this, creating a project in an ML workspace with git clone or with the API may result in API timeout errors. For workarounds, see Known Issue DSE-11837.
  13. If you intend to access a workspace over https, see Deploy an ML Workspace with Support for TLS.
  14. For non-TLS ML workspaces, websockets need to be allowed for port 80 on the external load balancer.
  15. Only a TLS-enabled custom Docker Registry is supported. Ensure that you use a TLS certificate to secure the custom Docker Registry. The TLS certificate can be self-signed, or signed by a private or public trusted Certificate Authority (CA).
  16. On OpenShift, due to a Red Hat issue with OpenShift Container Platform 4.3.x, the image registry cluster operator configuration must be set to Managed.
  17. Check if storage is set up in the cluster image registry operator. See Known Issues DSE-12778 for further information.

For more information on requirements, see CDP Private Cloud Base Installation Guide.

Hardware requirements

Storage

The cluster must have persistent storage classes defined for both block and filesystem volumeModes of storage. Ensure that a block storage class is set up. The exact amount of storage classified as block or filesystem storage depends on the specific workload used:
  • Machine Learning workload requirements for storage largely depend on the nature of your machine learning jobs. 4 TB of persistent volume block storage is required per Machine Learning Workspace instance for storing different kinds of metadata related to workspace configuration. Additionally, Machine Learning requires access to NFS storage routable from all pods running in the cluster (see below).
  • Monitoring uses a large Prometheus instance to scrape workloads. Disk usage depends on scale of workloads. Recommended volume size is 60 GB.
Local Storage (for example, ext4) Block PV (for example, Ceph or Portworx) NFS (for ML user project files)
Control Plane N/A 250 GB N/A
CML N/A 1.5 TB per workspace 1 TB per workspace (dependent on size of ML user files)

NFS

Cloudera Machine Learning (CML) requires NFS 4.0 for storing project files and folders. NFS storage is to be used only for storing project files and folders, and not for any other CML data, such as PostgreSQL database and LiveLog.

ECS requirements for NFS Storage

Cloudera managed ECS deploys and manages an internal NFS server based on LongHorn which can be used for CML. This is the recommended option for CML on ECS clusters. CML requires nfs-utils in order to mount longhorn-nfs provisioned mounts.

CML requires the nfs-utils package be installed in order to mount volumes provisioned by longhorn-nfs. The nfs-utils package is not available by default on every operating system. Check if nfs-utils is available, and ensure that it is present on all ECS cluster nodes.

Alternatively, the NFS server can be external to the cluster, such as a NetApp filer that is accessible from the private cloud cluster nodes.

OpenShift requirements for NFS storage

An internal user-space NFS server can be deployed into the cluster which serves a block storage device (persistent volume) managed by the cluster’s software defined storage (SDS) system, such as Ceph or Portworx. This is the recommended option for CML on OpenShift. Alternatively, the NFS server can be external to the cluster, such as a NetApp filer that is accessible from the private cloud cluster nodes. NFS storage is to be used only for storing project files and folders, and not for any other CML data, such as PostgreSQL database and LiveLog.

CML does not support shared volumes, such as Portworx shared volumes, for storing project files. A read-write-once (RWO) persistent volume must be allocated to the internal NFS server (for example, NFS server provisioner) as the persistence layer. The NFS server uses the volume to dynamically provision read-write-many (RWX) NFS volumes for the CML clients.