Nomad Enterprise is aimed at teams and organizations and addresses the organizational complexity of multi-team and multi-cluster deployments with collaboration and governance features.
This collection provides best practices and guidance for operating Nomad securely in a multi-team setting through features such as resource quotas and Sentinel.
The features explored in this collection are provided in the Enterprise versions of Nomad. In order to follow along you will need access to an environment running Nomad Enterprise with the Governance & Policy module. To explore Nomad Enterprise features, you can sign up for a free 30-day trial from here.
When many teams or users are sharing Nomad clusters, there is the concern that a single user could use more than their fair share of resources. Resource quotas provide a mechanism for cluster administrators to restrict the resources within a namespace.
Quota specifications are first class objects in Nomad. A quota specification has a unique name, an optional human readable description, and a set of quota limits. The quota limits define the allowed resource usage within a region.
Quota objects are shareable among namespaces. This allows an operator to define
higher level quota specifications, such as a
prod-api quota, and multiple
namespaces can apply the
prod-api quota specification.
Sentinel Policies - Policies are able to introspect on request arguments and use complex logic to determine if the request meets policy requirements. For example, a Sentinel policy may restrict Nomad jobs to only using the "docker" driver or prevent jobs from being modified outside of business hours.
Policy Scope - Sentinel policies declare a "scope", which determines when the policies apply. Currently the only supported scope is "submit-job", which applies to any new jobs being submitted, or existing jobs being updated.
Enforcement Level - Sentinel policies support multiple enforcement levels. The
advisorylevel emits a warning when the policy fails, while
hard-mandatorywill prevent the operation. A
soft-mandatorypolicy can be overridden if the user has necessary permissions.
Each Sentinel policy has a unique name, an optional description, applicable scope, enforcement level, and a Sentinel rule definition. If multiple policies are installed for the same scope, all of them are enforced and must pass.
Sentinel policies cannot be used unless the ACL system is enabled.
Sentinel policies specify an applicable scope, which limits when the policy is enforced. This allows policies to govern various aspects of the system.
The following table summarizes the scopes that are available for Sentinel policies:
|Applies to any jobs (new or updated) being registered
Sentinel policies specify an enforcement level which changes how a policy is enforced. This allows for more flexibility in policy enforcement.
The following table summarizes the enforcement levels that are available:
|Issues a warning when a policy fails
|Prevents operation when a policy fails, issues a warning if overridden
|Prevents operation when a policy fails
sentinel-override capability is required to override a
policy. This allows a restricted set of users to have override capability when
Nomad supports multi-datacenter and multi-region configurations. A single region is able to service multiple datacenters, and all servers in a region replicate their state between each other. In a multi-region configuration, there is a set of servers per region. Each region operates independently and is loosely coupled to allow jobs to be scheduled in any region and requests to flow transparently to the correct region.
When ACLs are enabled, Nomad depends on an "authoritative region" to act as a
single source of truth for ACL policies, global ACL tokens, and Sentinel
policies. The authoritative region is configured in the
server stanza of
agents, and all regions must share a single authoritative source. Any Sentinel
policies are created in the authoritative region first. All other regions
replicate Sentinel policies, ACL policies, and global ACL tokens to act as local
mirrors. This allows policies to be administered centrally, and for enforcement
to be local to each region for low latency.
Click the links below to jump to other topics: