Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

NRP Nautilus usage policies: research report

Checked 15 November 2026 against the official NRP documentation and its linked AUP. Completeness is not guaranteed: the site is large, and live namespace-specific admission rules, quotas, exceptions and enforcement configuration were not inspected. Quotes below are from the linked source; thresholds are as documented, not inferred from examples. The official llms.txt says its rules override conflicting examples.

Acceptable use, security and accountability

  • Access: “The first time you log in, you will be required to read and accept the NRP Acceptable Use Policy (AUP).” The AUP calls for re-acknowledgment every 12 months. Hierarchy; AUP (marked Draft 0.1).
  • Restricted data: “You agree to not introduce these types of data to NRP resources.” The AUP lists HIPAA, PII, CUI, FERPA, and other data protected by statute or data-use agreement. The landing page also names FISMA and PID. AUP.
  • Forbidden use: “You agree to not use NRP services nor allocated resources” for “financial gain”, “any unlawful purpose”, attempts to breach or circumvent controls, or violation of software/data agreements; comply with applicable law, including export controls. Cryptocurrency mining is not explicitly named in the AUP; mining for financial gain falls within its stated prohibition. AUP.
  • Credentials and incidents: “You agree to protect all accounts … and credentials”; “You agree to immediately report any known or suspected security breach or misuse of your access to security [at] sdsc.edu”. Users are responsible for backups. AUP violations may result in loss of access; illegal activity may be reported. AUP.
  • Publication acknowledgment: “In manuscripts submitted for publication and any other citable works, you agree to acknowledge NRP services and resource(s) contributing to research results” using the exact wording provided in the AUP.
  • Job-spec confidentiality: “MUST NOT put secrets, credentials or unpublished data in Job specs; every cluster user can read them.” NRP assistant rules; scheduling.

Workload type, lifetime and cleanup

  • Batch: “Batch computation MUST run as a Kubernetes Job whose command finishes on its own.” NRP assistant rules; jobs. Jobs are not subject to the documented two-week Deployment purge; the jobs page says they can run forever, provided their command ends by itself.
  • No interactive Jobs: “Running in interactive mode (sleep infinity command and manual start of computation) or any command that doesn’t end by itself is prohibited, and user can be banned.” The policy page states users running a Job with sleep infinity or equivalent will be banned. Jobs.
  • Bare interactive pods: “such pods will be destroyed in 6 hours unless you request an exception for your namespace”; “interactive pods are limited to 2 GPUs, 32 GB RAM, and 16 CPU cores.” Sleep is explicitly allowed in interactive pods. The exception example is an application controlling pods, such as JupyterHub. Policy.
  • Long-idle workloads: “Please don’t run such pods as Jobs”; “Such a deployment can not request a GPU.” Use a Deployment with minimal requests and Burstable QoS; sleep is explicitly shown for a pure shell in this Deployment context. Long idle; policy.
  • Deployment purge: “A periodic process removes workloads (deployments) older than 2 weeks”; “You will receive 3 notifications” before deletion; PVC data remains. Permanent services may request a namespace exception, but “Long idle pods can’t be added to the exceptions list”. Policy.
  • Job count and retries: “Do not submit more than 400 Jobs at once.” This is a Job submission count, not a verified spec.parallelism limit. Using Nautilus. “It’s a good idea to have [backoffLimit] more than 0”: recommendation, not a mandatory numeric value. Jobs.
  • Completed Jobs: “The pod and job will remain … for ttlSecondsAfterFinished=604800 seconds (1 week) by default, and you can adjust this value in your job definition if desired.” “Please make sure you did not leave any pods and jobs behind.” This is a documented default, not a mandate to set a particular TTL. Job tutorial.
  • JupyterHub: “A JupyterHub deployment MUST cull idle servers within 6 hours (timeout ≤ 21600) and MUST restrict sign-in with allowed_idps or allowed_users.” NRP assistant rules; JupyterHub.

Resource requests, utilization and GPUs

  • Requests and limits: “Your resource limits must be within 20% of the resource requests”, particularly CPU, memory and ephemeral storage. With “a large number (> ~100) of pods or jobs”, “resource limit = request.” Policy. The assistant rules require requests and limits for each container. Do not copy the more permissive ratios in example YAML as policy.
  • Usage violations: “A user can not submit more than 4 pods, violating” GPU utilization >40%, CPU usage 20–200%, and RAM usage 20–150% of requests. “These restrictions do not apply to pods requesting 1 CPU core and 2GB memory.” The precise interpretation of the exemption for values below those exact numbers and the measurement window are not specified. Policy.
  • Enforcement: “Admins monitor resource utilization and will ban namespaces with under-utilized resources requests.” “If a pod goes over its memory limit, it will be killed (OOM - Out of Memory).” Check the Violations page; the documentation does not specify the ban timing or averaging window. Policy.
  • GPU scaling: “The only reason to request more than a single GPU is when your GPU utilization is close to 100% and you can leverage using more GPUs.” For planned large jobs “(>50 GPUs) please present a plan in Nautilus Support.” Policy. The GPU page states up to 8 GPUs per node for Jobs, versus 2 for interactive pods; hardware descriptions elsewhere mention 16 GPUs per node, which is not a verified general per-Job entitlement.
  • Special GPUs: “By default every namespace has a quota of zero for each” of A100, H100, H200 and GH200. A100 quota can be requested; H100/H200/GH200 reserved quota is not user-requestable. “Any user, any of the four GPUs, no reservation” may use opportunistic or opportunistic2, bypassing quota but accepting preemption. GPU pods; priority classes.
  • Maximum ordinary Job pod size: Beyond the interactive caps and GPU-page statement above, a universal CPU/RAM/pod-size exception threshold was not verified in the reviewed documentation; do not assume one.

Storage

  • Retention: “You can only store the data actively used for computations”; “Any volume that was not accessed for 6 months can be purged without notification.” NRP storage must not be used as a long-term archive. Applicability of the volume-access rule to S3 buckets was not explicitly verified. Policy; purging.
  • CephFS: “Installing conda and pip packages on all CephFS (shared) filesystems is strictly prohibited!” Do not open the same file for writing from multiple parallel clients. For rook-cephfs-ucsd, data must serve current computation and be promptly deleted; admins may purge longer-lived data without notice at their discretion. RBD is preferred for data not requiring shared multi-pod access. Ceph.
  • Local scratch: “any information stored in it will be destroyed after pod shutdown.” Pods writing more than 50Gi per container can be evicted; explicitly request the scratch size when planning to exceed it. Disk starvation can evict all pods on a node. Local scratch. The local page calls the limit optional in that example; the stricter general resource rule calls for both request and limit.
  • S3: “The HA pool … is reserved for mission-critical services”; ask admins for credentials. General public pools use portal credentials. Ceph S3.
  • Large deletion: “In cephFS and RBD (not S3) … a folder with many (more than 10,000) files … might crash the metadata server”; the page recommends find-based deletion. Purging. Obtain authorization before deleting data.

Networking, images and scheduling

  • Ingress: “In general we don’t allow exposing non-http applications via TCP ports”; contact support for an exception. “The annotation to use is either haproxy-ingress.github.io or ingress.kubernetes.io. DO NOT use the haproxy.org annotation”. The example uses ingressClassName: haproxy; no exclusive class rule beyond the example was independently verified. Ingress.
  • Cross-namespace access: “To let other namespaces connect, add a NetworkPolicy in your namespace that selects the pods and allows them.” Once selected, retain rules for all required sources, including monitoring and ingress. Public LoadBalancer VIPs are an exception to the default cross-namespace block; use service authentication. Ingress.
  • LoadBalancer: “Check with cluster admins before reserving” a specific IPv4 address. A separate guide says admins can enable LoadBalancer services for a namespace; this is not a blanket verified ban on LoadBalancer. VIPs; virtualization example.
  • Tolerations: “Users may only tolerate values they are authorized for by cluster admins.” nautilus.io/system and nautilus.io/issue are prohibited to users; nautilus.io/reservation requires approved group membership; appropriate nautilus.io/hardware tolerations are permitted when needed. Unauthorized tolerations risk a ban. Special use; assistant rules. The GPU example still shows an older arm64 toleration; the special-use mapping identifies nautilus.io/hardware=arm64 as the new form.
  • Priority: “Everything else … is banned in user namespaces.” Allowed: unset, armada-default, owner-no-preempt, opportunistic2, opportunistic. Admission rejects banned classes; support can consider exemptions. Priority classes.
  • Images: A mandatory specific registry and numerical image-pull limit were not verified; examples using particular registries are not evidence of a mandate. Images tutorial; faster images.

Namespace administration and monitoring

  • Responsibility: “A Group Administrator is personally responsible for everything that runs in their Projects, and for keeping the member list current.” Projects map to Kubernetes namespaces. Hierarchy.
  • Namespace metadata and quotas: A universal required namespace annotation/label set or general CPU/RAM/storage quota was not verified. GPU-specific zero quotas and priority restrictions are documented above; actual namespace quotas may differ. Do not assume Kubernetes’ standard namespace-name label is an operator-mandated annotation.
  • Monitoring: “it’s your responsibility to make sure [jobs] are running as intended, without overrequesting the resources.” Monitor GPU >40%, and use Memory Usage (RSS) rather than cache-inclusive Memory Usage. Monitoring.
  • PodMonitor/Prometheus: A mandatory PodMonitor, Prometheus scrape interval, required metrics labels, or user-owned monitoring deployment was not verified. When a NetworkPolicy selects a pod, allow monitoring traffic if it must remain reachable. Ingress.

Documentation conflicts and limits

  • The long-idle guide includes sleep for Deployments, whereas the compact assistant rules say “sleep is allowed only” in interactive pods. The detailed long-idle page explicitly allows it for a GPU-free Deployment; seek support clarification before relying on this inconsistency.
  • Some examples omit resources or show request/limit ratios far outside 20%; the policy and assistant rules are stricter and should take precedence. Local-scratch examples also describe a limit as optional despite the general rule.
  • No claim above establishes actual live enforcement, exception approval, a blanket namespace suspension trigger, a universal backoffLimit/TTL mandate, a registry mandate, or a numeric Job parallelism cap. Ask Nautilus Support for current namespace-specific policy and unresolved ambiguities.