config.dev.yaml

# Development environment configuration

# Kubernetes namespace
namespace: krasa-ns

# Rancher project ID for user namespace annotation
rancherProjectId: "c-m-qvndqhf6:p-8rjpv"

# Global storage class for all persistent volumes
storageClassName: nfs-csi

# Container image registry. Helm charts are pushed under its /charts path.
registry: cerit.io/mddash
imageTag: dev

# Dashboard hostname
dashboard:
  hostname: mddash-dev.dyn.cloud.e-infra.cz

# OIDC provider for JupyterHub authentication
oidc:
  authenticator_class: "generic-oauth"
  oauth_server: "login.e-infra.cz"
  loginService: "e-INFRA CZ"  # the button label in UI
  authorizePath: oidc/authorize
  tokenPath: oidc/token
  userdataPath: oidc/userinfo
  usernameKey: preferred_username
  allowAll: true
  scope:
    - openid
    - profile
    - email

# MDRepo integration for publishing experiments
mdrepo:
  url: https://workflow-repo.test.du.cesnet.cz
  scopes: user:email

# MDPosit integration for publishing experiments
mdposit:
  url: https://mdrepo.eu/

# MetaDump API for extracting GROMACS TPR metadata before publishing
metadump:
  url: "https://gmd.ceitec.cz"

# Default notebooks repository (setup and analysis notebooks)
defaultNotebooksRepo: https://github.com/sb-ncbr/mddash-notebooks.git

# Notebook image (spawned by Dashboard pod)
notebook:
  image: mddash-notebook
  platform: linux/amd64

# MD analysis image used for on-demand analysis K8s jobs
analysis:
  image: cerit.io/mddash/mddb_wf:0.1.11-gmx2026

# GPU resource type used by mdrun-api jobs and Tuner workers
gpuType: "nvidia.com/mig-1g.10gb"

# Container images used for mdrun-api simulation jobs
gmxImage: "cerit.io/mddash/tuner-worker:gmx-2026.1_amber-26_ray-2.54.0_py-3.13"
amberImage: "cerit.io/mddash/amber:24"

# Landing page image (served at / on the dashboard hostname)
landing:
  image: mddash-landing

# Sidecar container images for user pod
sidecars:
  platform: linux/amd64
  pullPolicy: Always
  ui:
    image: mddash-ui
  proxy:
    image: mddash-proxy
  auth:
    image: mddash-auth
  api:
    image: mddash-api
  s3sync:
    image: mddash-s3sync
  mdrepoUploader:
    image: mddash-mdrepo-uploader

# S3 object storage configuration
s3:
  endpoint: https://s3.cl4.du.cesnet.cz

resources:
  singleuser:
    memory:
      guarantee: 512M
      limit: 4G
    cpu:
      guarantee: 0.2
      limit: 1
  notebookQuota:
    # Max concurrent notebook pods per user namespace
    maxConcurrent: 1
  notebook:
    # Idle timeout in seconds: kernels are culled and the server shuts down after this period
    idleTimeoutSeconds: 600
    # jupyter container — limits allow burst above scheduling baseline
    cpuRequest: "200m"
    memoryRequest: "512Mi"
    cpuLimit: "2000m"    # 10× request: interactive, prefer throttle over kill
    memoryLimit: "4Gi"   # 8× request: burst headroom for in-memory analysis

  analysisJob:
    # Dashboard mwf analysis jobs are mostly single-threaded; extra CPU limit rarely speeds them up.
    cpuRequest: "1000m"
    memoryRequest: "2Gi"
    cpuLimit: "1000m"
    memoryLimit: "8Gi"
  namespaceQuota:
    # requests = expected simultaneous usage; formula: see docs/resource-management.md
    requestsCpu: "2000m"
    requestsMemory: "5Gi"
    # limits = sum of all container limits at full load (1 notebook at 2x tier; 4x intentionally blocked)
    limitsCpu: "14000m"
    limitsMemory: "25Gi"
  pvcStorageSize: 10Gi

# JupyterHub image (hub + custom UI, includes EGI Check-in JWT login)
hub:
  image: mddash-hub
  resources:
    requests:
      cpu: "200m"
      memory: "512Mi"
    limits:
      cpu: "1"
      memory: "2Gi"

helm:
  package: mddash-dev

# mdrun-api service
mdrunApi:
  image: mdrun-api
  platform: linux/amd64
  storage:  # Persistent storage for mdrun-api database
    size: 2Gi
    # SQLite WAL needs a block-device-backed filesystem, not NFS.
    storageClassName: csi-ceph-rbd-du
  gunicorn:
    workers: 2
    threads: 4
  resources:
    limits:
      cpu: "500m"
      memory: "512Mi"
    requests:
      cpu: "100m"
      memory: "256Mi"
  polling:
    resources:
      limits:
        cpu: "250m"
        memory: "256Mi"
      requests:
        cpu: "50m"
        memory: "128Mi"
  jobHeadroom:
    # Maximum number of HPC jobs running simultaneously in the hub namespace.
    # Set cpuPerJob/memoryPerJob to the largest single job you want to allow.
    maxConcurrentJobs: 2
    cpuPerJob: "8"
    memoryPerJob: "16Gi"

# Tuner service
tuner:
  api:
    image: tuner-api
    pullPolicy: Always
    resources:
      limits:
        cpu: "1"
        memory: "2Gi"
      requests:
        cpu: "100m"
        memory: "256Mi"
    pricing:
      # Hourly resource rates (currency units) for trial cost estimation & early-stop pruning.
      cpuCoreHour: "0.04"
      gpuHour: "0.30"
      gbRamHour: "0.005"
  databaseStorage:
    size: 1Gi
    storageClassName: csi-ceph-rbd-du
  jobStorage:
    size: 10Gi
    storageClassName: nfs-csi
  ray:
    version: "2.54.0"
    head:
      resources:
        limits:
          cpu: "1"
          memory: "8Gi"
        requests:
          cpu: "250m"
          memory: "2Gi"
  worker:
    image: "cerit.io/mddash/tuner-worker:gmx-2026.1_amber-26_ray-2.54.0_py-3.13"
    pullPolicy: IfNotPresent
    maxReplicas: 1
    idleTimeoutSeconds: 3600
    resources:
      limits:
        cpu: "4"
        memory: "16Gi"
        nvidia.com/mig-1g.10gb: "1"
      requests:
        cpu: "4"  # should be same as limit
        memory: "8Gi"
        nvidia.com/mig-1g.10gb: "1"