On-prem Technical Requirements

Minimum System Requirements

Fiddler is horizontally scalable to support the throughput requirements for enormous production use-cases. The minimum system requirements below correspond to approximately 20 million inference events monitored per day (~230 EPS) for models with around 100 features, with 90 day retention.

  • Deployment: Kubernetes namespace in AWS, Azure or GCP
  • Compute: A minimum of 96 vCPU cores
  • Memory: 384Gi
  • Persistent volumes: 500 Gi storage across 10 volumes
    • POSIX-compliant block storage
    • 125 MB/s recommended
    • 3,000 IOPS recommended
  • Container Registry: Quay.io or similar
  • Ingress Controller: Ingress-nginx or AWS/GCP/Azure Load Balancer Controller
  • DNS: FQDN that resolves to an L4 or L7 load balancer/proxy that provides TLS termination

Kubernetes Cluster Requirements

As stated above, Fiddler requires a Kubernetes cluster to install into. The following outlines the requirements for this K8 cluster:

  • Node Groups: 2 node groups - 1 for core Fiddler services, 1 for Clickhouse (Fiddler's event database)

  • Resources:

    • Fiddler : 48 vCPUs, 192 Gi
    • Clickhouse : 64 vCPUs, 256 Gi [tagged & tainted]
  • Persistent Volumes: 500 GB (minimum) / 1 TB (recommended)

  • Instance Sizes

    Instance SizeAWSAzureGCP