Service Information

Prompt Keda service information.

Architecture

KEDA is deployed with the KEDA operator in the Prompt Processing vClusters in the keda Kubernetes namespace.

KEDA controls autoscaling for Prompt Processing by replacing the Kubernetes Auto Scaler. KEDA uses scalars to determine how aggressively to scale up and down. Prompt KEDA is configured to autoscale based on the number of Fanned Out events in Prompt Redis. This is called lag.

KEDA is not configured to scale down Scaled Jobs while they are running. There is a max run time setting that can be configured in KEDA, but it is not configured because Prompt Processing is configured to consume multiple rounds of Fanned Out messages before finishing. This is to reduce churn from new Pods being created.

Architecture Diagram

See Architecture Diagram

Associated Systems

See Associated Systems

Configuration Location

Config Area

Location

Configuration

lsst-sqre/phalanx

Vault Secrets Dev

secret/rubin/usdf-prompt-processing-dev/prompt-keda

Vault Secrets Prod

secret/rubin/usdf-prompt-processing/prompt-keda

Data Flow

See Data Flow

Dependencies - S3DF

Below are the S3DF Dependencies.
  • Kubernetes

  • SLAC LDAP to authenticate to vCluster

  • Prompt Redis for creating and scaling scaled jobs.

Dependencies - External

Below are external dependencies.
  • Internet access to pull Keda docker image.

Disaster Recovery

The application can be redeployed with not data being needed to be restored.