Production profile

The one Cluster configuration the project recommends for production until 1.0, the configuration every reliability test, alert and upgrade check runs on. Anything else is best effort.

The operate pages describe many valid configurations. Only one of them is the configuration the project recommends for production, and it is the one every reliability test, alert rule and upgrade check of a release runs on: a test matrix cannot cover an open set of combinations, it can cover one profile.

Until 1.0, the production profile is the configuration below. A Cluster install that departs from it is supported on a best-effort basis: bugs are fixed, but nothing in the release checklist proves that combination before a release ships.

The profile

AreaProduction profileWhy, and where the detail is
Edition and topologyCluster edition, non-split, replicaCount: 2, from values-ha.yamlA leader-elected scheduler with a warm standby and an active-active API. Split mode needs split.scheduler.replicaCount: 2 for the same failover (control-plane HA).
PlacementReplicas spread across nodes, and across zones on a multi-zone clusterTwo replicas on one node or in one zone are one replica (control-plane HA).
PostgresManaged Postgres 16 with high availability and point-in-time recoveryThe only durable state of the control plane. Direct connections, or PgBouncer in session mode, for every process that runs the scheduler: the leader election holds a session advisory lock (connection budget).
RedisManaged Redis with persistence, maxmemory-policy noeviction, reached through the provider’s failover endpointHolds XCom values of running DAGs and carries the live log tail (Redis requirements).
Task logsObject storage sink (logs.sink.provider: s3 or gcs), no log PVCA ReadWriteOnce log volume cannot follow a second replica (control-plane HA).
Task executionDedicated task podsWarm pools (execution.warmPoolsEnabled) are experimental and outside the profile (warm pools).
RetentionOn, with the starting configurationWithout it dag_runs, task_instances and audit_log grow without bound.
Monitoringmetrics.serviceMonitor.enabled and metrics.prometheusRule.enabledThe rules the chart ships page on the states that stop work silently, such as a schema that is not ready (helm chart). metrics.grafanaDashboard.enabled adds the overview dashboard (metrics).
VersionsA pinned chart version, whose appVersion pins the images, upgraded along the supported pathA minor is upgraded from the latest patch of the previous minor (versioning).
BackupsThe RPO and RTO targets of Cluster backupPostgres, Redis, the log bucket and the secrets are separate state, each with its own loss.

Install it

Start from the HA overlay, then set the datastores, secrets and bucket it marks CHANGEME:

helm upgrade --install dexaflow oci://ghcr.io/dexadata/charts/dexaflow \
  --version <x.y.z> -n dexaflow -f values-ha.yaml -f values-prod.yaml

with a values-prod.yaml that turns on what the overlay leaves to you:

metrics:
  serviceMonitor:
    enabled: true
  prometheusRule:
    enabled: true
  grafanaDashboard:
    enabled: true

and the retention settings through config or extraEnv. Leave image.tag and migrations.image.tag empty: they then follow the chart’s appVersion, so the pinned chart version also pins the images, and the release signature covers them (verify a release).

What best effort means

Outside the profile, for example a single replica, split mode, a log PVC, warm pools, Redis Sentinel or Redis Cluster, or a transaction-mode pooler in front of the scheduler:

  • the configuration is documented and expected to work, and a bug in it is a bug;
  • no reliability row of the release checklist and no row of the RC cluster validation runs on it, so a regression there can reach a release;
  • some combinations are refused at render time instead, with the reason, such as two replicas on a ReadWriteOnce log volume.

The profile changes only with a release, and the changelog names the change.