Skip to main content
Version: 1.0.0 (development)

Review a production deployment

Use these questions to describe the deployment you actually operate: its workload, topology, failure tolerance, and recovery procedure. Answers should identify the responsible operator and the observation supporting the decision. There is no universal replica count, storage size, or throughput figure that establishes production readiness for every workload.

Workload and topology

QuestionWhat to establishSupporting guide
Which message semantics does the application need?Normal/ordered/transaction/POP/delayed paths actually used; duplicate handling; completion and retry behaviorDelivery and retry, capabilities
Is this a sharded or replicated topology?Broker names identify groups; replica IDs belong to one group. Different Broker names alone do not provide copies of the same dataMulti-node layout
What permits a successful write?Flush policy, local durability, in-sync replica requirement, and Controller authority where enabledHA design
Which failure can the cluster tolerate while accepting writes?Lost Broker, Controller voter, disk, or host; correlated failure placement; two-of-two replication can stop writes on one replica lossHA deployment
Can applications reach every advertised endpoint?Route-returned Broker addresses, NameServer alternatives, Proxy address/TLS name, DNS and firewall behavior from the application networkKubernetes deployment, Proxy deployment

Record the chosen topology, host/zone placement, service identity, ports, and data roots in the deployment's own inventory. Keep the write availability decision consistent with the application's timeout and retry behavior. A timed-out send may have been accepted; retry safety belongs to the business operation.

Security and state

QuestionWhat to establishSupporting guide
Who can access each public, peer, health, and management endpoint?Actual network exposure and encryption per hop, not a single global TLS assumptionDeployment security
Are application, inner-client, and operator identities separate?Required allow/deny cases, outbound signing, receiver permissions, protected ACL/snapshot files, credential rotationDeployment security
Where does authoritative state survive process or Pod replacement?CommitLog paths, Broker metadata, offsets, identity, timer/transaction state, Controller storage, and PVC retentionBackup and recovery
Can one process accidentally open another process's data?Exclusive store ownership, stable identity-to-volume mapping, no shared writable root between replicasStorage design
Is the selected backend supported by the built artifact?Required features and native dependencies; authoritative log and derived-state boundaries; no assumed automatic backend migrationStorage backends

Document where secrets are stored without copying their values into the inventory. Keep recovery access available to the designated operators even when ordinary service authentication is impaired.

Capacity and visibility

QuestionWhat to establishSupporting guide
What is the retained data volume per replica?Stored bytes/message, sustained and peak rates, retention, retry/DLQ/timer backlog, indexes and operational headroomCapacity and performance
How quickly can consumers recover from lag?Available drain rate above ingress, downstream processing capacity, ordering/queue limitsCapacity and performance
Can operators distinguish the failing layer?Application latency/errors, Broker storage and replication, consumer progress, Controller quorum, Proxy admissionMonitoring, troubleshooting
Does telemetry actually leave the process?Compiled exporter features, effective runtime selection, collector reachability, bounded labels, shutdown flushObservability configuration
What triggers action?Workload-specific thresholds and a linked operator response; distinguish missing telemetry from a healthy zeroMonitoring

Resource limits and probe thresholds should accommodate observed startup recovery and shutdown duration. CPU/memory requests in example values are starting inputs, not measured sizing recommendations.

Maintenance and recovery

QuestionWhat to establishSupporting guide
What happens during a planned restart?Draining, admitted work, surviving write requirements, Controller majority, Pod disruption rules, and catch-up before the next nodeMaintenance
What is in the backup, and what consistency does it have?A coordinated state set or stopped-source copy, independent storage, identity and configuration needed for restoreBackup and recovery
Has restore been tried on an isolated destination?Recovered message bounds, offsets, permissions, routing, application reconciliation, measured elapsed timeBackup and recovery
Can the previous version read the resulting data and metadata?Storage/config/API compatibility, downgrade-preflight limits, and whether rollback means restoring an earlier state setUpgrade and rollback
Who responds when the expected result is absent?Operator contacts, environment access, diagnostic commands, and decisions to stop a rollout or restoreAdmin operations

State the scope of each exercise. A successful chart render, configuration parse, process health probe, or local first-message example does not demonstrate distributed recovery. Report RPO/RTO only for the failure and restore scenario actually measured; unresolved assumptions remain visible in the deployment record.