Production deployment: versions, storage, network, and acceptance

This page adds deployment decisions and acceptance checks against a v4.2.5 baseline. It does not certify arbitrary Kubernetes/Operator/storage combinations. Record the following before using the cluster deployment steps.

Reproducible version inventory

Layer

Record

Evidence

MatrixOne

Tag, digest, platform, configuration source/checksum

Release, actual image, runtime SELECT VERSION

Operator chart

Version, package checksum, complete values

Pinned Chart.yaml, values.yaml, rendered manifests

Controller/dependencies

Operator/OpenKruise/backup image tags and digests

Rendered manifests and running Pod imageIDs

Kubernetes

Server/client versions, CNI, CSI, StorageClass

Inventory and component support matrices

Storage/recovery tools

S3/MinIO, backup tool/format, encryption conditions

Vendor guidance and same-version restore drill

Chart, controller image, and MO versions are separate values. The official chart 1.3.0-alpha.4 downloaded on 2026-10-14 renders the default controller image matrixorigin/matrixone-operator:1.2.0-alpha.5, plus matrixorigin/mobr:1.0.0-rc1 and OpenKruise dependencies. A chart tag alone does not identify the controller and does not establish certification with MO v4.2.5.

This review linted and rendered that chart using Helm 3.19.0 without installing it on Kubernetes. Validate the chosen combination rather than treating “latest” as a support matrix:

MO_OPERATOR_CHART=/path/to/pinned/matrixone-operator
helm show chart "$MO_OPERATOR_CHART"
helm show values "$MO_OPERATOR_CHART"
helm lint "$MO_OPERATOR_CHART"
helm template mo-op "$MO_OPERATOR_CHART" -f approved-values.yaml > rendered.yaml
kubectl get crd matrixoneclusters.core.matrixorigin.io
kubectl explain matrixonecluster.spec --api-version=core.matrixorigin.io/v1alpha1
kubectl apply --dry-run=server -f mo.yaml

Prepare approved-values.yaml and mo.yaml against the selected versions; this page does not provide directly applicable manifests. Rendered files may contain Secrets: inspect them in a controlled location rather than posting them publicly. Verify CRD upgrades and existing-resource compatibility separately; rolling back a Helm revision does not establish CRD rollback compatibility.

Failure domains and capacity

  • Standalone deployment, multiple CNs, or three machines do not automatically provide HA. Assess Kubernetes control plane, Logservice quorum, TN state recovery, CN/Proxy endpoints, and object storage separately.

  • Schedule replicas across independent nodes/zones. Multiple Pods on one machine do not isolate machine failure. CN scaling does not imply uninterrupted availability during arbitrary TN/storage failures.

  • Size CPU/memory requests and limits, disk capacity/IOPS, cache/spill capacity, and network separately. Hardware tables are starting points, not data-volume or concurrency performance guarantees. Benchmark actual SQL, transaction sizes, indexes, and vector/fulltext workloads.

  • New initialization differs from taking over existing data. Preserve UUIDs, Logservice hostname/identity, backend, and paths; do not overwrite old data configuration with newer examples blindly.

Storage and network checks

Object

Before deployment

S3/MinIO

DNS from required Pods, TLS/CA, region, credentials, bucket/prefix, permissions, recovery targets

Local/PVC

Persistent state versus rebuildable cache; mount paths, capacity, StorageClass, topology, retention, access

Data format

FileService backend matches existing files; v4.2.5 examples use DISK-V2 and must not blindly replace existing DISK configurations

SQL entry

Client routing to Proxy/CN, 6001 or configured port, identity, TLS; 6001 is not all internal communication

Internal traffic

Version-specific Logservice/Raft, HAKeeper, CN/TN RPC/discovery addresses; cross-node DNS and routes

Operations entry

Restrict metrics/debug/pprof and Kubernetes management access; SQL connectivity does not justify public management access

Distinguish prefixes and permissions for shared objects, backups, exports, and CDC state within a bucket. PVC deletion and object deletion differ; Retain is not an independent backup. Never use the entire shared-data prefix as a routine cleanup target.

Acceptance

MO_NAMESPACE=mo
MO_CLUSTER=mo
kubectl get matrixonecluster "$MO_CLUSTER" -n "$MO_NAMESPACE" -o wide
kubectl get pods,pvc,svc -n "$MO_NAMESPACE" -o wide
kubectl get events -n "$MO_NAMESPACE" --sort-by=.lastTimestamp
kubectl describe matrixonecluster "$MO_CLUSTER" -n "$MO_NAMESPACE"

Replace names/namespaces. Check Pod Running, resource Ready, and business correctness separately:

  1. Stable resource conditions, approved images/configuration, no persistent restarts, Pending state, or storage errors.

  2. Authentication through the real client entry and test-table transactional writes/reads. With multiple CNs, verify visibility and failures across connections/CNs.

  3. Baseline representative latency, errors, throughput, resources, and storage. Record cold-cache and steady-state behavior separately.

  4. In a test topology, drill endpoint switching and CN/TN/Logservice/storage failure and recovery. Measure interruption; replica counts do not prove zero downtime.

  5. Restore to an independent target, verify counts, key aggregates, privileges, and external dependencies, then accept RPO/RTO.

Mark combinations without compatibility, failure-drill, and restore evidence as unverified. Helm lint, YAML rendering, and single-CN tests do not replace production acceptance.