Skip to main content
Version: 3.1.0 (Preview)

Operations And Recovery

Operate on the explicit cluster and the exact saved operation. A successful HTTP submission, a zero process exit, a log message or an unrelated completed request does not prove that the intended workflow finished. Use the Server's structured state and identifiers.

This runbook complements cluster creation and managed HBase dependencies. The architecture guide distinguishes historical runtime evidence from remaining qualification.

Read-only API Examples​

The following are HTTP request targets for an already authenticated API client. Replace the example cluster name and identifier placeholders with values returned for your operation; preserve the deployment's authentication and proxy settings. Do not put credentials in scripts or shared diagnostic output.

GET /api/v1/clusters?fields=Clusters/cluster_id,Clusters/cluster_name,Clusters/provisioning_state
GET /api/v1/clusters/analytics-a/hosts?fields=Hosts/host_name
GET /api/v1/clusters/analytics-a/services/HBASE/dependencies
GET /api/v1/clusters/analytics-a/services/HBASE/dependencies/{bindingId}
GET /api/v1/clusters/analytics-a/services/HBASE/dependencies/deployments/{deploymentId}
GET /api/v1/clusters/analytics-a/requests/{requestId}
GET /api/v1/clusters/analytics-a/requests/{requestId}/tasks
ObservationRead and validate
Cluster inventoryClusters.cluster_id and Clusters.cluster_name; a name change must preserve the numeric identity.
Binding listitems; ownership may be managed, local, unmanaged or unknown. A missing managed binding does not prove a local dependency is healthy.
Binding detailbinding_id, consumer, provider, phase, row_version, operation_epoch, desired/applied snapshot versions and operation identity.
Binding readinessreadiness identifies required, prepared and verified daemon hosts; capabilities, allowed_actions and next_action describe permitted progression.
Deploymentdeployment_id, cluster_id, attempt_id, request_id, history, state, phase, failure_code, retry_allowed, completed and install_only.
Task completionRead the exact request's tasks and owning cluster. Confirm the task/host/operation association and terminal results rather than searching all recent requests.

A structured dependency error has code and message. Branch on the HTTP status and exact error code; the message is diagnostic text. Missing fields, malformed values or foreign identities must remain an error or unresolved observation.

Mutation Contract For Integrators​

Prefer the UI for ordinary administration. For automation, use the pinned API client and REST service as the payload contract.

In the table, D means /api/v1/clusters/{clusterName}/services/HBASE/dependencies.

Method and resourceContract
GET D/candidates?type=HDFSDiscover authorized provider candidates for an existing HBase consumer; ZooKeeper uses its own dependency type.
POST D/previewPreview one selection or the complete selection set; not approval or provisioning.
POST DApprove exact binding IDs, provider identity, preview schema version and returned fingerprints, with an operation UUID. Supports a one- or two-item collection.
POST D/deployments/{deploymentId}Persist/launch exact targets and an explicit install-only mode under an immutable deployment UUID.
POST D/deployments/{deploymentId}/actions/retryRetry that deployment with an operation UUID only when the returned capability permits it.
GET D/{bindingId}/preview-updateRead a fresh update proposal before approving provider configuration changes.
POST D/{bindingId}/actions/updateSubmit the fresh preview's row/snapshot versions, fingerprints, approval and schema version with the operation UUID.
POST D/{bindingId}/actions/retrySubmit an operation UUID and the expected current row version.
POST D/{bindingId}/actions/verify-credentialsVerify credentials against the expected current epoch, when the workflow permits manual credential handling.
DELETE D/{bindingId}Detach with an operation UUID and expected row version after satisfying lifecycle guards; does not erase provider data.

Draft and Add Service discovery use /api/v1/service-dependencies/candidates and /api/v1/service-dependencies/preview, with the appropriate draft or service-plan scope and expected revision. Do not create a dummy HBase service simply to make discovery work.

Provider-side reads use /api/v1/clusters/{clusterName}/services/{serviceName}/dependents and /api/v1/clusters/{clusterName}/services/{serviceName}/dependency-impact. Stop/Restart confirmation is bound to the current impact revision and exact provider/action. Re-read impact if another consumer changes.

Persist launch/retry IDs before submission. Following a timeout, read the original resource before deciding whether to retry; retain an immutable retry identity for the same attempt. An acknowledged deployment later returning 404 must not trigger a replacement launch.

Binding Phases​

PhaseOperator interpretation
PREVIEWEDThe binding has not established provider preparation.
PROVIDER_PREPARINGProvider work is in progress; inspect its associated request.
PROVIDER_PREPAREDPreparation is available; consumer installation and verification still have gates.
ZOOKEEPER_HANDOFF_RECONCILINGFollow the specific ZooKeeper handoff reconciliation; do not treat it as ready.
CONSUMER_VERIFYINGCurrent client, identity and connectivity evidence is being collected.
READYBinding readiness is established for the current approved inputs; deployment completion is a separate observation.
STALEApproved inputs/evidence no longer match current provider or consumer state; review changes.
FAILEDInspect failure details and permitted retry actions.
FENCING_UNCERTAINExisting operation ownership cannot be established safely; investigate before further mutation.
DETACHING, RETIRED, DETACHED, TOMBSTONEDRetirement/detach lifecycle phases; none imply automatic data deletion or permission to restart the old dependency.

Use returned capabilities with the phase. A phase name alone is not a substitute for topology, snapshot, identity and permission checks.

Recovery Cases​

Symptom or exact codeInspectRecovery
Installer refresh or lost create responseOriginal draft UUID, owner, revision and saved cluster association.Resume the original draft; re-enter requested secrets. Do not adopt a same-name cluster without the association.
Workflow revision conflictLatest saved scope/revision and another active editor.Reload the authoritative checkpoint, review differences, then submit current edits.
DEPENDENCY_OPERATION_STALECurrent operation, row version, snapshot and epoch.Reload; preview again when inputs changed. Do not replay an old approval against new state.
DEPENDENCY_VERSION_UNSUPPORTEDBoth active repository records and resolved version metadata.Select a compatible provider or correct repositories; do not bypass the validator.
DEPENDENCY_SECURITY_MISMATCH or CROSS_REALM_NOT_SUPPORTEDConsumer/provider security modes and realms.Use a supported topology; do not copy provider keytabs to bypass the mismatch.
DEPENDENCY_CONSUMER_NOT_READYEvery required daemon host's package, config, credentials and verification evidence.Repair the failing prerequisite and use the offered preparation/verification action.
DEPENDENCY_NAMENODE_RPC_FAILED or DEPENDENCY_DATANODE_READ_WRITE_FAILEDConfigured endpoints, network, identity and exact structured probe result.Correct the reported cause, then reverify the affected current plan.
DEPENDENCY_ZOOKEEPER_SESSION_FAILED or DEPENDENCY_ZOOKEEPER_AUTHORIZATION_FAILEDEnsemble connectivity, client identity and private-znode policy.Correct connectivity/authorization and follow the supported verification path.
DEPENDENCY_FENCING_UNCERTAINPersisted operation and provider ownership records.Preserve uncertain state and obtain operator investigation; never delete journals to force success.
Deployment UNRESOLVEDExact deployment history and missing/mismatched request/task lineage.Preserve evidence and investigate; retry only if the Server explicitly permits it.
Provider stop/delete blockedCurrent dependents and impact; consumer stop/detach requirements.Coordinate maintenance, review the current impact or retire/detach consumers through supported flows.
One cluster's directory data failsThat cluster's authorization and failed read.Retry that scope without discarding successful authorized rows.

Capture exact cluster, draft, binding, deployment, operation, epoch, request and task IDs when relevant, plus sanitized structured responses. Logs can help explain failure but must not establish workflow success. Never include credentials or raw keytabs in an issue.

Blueprint Handoff​

A managed Blueprint declares required dependency types through its settings:

{
"settings": [
{
"managed_dependencies": [
{"consumer_service": "HBASE", "dependency_type": "HDFS"},
{"consumer_service": "HBASE", "dependency_type": "ZOOKEEPER"}
]
}
]
}

Provision with provision_action: PREPARE_ONLY, wait for actual host placement and topology configuration, then preview the complete HBase selection, apply reviewed configuration, approve the exact binding fingerprints and launch a durable deployment with exact targets. Provider IDs and credentials do not belong in the reusable template. The ordinary local Blueprint path retains its default actions.

This is the handoff contract, not a complete provisioning payload. Use the Blueprint implementation and the general Blueprint guide when building an integration.

Upgrade And Acceptance Work​

Back up the shared database and Server configuration before an upgrade. Validate exclusive host ownership and the scoped-workflow/dependency schema migration. Do not “repair” duplicates by deleting membership rows without establishing the correct owner.

Before relying on a deployment, verify two-cluster host and permission isolation, two-tab navigation, exact draft recovery, per-host managed readiness and the provider maintenance paths you will use. Secure installations additionally need real KDC/credential tests. Exercise network loss and Server restart using saved operation IDs.

Still-needed release documentation includes the final artifact/tag matrix, supported upgrade/database combinations, real-KDC results, managed detach/data-retention evidence, and failure-injection/broker-revocation results. Add those only after the corresponding tests run; the historical source review is not a substitute.