Features and their limits
This map describes the implemented scope targeted for Replicove v0.3.0-alpha.2, including its owned-PVC cleanup fix. All 16 exact-source CI jobs and all 11 release jobs passed; validation records those gates and the separate scenario evidence. PostgreSQL, chaos, test recipes/pools, diagnostics and local agent interfaces were introduced in v0.3.0-alpha.1. That version's historical 16-job source pass at 93bcfbc remains separate from the new regression and published-artifact checks.
Use the ten executable scenarios and coverage matrix to try these feature families in disposable clusters. The scenarios cover supported workflows, not every possible API field combination or host distribution.
Installation and lifecycle
| Feature |
What it does |
Guide and limits |
| Installation |
Helm, CLI installer, or native YAML with protected key bootstrap |
Installation; all six CRDs share one operator image/chart |
| New vCluster |
Provisions a pinned persistent control plane or disposable lab profile |
Helm quickstart; one managed vCluster 0.37.1 runtime per destination |
| Existing vCluster |
Uses an explicitly registered guest with pinned identity and credentials |
Existing targets; no runtime adoption or deletion |
| Bounded lifetime |
TTL includes planning, admission, approval and provisioning; cleanup runs on expiry/deletion |
Cleanup; refresh/reset does not renew the lifetime |
| State persistence |
Encrypted captures, access, capacity and ownership records in a protected namespace |
Key persistence; back up key and state together |
| Optional modules |
Mirrors, PostgreSQL and chaos are disabled by default and can be enabled later |
Installation; update CRDs, image, values and RBAC together |
| Capacity and pools |
Durable destination reservations and grant caps; offline pool rendering and least-occupied destination selection |
Pools; selection is a hint, not a reservation; queue time consumes TTL |
Selecting and reproducing configuration
| Feature |
What it does |
Guide and limits |
| Administrator policy |
Bounds source namespaces/kinds, secrets, targets, profiles, TTL and guest roles |
Grants and RBAC; does not replace host RBAC |
| Selection and dependencies |
Namespace/name/label selectors, known references, mappings and explicit patches |
Selection; arbitrary operator dependencies and embedded strings are not inferred |
| Helm and operators |
Reconstructs selected Helm desired resources and values, CRDs and authorized operators |
Operators; no generic lifecycle-hook or external-side-effect cloning |
| Secrets |
Explicit snapshot/follow modes within exact administrator grants |
Secrets and storage; host service-account tokens and cloud identities are not cloned |
| Plan approval and evidence |
Approves a captured revision; reports source UID/resourceVersion, dependency paths, transformations and observed omissions |
Diagnostics; metadata evidence is not a historical replay bundle or an atomic source snapshot |
| Preflight |
Checks API discovery and caller authorization before provisioning |
Diagnostics; unknown results are explicit; no proof of operator permissions, CNI, CSI or cloud identity |
| Refresh and drift |
Recaptures source configuration; preserves foreign ownership and guest additions |
Lifecycle; database-enabled replicas require recreation instead of refresh |
| Ordinary PVCs |
Explicitly authorized empty destination claims |
Secrets and storage; copying contents requires a separate data grant |
Data copies
| Feature |
What it does |
Guide and limits |
| CSI mirrors |
Copies exact granted filesystem PVCs into independent writable generations |
Mirrors; qualified driver/classes and host NetworkPolicy enforcement required |
| Mirror reset and schedules |
Manual/scheduled latest-source sync or reset to a retained revision |
Mirrors; per-volume recovery points, no atomic multi-volume or database consistency guarantee |
| Mirror leases and retention |
Bounds testing and retained captures; access resolves one active generation |
Mirrors; whole-mirror TTL wins; reconnect after replacement |
| PostgreSQL 17 |
Consistent logical dump, isolated staging, sanitized second dump and new persistent target |
PostgreSQL; explicit database/source credentials, managed target only, no mirror combination or refresh |
| Approved masks |
Null, constant or salted text token masks with declared equality domains and relationship validation |
PostgreSQL; no general anonymization; undeclared fields may remain sensitive |
| Approved subsetting |
Equality filters delete rows only from explicitly selected staging tables before masking |
PostgreSQL; undeclared tables remain intact; no inferred tenant graph; invalid relationships reject publication |
| Data cleanup |
Durable ownership records, staging removal and owned storage cleanup |
Cleanup; no source deletion or claim of physical erasure of provider backups |
Repeatable tests and bounded faults
| Feature |
What it does |
Guide and limits |
| Test recipes |
replicove run creates a fresh replica or mirror, waits, obtains access, runs a local command and verifies cleanup |
Test runs; immutable TTL; no existing-target adoption; local command is not sandboxed |
| Reports |
Automatic metadata/provenance JSON and JUnit artifacts distinguish execution and cleanup outcomes |
Test runs; no raw captures, credentials, logs or exact source replay |
| Faults |
PodDelete, ScaleZero, NetworkIsolation, CPUStress, MemoryStress, CustomJob |
Chaos; 1–8 simultaneous faults, one active experiment per replica, exact owned targets and bounded duration/resources |
| Fault rollback |
Durable journals, cancellation/expiry rollback and verified Job cleanup |
Chaos; shared-worker restrictions, no privileged host/node faults; isolation/Jobs reject overlapping host allow policies |
Humans and agents
| Feature |
What it does |
Guide and limits |
| Guest authorization |
Expiring viewer/deployer/admin sessions and exact credential-Secret reader RBAC |
Access; host request permission and guest roles are separate |
| CLI and YAML |
Same Kubernetes resources for humans, CI and agents |
CLI, YAML, GitOps; no dedicated GitHub Action product integration |
| Stdio MCP |
Namespace-scoped listing, plan inspection and readiness waiting; mutations require --allow-write |
Agent interface; caller's Kubernetes identity; no remote transport, CA, certificate issuance or Secret-reading tools |
| Local dashboard |
Loopback-only, read-only lifecycle, plan, mirror and experiment views |
Dashboard; private session URL; no credentials/data payloads or shared hosted service |
| References |
Fields for all six CRDs plus command and operator help generated from source |
API, CLI, operator; TestRecipe and ReplicaPool are local CLI documents, not CRDs |
Not implemented or not qualified
- Automatic cloud identity exchange, generic external database/bucket/queue cloning, PITR, atomic multi-volume capture, and automatic relational tenant extraction.
- A remote multi-user MCP gateway, Replicove certificate authority, per-user request ownership, or hosted dashboard.
- Exact arbitrary-cluster cloning, independent host-kernel/worker isolation, vCluster Platform or distribution-specific certification.
- Production support and large-scale recovery qualification. Disposable results apply only to their recorded revisions, fixtures and host capabilities.
The implementation ledger, project status and roadmap track implementation and remaining gates separately.