Skip to content

Roadmap

Where the project is going, by release. This page is the only place that orders work and names phases; the backlog holds the outcomes and acceptance criteria, the tech-debt register the known gaps, and the architecture the design. Dates are not promised.

Want to help? Items marked help wanted are good places to start; ask in the linked issue or in Discussions before you begin a large one.

Where we are

  • v0.1.1, v0.2.0, v0.3.0 and v0.3.1 are released: a multi-arch image, the Helm chart, install.yaml and crds.yaml, signed with keyless cosign, from the GitHub releases. Since v0.2.0 a release also carries an examples bundle, and since v0.3.0 the chart has a GPG Helm provenance file. The changelog lists the changes of each release; v0.1.0 has no GitHub release.
  • The API group domain openbaoconfig.io is final (ADR 0008).
  • 74 built-in kinds implemented and integration-tested against bao server -dev, and 52 experimental plugin kinds behind --enable-plugin-kinds (kind catalogue).
  • Engine: path-index ownership with a server-side marker, Create/Orphan defaults, the destructive-delete guard, drift tiers with a per-server request budget, observe mode with the obc-import CLI, the write circuit breaker and pause (architecture).
  • Evidence: API coverage per OpenBao release (92.2% of configuration paths typed on 2.7.2, built-in kinds only), the conformance run (no-write and default tests per kind: on pull requests for the kinds a change touches, every kind weekly and on main), the two-tenant isolation e2e test, and a load test up to 5,000 objects with zero writes at rest.
  • The chart enables NetworkPolicy by default. Set explicit API server and OpenBao peers; empty peer lists leave egress unrestricted by destination on the configured ports (hardening guide).

Phases

The kind catalogue tags each kind with one of these phases. A phase ends when its exit test passes in CI. Phases describe exit criteria, not build order: the kinds of phases 2, 3, 4 and 7 were built before the current engine and are retrofitted to its rules as the engine lands.

phase scope exit test release state
1 engine and tenant boundary: OpenBaoServer, connections, field classes, ownership and path index, observe mode, write-once fields, write circuit breaker, Policy, OpenBaoEndpoint, chart, CI no-write and default tests for built-in kinds; a namespaced connection cannot reach the operator identity (e2e); the self-init example boots with no root token left next conformance done (no-write and default tests on pull requests); operator-identity e2e done (test/e2e/isolation_test.go); self-init CI gate open
2 tenant and cluster onboarding: OpenBaoNamespace, mounts, Kubernetes, JWT and AppRole auth, TokenRole, KubernetesSecrets* one commit onboards a tenant namespace and a remote cluster, including dynamic tokens next kinds done; exit test open
3 identity with the ownership marker wiping status then reconciling makes zero writes for marked kinds next kinds and marker done; exit test open
4 secrets engines, the rest of sys, OIDC and MFA 10,000 objects hold a 1 h drift interval inside the request budget after next kinds done; scale test open
5 plugin support: Plugin registers OCI plugins, mounts wait for it the LDAP plugin from OCI and OpenLDAP in CI later open
6 plugin-backed kinds against real backends, LDAP, Kerberos and RADIUS first each kind runs against its backend in CI before it is un-gated later kinds in code, experimental and gated
7 access models, CertAuthRole, UserpassUser kinds integration-tested done done

after 6 in the catalogue marks plugin kinds outside the core plan (cloud, GitHub, Okta, OCI, CF, Consul, Nomad, Terraform Cloud, MongoDB Atlas, Quay); they stay gated until each has a plugin and backend in CI.

v0.1.1 to v0.3.1: released

v0.1.1 was the first complete GitHub release, then v0.2.0 (the destructive-delete guard for key material, faster drift reads, the documentation site) and v0.3.0 (the signed chart and the role tutorials). v0.3.1 tightened chart value validation and namespace-scoped watches; it made no CRD changes. The API group domain was decided before v0.1.0 (BL-001). The goals in the next two sections were planned for v0.2 and v0.3; those releases shipped without them. Still open from the v0.1 plan:

item source status
kubectl short names for every kind, unique against core resources BL-007 decision pending

Next: tenant onboarding you can trust

Goal: the tenant boundary is proven, one commit onboards a tenant namespace and a remote cluster, and a restore from Git rebuilds ownership.

item source help wanted
phase 1 exit test: the self-init example as a CI gate (the operator-identity e2e test is done: test/e2e/isolation_test.go) BL-012
phase 2 exit test: one commit onboards a tenant namespace and a remote cluster, including dynamic tokens phases
phase 3 exit test: wiping status then reconciling makes zero writes for marked kinds phases
restore without status: ownership of memberships and access children rebuilt from the server-side marker BL-011
ownership marker on KV version 2 metadata ADR 0012
typed kinds CertAuthCRL, TransitCacheConfig, TransitKeyConfig, KVSecretMetadata, PKICELRole and PKIKeyConfig BL-009 done
generic test that each kind's Desired covers its CRD schema (full-field comparison itself is done, ADR 0011); MaxLength on strings read by CEL BL-012 yes
DatabaseConnection and LDAP rotated root credentials on WriteOnce instead of CreateOnly (the engine supports it; the kinds do not use it yet) ADR 0017 yes

After that: scale and operations

Goal: large shared servers with many tenants.

item source help wanted
phase 4 exit test: 10,000 objects hold a 1 h drift interval inside the request budget, on CI hardware phases, load testing
pin the read after a write to the active node architecture
drift reads on standby nodes (OpenBao horizontal read scaling) architecture
per-namespace fairness for the write budget BL-016
integration suite on every supported OpenBao minor, not only the pinned one architecture, BL-014 yes
audit devices: an ADR on when the kind applies and how it reports a server that refuses API creation BL-003
guide for adopting existing configuration: extend migration.md with worked examples architecture yes
tenant quota e2e: a ResourceQuota count blocks the next object BL-012 yes
namespace-scoped install as the chart default TD-005

Later: plugins and the road to v1beta1

item source
phase 5 plugin support phases, architecture, BL-005
phase 6 plugin-backed kinds against real backends, each un-gated when its backend runs in CI phases, TD-004
hash key rotation without rewriting every write-only field known limitations
sharding by OpenBaoServer label for very large fleets architecture
namespace sealing support in OpenBaoNamespace, with shares through writeSecretRef architecture
v1beta1, when the conditions in the release process hold ADR 0020

Open questions

  • Are LDAP, Kerberos and RADIUS still built in on 2.7? RabbitMQ is built into 2.7.2; does it stay?
  • What is the sealed-namespace API (create parameters, key share return)?
  • Which kinds have metadata for the ownership marker (entities, groups and namespaces do; KV version 2 to test)?
  • How can a read be forced to the active node after a write?

Not planned

  • Kinds for features OpenBao does not have (Sentinel, replication, KMIP, Transform) (ADR 0009).
  • Syncing secret values into Kubernetes Secrets: use External Secrets Operator (ADR 0007).
  • Server lifecycle (deploy, initialise, unseal, upgrade): use the OpenBao Helm chart or an operator for that.
  • One-off actions as resources (login, issue, sign, rotate, tidy).
  • Admission webhooks (ADR 0005).

Changing the roadmap

Open a pull request against this page. New outcomes go into the backlog first with acceptance criteria; this page only orders them. Maintainers review the roadmap at each minor release.

Hash-key rotation design constraints

Rewriting-free rotation is deferred. A future implementation must verify a recorded HMAC against an explicitly bounded previous-key set before moving status to the new key, without writing unchanged remote inputs. It must still write a changed Secret, never accept an unmatched old hash as equal, and distinguish key retirement from restoring lost status. Multi-replica rollout, status patch failure and rollback need tests before enabling this behaviour.

The current key also participates in adapter credential fingerprints, beyond the generic writeOnlyHash check. A keyring cannot safely be introduced only in that generic comparison. Compromise recovery still requires retiring the old key; keeping it indefinitely to avoid writes is not a security fix. Until these contracts are designed and tested, use the existing controlled replacement procedure, which deliberately rewrites current write-only inputs. Out-of-band drift in values OpenBao never returns remains unobservable, regardless of key rotation.