Skip to content

Release v0.12.0

Latest

Choose a tag to compare

@homer6 homer6 released this 16 Nov 08:45
075b6d1

Released: 2026-11-16 · Branch: main · Previous: v0.11.12 (2026-10-31) · Tracking: #1206

A minor release that gets the operator ready to publish as open source. It adds the license and contributor documents and a public README, and it removes Pixie and the observability kinds built on it. It also carries security fixes for tenant isolation, datastore credentials and Zetta's secrets, moves Orama's own cluster resources out of the repository, and pins zetta-tunnel v0.2.0. It was cut without a planned release epic, as part of the open-source release (#1148).

Measured Value
Resource kinds 49 CRDs in chart/crds (54 in v0.11.12; five removed)
Pull requests 32 merged since v0.11.12
CRD changes removed: ObservabilityProvider, ObservabilityHub, ObservabilityAgent, PixieCloud, PixieVizier. Added fields: OramaZettaCluster spec.egressAllowConfigMap and the identity-root rotation; Lake fields for the tenant-isolation and service changes; OramaCloudV1 spec.components.controlDB.referenceRows and spec.components.pgAdmin.email
License MIT, with COPYRIGHT, SPDX headers on every source file, THIRD_PARTY_LICENSES and an SBOM
zetta-tunnel v0.2.0 (main-10acaa2, by digest), from v0.1.0
Tests see Verification

What it includes

Open-source release

  • License and notices (#1194): MIT LICENSE and COPYRIGHT (OramaSearch, Inc.), SPDX-License-Identifier: MIT on every source file (tools/spdx-headers.py), THIRD_PARTY_LICENSES from go-licenses, and sbom.spdx.json.
  • Contributor documents: SECURITY.md, a private reporting address and response times (#1161); CONTRIBUTING.md, CODE_OF_CONDUCT.md and CLA.md (#1190).
  • A public README (#1202) with a quickstart that was run on a throwaway cluster. make run now passes --control-plane-bind-address="", so it starts without a control-plane certificate.
  • No internal material in the tree: comments, CRD descriptions, samples and chart values (#1193), test fixtures (#1195, #1188), and the public docs (#1196). The design log, the security reports and the research notes moved out of the repository (#1190).
  • Cluster resources are an example (#1149): config/clusters/ holds a cleansed example deployment. The tests read it, or another directory named by ORAMA_CLUSTERS_DIR.
  • Cloud v1 configuration instead of embedded data. A stack's plan catalog is spec.components.controlDB.referenceRows. The built-in rows are a made-up example that refuses to run on a database that already has a catalog (#1185). pgAdmin's sign-in is spec.components.pgAdmin.email, admin@example.com by default (#1189).

Pixie removed (#1199)

ObservabilityProvider, ObservabilityHub, ObservabilityAgent, PixieCloud and PixieVizier are removed, with their controllers, CRDs, RBAC, the vendored Pixie manifests and the Pixie build material. The hub Prometheus and its federation, and the per-cluster alert routing, go with them; nothing replaces them in this release. The platform layer's Prometheus, Alertmanager and Perses stay, with each Cloud v1 stack's rules and dashboards. The manager also loses its alertmanagerconfigs and prometheuses grants. make check-crds now fails while an object of a removed kind remains (hack/retired-crds.txt).

Security

  • A person's service launch gets its own token, never the Lake's agent service token (#1150).
  • An explicit CIDR egress rule reaches public addresses only (#1153), and a tenant-named peer remote must be public (#1154).
  • The default images that run tenant code are pinned by digest (#1155).
  • The operator verifies a workspace Postgres's certificate before signing in as superuser (#1156).
  • Image builders never mount the platform pull Secret (#1159).
  • An app's bindings copy only its own workspace's services (#1160).
  • Tenant services never run in the Lake namespace. Leftovers there are reported as a condition and a Warning event, never deleted (#1181).
  • Five findings in one batch (#1162):
    • TiDB's root password is set while TiDB bootstraps;
    • a lake secret's projection is deleted when lake says the secret is gone;
    • a row's workspace is checked against its customer;
    • methodology rows and agent pods are bounded on the operator's side;
    • mesh projections carry the pod UID and namespace inode.
  • TiDB and Cassandra tenants get a user of their own, never the administrator (#1171). Each service's administrator password is then rotated once (#1173).
  • Zetta: a connector-credential root Secret mounted outside the data volume, and the egress allowlist as a file the engine re-reads (#1167). Identity-root rotation as a spec generation, and the enrollment CA (#1177). Each is switched on by what the image's own --help lists.

Datastores and Cloud v1

  • lake-duckdb v0.2.0 is the default DuckDB image (#1204). It fixes lake-duckdb's security findings, whose reproductions are public. lake-sqlite v0.2.0 is the default SQLite image (#1203). Both are pinned by digest.
  • Cloud v1:
    • the embeddings server requests the 4Gi it holds, limited to 5Gi (#1147);
    • a stack's analytics going stale raises OramaAnalyticsStale;
    • ClickHouse's own memory cap is three quarters of its pod's limit (#1141). The default renders the same file as before.

Other

  • zetta-tunnel v0.2.0 (#1138), which the mesh projection needs.
  • Seventeen tests the untested merges left failing, fixed on dev before the cut:
    • seven in #1205, one of them a real fix: spec.env can no longer override the two Zetta file variables the operator mounts;
    • ten in #1208: fixtures for #1162's workspace routing check and its revocation.

What it does not include

  • Any replacement for Pixie, the hub metrics, or alert delivery.
  • Public images. Every default image, and the operator's own, is in a private registry (#1186).
  • These open pull requests:
    • #1164 with #1170, #1175 and #1176: apps stop being served until each Lake has its own wildcard domain.
    • #1178: moves DNS for every Cloud v1 host.
    • #1163 and #1157: their lake halves are not merged.
    • #1180: waits on a zetta-tunnel change.
    • #1137: denies the cluster's own addresses in the egress gateway snapshot, not yet run against zetta-tunnel v0.2.0.
    • #857, #860 and #864: on hold for a planned rollout.
    • The backup system: #1182 to #1184.

Before it rolls

Each step needs the cluster owner's consent, in this order:

  1. Remove the observability resources under the operator that still ships them (v0.11.12). Delete the ObservabilityAgents, ObservabilityHubs and the ObservabilityProvider from their committed files as they were before this release, and wait until they are gone. The managing operator tears down what each one wrote on its clusters.
  2. Apply the Cloud v1 set files that carry the new fields before the operator rolls:
    • controlDB.referenceRows with each stack's current catalog. Otherwise each stack's reference-data Job fails its guard, and its catalog stays unchanged.
    • pgAdmin.email on the stacks that run pgAdmin. Otherwise pgAdmin signs in as admin@example.com from its next restart.
  3. make check-crds CONTEXT=<context> on every cluster. It fails while an object of a removed kind remains.
  4. Roll the operator, with the CRDs applied widened first and exactly after (docs/operations/crd-rollout.md). Then apply the platform layer by both installation paths: kube-state-metrics stops listing the Pixie kinds.
  5. lake must carry its half of #1150 (lake#1276) in the same roll. Until both run it, a person's service launch answers 401.
  6. Roll this operator before the Zetta images that read the connection-root file and the identity rotation (zetta#1509, zetta#1512, zetta#1514).

What changes on the roll:

  • Each Zetta StatefulSet rolls once, for the connection-root mount.
  • Each existing TiDB or Cassandra service changes its delivered user once, and has its administrator password rotated once.
  • The peer gateways and the mesh agents move to zetta-tunnel v0.2.0.
  • A build FROM a private base image fails until the Lake names its own base-image credential.
  • Every DuckDB and SQLite service rolls once onto v0.2.0. After it:
    • DuckDB refuses multi-statement sql (use /v1/batch), and refuses COPY, EXPORT DATABASE and similar;
    • SQLite refuses ATTACH, VACUUM and connection-changing pragmas. Neither image has run on a cluster.
  • Each Cloud v1 embeddings pod restarts once, for its new memory request.
  • After the roll, apply orama-infra's v1-eu analytics change: an 8Gi limit, so a 6 GiB ClickHouse cap.

Known issues

Security first:

  • Open security issues: 18. Among them:
    • tenant builds still run without process isolation (#792, #759);
    • tenant app pods share nodes with Lake (#763, #818);
    • ClickHouse's built-in users (#772);
    • NetworkPolicy enforcement is not verified (#699).
  • Alerts are not delivered: no AlertmanagerConfig is written since Pixie's removal. Each cluster's Alertmanager still evaluates the rules.
  • The manager exits on a cluster without Cilium or CloudNativePG (#1200).
  • Not run on the dev environment. By the cluster owner's decision, this release was not built and run on the dev environment before the cut.
  • Defaults that point at Orama: the chart's default operator image tag is the mutable v1-sandbox. The operator's repository and image oramasearch/orama-operator are public from this release, but the other built-in default images are private oramasearch/* images (#1186). Three Cloud v1 default images are not open source: mongo (SSPL), typesense (GPL-3.0), and redis from 7.4 (RSALv2/SSPL).

Verification

Check Result
go build ./..., go vet ./... pass
go test -timeout 45m ./... with envtest (Kubernetes 1.30.3), macOS arm64 3,236 passed, 0 failed, 20 skipped. Every package ok: cmd, cmd/check-image-digests, internal/backup, internal/controller, internal/crdrollout, internal/imagepolicy, internal/meshprojection. The 20 skips are the tests that need a live service: 17 a Postgres (next row), one the Prometheus release (PROMETHEUS_BINARY_DIR), one a zetta-tunnel build (EGRESS_GATEWAY_BINARY), one real chart downloads (LEAF_LIVE)
make generate manifests reproduces the committed deepcopy and CRDs (two CRD descriptions regenerated on this branch)
python3 tools/spdx-headers.py --check 0 missing
helm lint chart 1 chart(s) linted, 0 chart(s) failed
The 17 Postgres-backed tests, against a throwaway local Postgres 17 all 17 pass with #1210. Before it, four pgx store tests failed: their fixtures lacked workspaces.customer_id, which #1162's join needs and lake's schema has (test-only). TestLakeSchemaEnvtest passes against the default local Postgres (it hard-codes port 5432)
orama-infra check.sh (the operator's tests against the real cluster resources) pass: cmd and internal/controller ok against orama-infra's files
Linux image build from the tag oramasearch/orama-operator:main-075b6d1f, built on the cluster BuildKit from a detached checkout of v0.12.0 (linux/amd64); Docker Hub v0.12.0 names the same digest, sha256:fb24a348433a84bee2e28ac992e6b72fde22aab920b25a8832075eae309805f4
Live on a cluster not deployed: the release is cut without a roll. The order for one is under Before it rolls