0.15.0

Every release now reaches both OpenShift operator catalogs by itself

Submitting the operator bundle to the two community catalogs was hand work, twice per release, and the cost was measurable rather than theoretical: both catalogs sat at 0.9.59 while the product reached 0.14.1, so 0.14.0 was never submitted at all.

The submit-catalogs job in operator-release.yml now opens the bundle pull request against k8s-operatorhub/community-operators and redhat-openshift-ecosystem/community-operators-prod on every release, switched from distribution/channels.yaml like every other optional channel. Both catalogs carry 0.14.1 as of today, which was the last submission opened by hand.

The load-bearing part is spec.replaces. Its only correct value names the version immediately preceding this one in the catalog being submitted to, which is external state this repository cannot derive: it lags releases by however many submissions are unmerged, and the two catalogs disagree with each other. So the value is derived from the catalog at submission time and injected into the bundle copy, and nothing in the repository declares it.

Both upstream gates reject an unlinked graph and neither accepts olm.skipRange as a substitute. opm index add --mode replaces prunes the previous bundle on operatorhub.io; opm validate reports multiple channel heads found in graph on the file-based catalog side. Both were measured, the second on a real submission whose FBC dry run failed exactly that way.

A submission that cannot happen is a notice rather than a red release: the channel switched off, no token, the operator not listed yet, the version already carried, an open submission for another version. Three conditions do fail the job, because each is a misconfiguration that would otherwise break the channel quietly: the controller image is not anonymously pullable, the fork's parent is not the upstream, and a GitHub API read failed.

Database sizes that were never measured

An overview size the server declines to answer is now reported as absent instead of as 0 bytes, on MySQL, PostgreSQL, SQL Server and Oracle. An unmeasured database read as an empty one, which is a different and much more alarming claim.

CREATE TABLE speaks the engine it is aimed at

The generated statement was PostgreSQL's DDL whatever the target, so it did not run where it was generated. Each engine now emits its own.

Agent

Grounding capture no longer pulls in engine and extension objects, so a run is grounded in the schema the question is about. A stop that read nothing is answered for whoever stopped it rather than left silent.

Distribution

The AWS Marketplace, Azure Marketplace and CasaOS App Store channels are live. The deb and rpm postinstall uses systemctl try-restart, so installing over a stopped service no longer starts it.

Documentation

A sweep across the provider, agent, security and architecture pages replaced line-number citations with names, and corrected the counts several pages stated from memory. A contributor devcontainer landed, with CI named as the gate it actually is.

Contributors

Named for what they did, in the order their work appears above.

• voidofrgestudio found that an unmeasured database size was being published as 0 bytes on MySQL and PostgreSQL, and fixed it to report absence instead. An unmeasured database reading as an empty one is a different and much more alarming claim, which is what makes this a correctness fix rather than a display one.
• be-student did the same for SQL Server and Oracle, preserving an absent overview size rather than filling it in, and replaced the stale line-number citations in the Redis provider doc with named ones.
• 7487 made CREATE TABLE emit the DDL of the engine it is aimed at instead of PostgreSQL's, so the generated statement runs where it was generated, and replaced line-number citations with named ones across the PostgreSQL, ClickHouse, Couchbase and Apache Druid provider docs and the provider factory page.
• koraysrn stopped agent grounding from capturing engine and extension objects, so a run is grounded in the schema the question is about rather than in the server's own furniture.
• yusuf-gundogdu made an agent stop that read nothing answer for whoever stopped it, measured gemma4:31b against all six evaluation gates, brought the AWS Marketplace, Azure Marketplace and CasaOS channels live, and wrote the guard that derives the model roster count instead of restating it.
• XiaoZ-0218 removed the dead getPlaceholder mentions from the provider docs and stopped the architecture pages crediting SQLBaseProvider with placeholders it does not implement.
• SyedMuhamadYasir replaced the stale source line citations in the SQLite provider doc, and added the contributor devcontainer along with the clarification that CI is the gate.
• saad-works resolved the two static process.env bracket shapes the documentation guard could not see, and gave its boundary test a control so the guard cannot pass vacuously.
• dvd233 pinned the measured aggregate helper citation to the source it describes.
• dchaudhari7177 documented the bundled Node.js runtime that the SBOM does not report.
• SatvikMishra08 added Helm to the CONTRIBUTING prerequisites, which is what the chart tests actually need.

Helm chart: 0.1.61

No template or value moves. artifacthub.io/changes records seven entries and containsSecurityUpdates stays false.

Full changelog: 0.14.1...0.15.0