Thanks for looking. Read the first section before you spend time on a change — the repository is mirrored, and that changes how a contribution reaches us.
The upstream is a GitLab instance inside Gluesys. GitHub
(gluesys/daos-operator) is a push mirror of it, refreshed about every five
minutes.
That has one consequence you cannot work around: a commit pushed to GitHub will be overwritten by the next mirror run. A pull request opened against the GitHub repository cannot be merged there either. This is not a policy about outside contributions; it is how the mirror is configured.
So:
- Bug report, question, or a design proposal — open a GitHub issue. That works normally and is the right first step for anything larger than a typo.
- A patch — open a GitHub pull request anyway. We cannot press merge on it,
but we can read it, discuss it in place, and carry the commits to the internal
upstream with your authorship kept (
git amof your patches, or a cherry-pick that preservesAuthor:). We will tell you in the PR when it has landed, and the mirror will bring it back to GitHub within a few minutes. Then the PR gets closed as merged-via-mirror, not rejected. - If your change is large, please open the issue first. Being asked to restructure a big patch after the fact is a waste of your evening.
Code, comments, API documentation, commit messages and anything a user reads are
English. The repository started inside Gluesys and some of its history and
internal notes are Korean; files that are deliberately Korean carry a .ko.md
suffix or are marked (Korean) where the README links them. Those are working
records, not translations waiting to happen — there is no English counterpart.
Please do not rewrite old Korean commit messages: the GitHub mirror and the published chart provenance both depend on the existing commit ids.
Everything the CI checks before it touches hardware runs locally. You need Go
1.26 and Helm; make fetches controller-gen and setup-envtest on demand.
go build ./...
go vet ./...
gofmt -l api/ internal/ cmd/ # must print nothing
make manifests generate # regenerates CRDs and deepcopy
git diff --exit-code -- config/crd api/
make helm-sync # copies CRDs, RBAC and the dashboard into the chart
git diff --exit-code -- charts/
make test # unit tests + envtest
helm lint charts/daos-operatorTwo generated files have their own commands, and CI fails if they drift from their source:
charts/daos-operator/values.schema.json—hack/gen-values-schema.py, with--checkto verify. It carries each key's comment fromvalues.yamlas the schemadescription, and Artifact Hub renders those. If you edit a comment invalues.yaml, regenerate.charts/daos-operator/templates/clusterrole.yaml—make helm-sync, fromconfig/rbac/role.yaml.
Plus the license check: every .go, .py, .sh, .c, .cpp, .h and
Dockerfile* must carry
SPDX-License-Identifier: Apache-2.0
in its first five lines, with your own copyright line below it if the file is substantially yours.
The end-to-end tests drive a real DAOS system: NVMe handed to SPDK, hugepages, a fabric, several nodes. They run on a dedicated cluster from the internal CI, on two lanes that are reset from a snapshot before every run. There is no way to run them from a fork.
If your change touches a path only the cluster exercises — engine lifecycle, formatting, rank membership, upgrade — say so in the PR and describe what you did run. We will run the tiers against the testbed before landing it. You are not expected to own hardware to contribute.
test/e2e/ is readable without the hardware, and the profiles under
charts/daos-operator/ci/ show what each configuration looks like.
This project has retracted several of its own conclusions — a transport was blamed for data corruption that turned out to be two server ranks formatting the same drives, and a "hardware ceiling" turned out to be one transport's loss. The retractions are in the docs on purpose.
So if a change comes with a number, please give the conditions with it: the storage class and tier layout, the fabric and provider, the object class and chunk size, how many concurrent requests, and how many runs the number came from. A measured range beats a single best result. If a claim rests on one trial, say that too.
- One logical change per commit. The message should say why, since the diff already says what.
- Sign off your commits (
git commit -s), which adds aSigned-off-by:line. It is the Developer Certificate of Origin: you are stating you have the right to submit the work under this project's license. There is no CLA. - Use a real name and a working email in
Author:. Some of our own early history has the email in the name field; do not copy that. - Contributions are accepted under Apache-2.0 (see
LICENSE§5).
| Path | What it is |
|---|---|
api/v1alpha1/ |
the CRD Go types — DaosSystem, DaosPool, DaosContainer, S3Service. This is the public API surface |
internal/controller/ |
one reconciler per kind, plus the engine, format, rank and upgrade paths |
internal/dmg/, internal/certs/, internal/discovery/, internal/render/ |
the dmg wrapper, certificate handling, node facts, manifest rendering |
cmd/ |
the manager (cmd/main.go), the hostprep DaemonSet binary, the kubectl-daos plugin |
charts/daos-operator/ |
the Helm chart, its values schema and the ci/ profiles |
config/ |
kubebuilder manifests — CRD bases, RBAC, the Grafana dashboard |
test/e2e/ |
the end-to-end suite the CI tiers run |
doc/adr/ |
the decisions behind the deployment, device-binding and upgrade models (Korean) |
hack/ |
generators and CI cluster scripts |
images/hostprep/ |
the hostprep image build |
The storage classes are not equally mature: nvme is what the tiers exercise
most, while kdev and file exist for testing against upstream defaults and are
not meant for production — charts/daos-operator/README.md says so where it
describes them.
Open an issue for anything public. For a security report, do not use the issue
tracker — see SECURITY.md.