Repository navigation
Conversation
Separate retained initialization counts from desired node replicas. Import live chart-managed bootstrap inputs and reject conflicting history. Document PD compatibility, controlled membership and startup limits.
…-stable-bootstrap-pr
bitflicker64
left a comment
bitflicker64
left a comment
There was a problem hiding this comment.
Blocking: yes. Summary: helm/test_helm_bootstrap.py is never run by CI, so its Pod template invariance checks for Store growth and PD growth or shrink do not guard the feature. The bootstrap resolver, legacy import, record guards and default-disabled renders read correctly, and the 13 new helm unittest suites pass.
|
|
||
| CHART = Path(__file__).resolve().parent / "hugegraph" | ||
|
|
||
|
|
There was a problem hiding this comment.
Important: .github/actions/helm-checks/action.yml runs only named Python files (test_helm_package.py, test_helm_seatunnel.py, test_helm_vermeer*.py, test_helm_metadata_cluster.py, test_helm_hook.py), so these lifecycle tests never run in the build. Add a python3 -m unittest discover -s helm -p 'test_helm_bootstrap.py' -v step to that action.
bitflicker64
left a comment
bitflicker64
left a comment
There was a problem hiding this comment.
Blocking: yes. Summary: two new helm unittest suites fail at this head, so the helm unittest CI step fails and the legacy-import success path has no passing coverage. A PD shrink below the recorded bootstrap majority also leaves the Store startup wait unsatisfiable.
| template: | ||
| spec: | ||
| containers: | ||
| - name: store |
There was a problem hiding this comment.
Important: this live Store fixture has no HG_STORE_PD_ADDRESS env, so the consumer guard added in 0519f2b fails every test in this suite with bootstrap import requires the existing store template to expose HG_STORE_PD_ADDRESS as a literal value. Add HG_STORE_PD_ADDRESS with the three original PD gRPC seeds to this container's env.
| template: store-statefulset.yaml | ||
| asserts: | ||
| - failedTemplate: | ||
| errorPattern: bootstrap import only supports the chart |
There was a problem hiding this comment.
Important: this pattern never matches: 96ef3e2 renamed the message to bootstrap only supports the chart's ..., and the one-entry custom-store:8500 list with count 3 fails first on initialStoreCount cannot exceed bootstrap.storeReplicas. Match the renamed message and give the fixture a custom list with at least three entries.
| - | | ||
| set -eu | ||
| REQUIRED={{ include "hugegraph.pd.quorum" . }} | ||
| REQUIRED={{ include "hugegraph.pd.bootstrapQuorum" . }} |
There was a problem hiding this comment.
Important: after a controlled PD shrink below the recorded majority, such as bootstrap.pdReplicas=3 with pd.replicas=1, this renders REQUIRED=2 against seeds where only pd-0 exists, so every restarted Store times out in the wait init container. Fail the render when bootstrap.enabled and pd.replicas is below hugegraph.pd.bootstrapQuorum, or document that this shrink is unsupported.
Correct legacy fixtures and isolate fresh topology guards. Cover retained-record checks in enabled and disabled installs. Clarify recovery, consumer drift and static seed limits.
bitflicker64
left a comment
bitflicker64
left a comment
There was a problem hiding this comment.
Blocking: no. Score: 9/10. Summary: The disabled path renders the same PD peers, Store allowlist, initial Store count, seed quorum and shard count as before, and the enabled resolver rejects conflicting pins, foreign or corrupt records and customized live endpoints. The earlier comments are addressed at this head, and helm-checks passes with test_helm_bootstrap.py and all 60 helm unittest suites.
Before → After
Changing PD or Store replica counts also regenerates initial peer lists and client seeds. This change adds an opt-in, immutable retained bootstrap ConfigMap so controlled node operations can change desired capacity without changing existing Pod startup inputs. The default remains disabled.
Adoption imports chart-managed live initialization inputs and rejects ambiguous history; offline upgrades require explicit retained counts. Membership changes, Pending Store admission, safe draining and rollback remain controlled procedures.
Dependency and validation boundary
Node-operation use requires the PD runtime-quorum change to be merged and included in the deployed PD image: #3428 . Static startup-seed and one-PD probe limits remain documented. The source PR reports Helm/Python/render regressions and unchanged default-disabled Pod templates; these checks are not cluster-level node-operation acceptance.
Source and detailed validation: hugegraph#352
Visual overview
Opt-in retained bootstrap counts keep initialization inputs separate from desired capacity. Runtime membership changes remain controlled operations and require a deployed PD image containing #3428. This is a design illustration, not cluster acceptance.