Abstract

A method is disclosed for detecting drift in the shape of a fleet of long-running workload units on a container orchestrator, in which a clean result is possible only after both sides of the comparison have actually been read. The expected shape of each unit, principally the set of container names it must run, is stated in a separate expectation ledger. The ledger names, but never parses, the manifest that would produce the unit, so editing a manifest to match a broken cluster cannot make the comparison agree with itself. The live side is enumerated without reference to the ledger's entries, by a name prefix or a fixed membership label, and the ledger is subtracted from it. Every enumerated unit identity therefore ends as a managed entry, a recorded exception carrying a mandatory reason, or a typed finding, including a finding for a live unit that no entry names. Each live unit matched to a ledger entry is also compared against the last-applied configuration recorded on the object itself, which forecasts what the next declarative apply will revert without reading any repository. The detector distinguishes clean, drift and could-not-look. Each of the following raises one distinguished self-check error, which produces no report and the could-not-look outcome: an unreadable or malformed ledger, a failed or unparseable listing, a zero-unit listing, or a finding count above a configured bound. The detector has no write path.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS