nolar/kopf 0.23

on GitHub

Originally released on 2019-11-20 10:15:07+00:00 (link).

TL;DR: Massive refactoring, renames, code moves. Generally, should be backward-compatible.

RISKY CHANGES (can be BREAKING, should be not):

• Authentication and API talking are fully reworked: pykube-ng and kubernetes clients are now piggybacked by default to extract the endpoints and credentials, but are not used for the API communication. This can break API connectivity in some cases. #226 #227
• In-memory structures were introduced to carry the technical information/objects/flags for individual resources. This can lead to increased memory usage in huge clusters with thousands of monitored resources. #230
• @kopf.on.resume() handlers were fixed, and now execute when previously they were not executed by mistake, but this could be taken as an expected behaviour. This can lead to massive patches of all resumable objects on every operator startup (if there are 2+ handlers), which can be a problem in huge clusters. #230 #236
• Massive internal renames and code moves; not exposed publicly, unless hacks or monkey-patching were used. #209 #210 #211

New features:

• @kopf.on.startup() handlers for operator initialisation. #225
• @kopf.on.cleanup() handlers for operator shutdown. #225
• @kopf.on.login() custom authentication handlers. #226
• @kopf.on.probe() handlers for liveness metrics. #226
• Liveness probes with kopf run --liveness. #228
• memo kwargs to keep runtime-only operator-lifetime-limited arbitrary values. #234
• retries= limiter for handlers in addition to timeout=. #222
• errors=TEMPORARY, errors=PERMANENT, errors=IGNORED modes for handlers. #222
• Explicit owner is not needed in kopf.adopt() and hierarchy methods, current object is used by default. #203

Bugfixes:

• @kopf.on.resume() handlers fixed:
  • There can be more than one resuming handler. #230
  • Are executed if they go after the on-create/on-update handlers. #230
  • Are not repeated every few minutes for no reason, guaranteed to be called once. #229 #230
  • Can be retried in case of temporary or arbitrary errors. #230
  • Can have sub-handlers. #230
  • Are not invoked if the object is being deleted. #233
  • Can be explicitly marked as deletion-safe, and hence be invoked on resuming during deletion. #233
• spec, status, metadata fields are not added to the body when originally absent. #198
• OS signal handlers are ignored on Windows (not supported by asyncio). #220

Internal changes:

• Fully type-annotated in the code (not in the tests). #195 #196 #197 #200 #201
• Fully asynchronous internally (asyncio/aiohttp); synchronous pykube-ng client is removed. #217 #227
• Handler outcomes and persistent state management separated from each other. #221
• E2E tests are now isolated from each other, preventing false-positives. #205
• The latest k8s & minikube versions are used in e2e tests. #193
• No logging of empty meaningless diffs in on-resume and other handlers. #236
• No patching of the objects if there is no need to patch (and no logging). #236
• No purging of the handler progress if there is only one handler (and no logging). #236
• cause.event is renamed to cause.reason to avoid terminology conflicts. #201
• kex is added as an alias for KopfExample CRD for demos/docs. #235