Storage: fixed-width millisecond timestamp format with a one-time migration
so lexicographic comparisons stay chronological; snapshot auto-IDs get
nanosecond precision; Close is synchronized; commit failures roll back; DSN
paths are URI-escaped; the dedup index is no longer rebuilt every startup;
the in-memory store now dedups by event_id and honors Limit/Offset/SinceSeq.
Event/discovery: stream IDs URL-escape the selector so slash-containing label
selectors round-trip and cannot collide; alias resolution fills empty fields
only and no longer clobbers explicit group/version/kind; rc maps to
replicationcontrollers and rs to replicasets.
Watch: the label selector is now actually applied to List and Watch; the 410
relist path is backoff-throttled, backoff resets after a healthy watch, an
empty list resourceVersion backs off instead of looping forever, non-410
watch errors reconnect without writing a spurious permanent gap, an idle
watch timer forces reconnects, and Run cancels sibling streams and recovers
from panics. Synthetic relist events carry a distinct event_id so unchanged
objects re-listed after a gap survive dedup.
Materialize: a baseline now resets object state (objects deleted during a gap
disappear) and heals open gaps; diff ignores server metadata by path, not by
key name, so user fields named status/uid are kept; nested add/remove carry
Added/Removed flags; dotted paths are escaped; Diff validates the window and
treats a zero before as empty; Snapshot records per-stream watermarks and
writes a TypeSnapshot journal record.
Replay: apply requires a successful dry run; gvrFor covers every policy kind
so Include* toggles work; unsupported kinds are reported as skipped instead
of silently dropped; NodePort services lose clusterIP and nodePort while
headless keeps clusterIP: None; targetNS is always honored; namespace mapping
collisions are detected; plan IDs and field managers are collision-safe and
plans are mutex-guarded against concurrent dry-runs/applies.
API: cursor pagination pushes ingest_seq into the SQL filter (export --server
no longer truncates); gap records expose their payload; /v1/diff requires a
cluster_id; dry-run/apply no longer accept a client-supplied kubeconfig;
plans carry target_context; HTTP server gets timeouts.
Metrics: the watch collector and replay planner now bump the registered
counters, and the server refreshes store-derived gauges on a ticker.
CLI: replay apply works locally and refuses to apply after a failed dry run;
coverage --server follows cursors and shows gap details; timeline
auto-detects namespaces and resolves cluster-scoped objects; diff validates
the window; export writes 0600 files; PVCs are no longer treated as
cluster-scoped; krply-server reports the build version.
- grpc query server for events, snapshots, diffs, and replay plans
- server metrics, request audit logging, and identity endpoint
- krply-server and krply cli entrypoints
- dependency-light dark web ui backed by the query api
