Skip to content

feat: wire the periodic reconciliation pass into the EtcdMirror engine#408

Open
xrl wants to merge 6 commits into
etcd-io:mainfrom
xrl:pr/etcdmirror-reconciliation-pass
Open

feat: wire the periodic reconciliation pass into the EtcdMirror engine#408
xrl wants to merge 6 commits into
etcd-io:mainfrom
xrl:pr/etcdmirror-reconciliation-pass

Conversation

@xrl

@xrl xrl commented Jul 7, 2026

Copy link
Copy Markdown

Wires the periodic reconciliation pass into the EtcdMirror engine's steady-state loop, taking Config.ReconcileInterval / Config.ReconcileDeleteOrphans from documented no-ops ("NOT YET WIRED") to live behavior. The diff-and-repair pass itself shipped with the engine (it already backs the forced-resync mark-and-sweep, the OverwriteAndPrune genesis pass, and the Drain verification); this PR promotes it to a schedulable, continuous integrity signal.

Depends on: #407 (the pkg/mirroragent engine — previous rung of this series).

Scheduler design

Deadline-driven and single-goroutine: the steady-state consume loop gains a fourth select case on a timer armed from a persistent nextReconcile deadline (nil channel when disabled, so the case never fires). The pass runs inline on the Run goroutine with repair=true and orphan deletion gated on ReconcileDeleteOrphans. Non-overlap is structural, not locked:

  • it can never overlap the genesis scan or a forced-resync sweep (those run outside the tail, where consume isn't running);
  • it can never overlap a drain (a requested or configured Drain gates the pass — drain's own verification supersedes it);
  • it can never overlap itself (synchronous on one goroutine).

Every counts-producing pass — periodic or mandatory — re-arms the deadline, so a forced-resync sweep pushes the next periodic pass out a full interval (it just produced the same signal). Errors escaping the pass ride the existing classification taxonomy untouched: transient/throttle are absorbed inside by the retry/backoff machinery, ErrCompacted propagates to the forced-resync arm, and permanent errors (PrefixConflictError, FenceError) fail the agent. Every repair and prune delete rides the same fenced mod-revision-compare Txn path as ordinary applies, so a concurrent fence bump aborts the pass loudly.

Always-on key counts

The per-side key counts (Snapshot.SourceKeyCount/TargetKeyCount) are the input to the controller's InvariantsHeld condition and must stay meaningful even when periodic reconciliation is disabled — so recordKeyCounts now stamps LastReconcileTime alongside the counts, and every mandatory pass (forced-resync sweep, OverwriteAndPrune, drain verification) dates them. Counts, timestamp, and drift are published in a single snapshot update so no reader can observe fresh counts with stale drift.

Bounded memory during a paced repair

A long, healthy, rate-limited repair pass would previously let the source watch buffer grow for the duration of the pass (the existing watch-cancel fired only from backoff/quota branches, which a healthy pass never enters). The pass now cancels the source watch once queued repair work exceeds a few seconds' worth of MaxOpsPerSecond pacing and resumes from the watermark through the normal transient re-watch path — the same bounded-retry posture the engine takes everywhere else. The genesis sweep keeps its watch: that one drains into the byte-bounded replay buffer by design.

Validation and tests

Config.Validate() rejects a negative ReconcileInterval (0 = disabled). New unit tests cover the gating table (reconcileDue), deadline re-arm, validation, and the watch-cancel threshold; new embedded-etcd integration tests cover: injected divergence repaired by the periodic pass with drift reported; orphan policy in both report-only and delete modes; the pass never firing before steady state; a fence bump aborting the pass cleanly; mandatory passes keeping counts fresh with periodic reconciliation disabled; and a paced 100-key repair that forces a mid-pass watch cancel, then converges and still replicates fresh writes. The new tests are race-detector-clean over repeated runs.

Verified with make test and make verify.


PR series — operability fixes, TLS & EtcdMirror

Small single-purpose PRs from live kind-cluster testing of the operator. Each stands alone unless an After is listed. → = this PR.

PR Lands After
Reviewable now — small, any order
🟢 #374 Requeue instead of swallowing the client-cert provisioning error
🟢 #391 Grant events.k8s.io RBAC so operator Events are actually recorded
🟢 #392 Correlate members[]/leaderID from one health snapshot (consistent leader)
🟢 #393 Propagate user-supplied altNames.ipAddresses into certificates
🟢 #394 Accept day-suffix validityDuration (365d, 100d12h) as documented
🟢 #395 Surface early reconcile errors as a Degraded condition (was empty status)
🟢 #379 Configurable reconcile worker pool (--max-concurrent-reconciles)
🟢 #369 kind-based stress/scale e2e harness (1/3/7 members, churn, quorum watcher)
🟢 #398 podTemplate.spec scheduling fields: topologySpread, resources, priorityClass, schedulerName
🟢 #397 First-class Helm chart covering the full config/ surface
🟢 #399 Quorum-aware PodDisruptionBudget per EtcdCluster
🟢 #400 Scale-in removes the right member and transfers leadership first
🟢 #401 Backend quota + auto-compaction spec fields, NOSPACE alarm surfacing
🟢 #402 Reconcile-loop gates replace the blocking StatefulSet-ready wait
🟢 #403 Quorum-gated one-pod-at-a-time version upgrades via OnDelete
TLS stack — in order
🟢 #376 Independent spec.tls.{peer,client} surfaces (breaking alpha API)
#377 TLSReady condition + TLS lifecycle Events #376
#378 Multi-member TLS quorum e2e + PeerCANotShared #377
EtcdMirror — in order
🟢 #406 EtcdMirror CRD: one-way cross-cluster replication API (types + CEL)
🟢 #407 pkg/mirroragent replication engine (fenced checkpoint-in-target) #406
🟢→ #408 Periodic reconciliation pass wired into the engine's steady-state loop #407
🟢 #412 mirror-agent binary: config/TLS-reload/statusz/metrics + kind e2e #408
🟢 #413 EtcdMirror controller: Deployment rendering, statusz-driven conditions, guards, fenced finalizer #412
🟢 #414 CRD/sample/editor+viewer-role kustomize wiring #413
🟢 #415 Controller-side phase/condition gauges, shipped PrometheusRule + dashboard #414
🟢 #416 Full-lifecycle kind e2e: fence overlap, compaction+prune, restart resume, cert rotation, cutover/reversal #415
Parked as drafts pending #363
#382 Per-cluster domain metrics on the operator /metrics endpoint
#384 EtcdCluster admission webhooks (consolidating with #328) #363
#386 EtcdBackup CR → object storage (S3/GCS) #363
#387 Automatic quorum-loss disaster recovery (bootstrap-latch guarded) #363

🟢 ready · ⚪ draft

@kubernetes-prow

Copy link
Copy Markdown

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: xrl
Once this PR has been reviewed and has the lgtm label, please assign jmhbnz for approval. For more information see the Code Review Process.

The full list of commands accepted by this bot can be found here.

Details Needs approval from an approver in each of these files:

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

@kubernetes-prow

Copy link
Copy Markdown

Hi @xrl. Thanks for your PR.

I'm waiting for a etcd-io member to verify that this patch is reasonable to test. If it is, they should reply with /ok-to-test on its own line. Until that is done, I will not automatically test new commits in this PR, but the usual testing commands by org members will still work.

Regular contributors should join the org to skip this step.

Once the patch is verified, the new status will be reflected by the ok-to-test label.

I understand the commands that are listed here.

Details

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

This was referenced Jul 7, 2026
@xrl xrl force-pushed the pr/etcdmirror-reconciliation-pass branch from 4f3ad2c to adcdde4 Compare July 7, 2026 23:43
xrl and others added 6 commits July 10, 2026 02:05
Introduces the EtcdMirror CRD (schema only, no controller/agent) that
will describe a continuous one-way key-range sync from a source etcd
cluster to a target etcd cluster. Adds EtcdMirrorSpec/Status, the
EtcdMirrorEndpoint/TLS/Auth/Sync/Checkpoint/Reconciliation sub-specs,
phase and condition constants (including TargetThrottled and
ReplicationLagExceeded), printer columns, and CEL XValidation rules
for the destPrefix/noDestPrefix mutual exclusion, the
endpointList/serviceRef oneOf, and the insecureSkipVerify/
insecureSkipVerifyAcknowledgeRisk companion-field requirement.

Registers EtcdMirror/EtcdMirrorList in the scheme, regenerates
zz_generated.deepcopy.go and the CRD manifest via controller-gen, and
adds a realistic sample CR. CEL rules are covered by a table-driven
envtest suite in etcdmirror_cel_test.go, following the
tls_cel_test.go pattern, exercised against a real envtest apiserver.

The design's Metrics field (reusing EtcdClusterSpec's MetricsSpec) is
intentionally omitted: MetricsSpec does not exist on this branch yet
and will land with pr/domain-metrics; it will be added in a later PR
once that type is importable.

No controller or agent logic in this PR -- types, generated code, and
CRD schema only.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Signed-off-by: Xavier Lange <xrlange@gmail.com>
Reshape the EtcdMirror API for the Design-3 engine (checkpoint stored in
the target etcd, stateless agent) per the 48-finding cross-cloud review.

Deletions:
- checkpoint.storageSpec (no PVC; checkpoint lives in the target at a
  reserved fenced key)
- expectEmptyPrefix (superseded by initialSync.mode)
- noDestPrefix (redundant; one anchored rewrite formula
  key' = target.prefix + destPrefix + TrimPrefix(key, source.prefix))
- syncInterval

Additions/renames:
- spec.mode: Sync|Drain; status.cutover{drainTargetRevision,
  drainedRevision, verifiedTime, counts, leasedKeyCount}
- initialSync.mode: RequireEmpty|Overwrite|OverwriteAndPrune,
  startRevision
- sync: requestTimeout, excludePrefixes, pageKeyLimit,
  watchBufferBytes; batching flushes only at source-revision boundaries
- TLS: secretRef pointerized (nil = system trust roots), caBundleRef
- pod resources; configurable reserved checkpoint key
- watermark-derived lag/liveness semantics; frozen condition vocabulary
  (TargetQuotaExhausted, ResyncLoopDetected, CutoverReady,
  InvariantsHeld, LearnerEndpoint, Prefix/DirectionConflict) and Event
  Reason constants as API contract
- status: source/target versions and cluster IDs, leaseBackedKeyCount,
  always-on source/target key counts, initialSyncTotalKeyCount
- CEL: scheme-vs-TLS rules, prefix/rewrite immutability
- docs/etcdmirror.md (fidelity caveats, one-way contract) and
  regenerated API reference

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Xavier Lange <xrlange@gmail.com>
Purpose-built replication engine, zero k8s dependency; replaces the
clientv3 mirror.Syncer plan.

Architecture: unpinned chunked scan + watch-replay from R0 (reflector
pattern) — the watch starts at the revision observed before the scan
and events are buffered/replayed over the scanned base, eliminating
mid-scan compaction as a failure class. Pull-based single-page scan
with byte-bounded pages keeps agent memory bounded; the source watch
is cancelled on sustained target backoff and resumed from checkpoint.

Fence protocol: checkpoint/watermark stored in the target etcd at a
reserved key (\x00-after-prefix convention, exact-match excluded from
scans/counts/prune/RequireEmpty), written in the same Txn as each
applied batch and guarded by a mod_revision compare on every write
path (applies, reconciliation repairs, prune deletes). Fence value
carries {linkUID, epoch, role}; role flips to Primary at cutover so
straggler applies fail their compare loudly. Undecodable checkpoints
fail closed; prune-pending is durable across restarts.

Batching: flush only at source-revision boundaries (whole revisions
coalesced, never split), ~1MiB byte watermark, one MaxTxnOps slot
reserved for the checkpoint write.

Error taxonomy: ErrCompacted classified distinctly;
rpctypes.ErrNoSpace -> TargetQuotaExhausted (permanent until operator
acts); oversized-Txn InvalidArgument and >2MiB client send-cap
ResourceExhausted both permanent with the offending key surfaced
redacted; throttle-distinct backoff; N-consecutive-forced-resync
livelock detector.

Liveness: watermark-derived progress via WithProgressNotify +
WithRequireLeader, client-driven RequestProgress, gRPC keepalives on
both clients, per-unary-RPC deadlines (watch excluded). Version probe
at connect enforces the >=3.4 source floor and gates progress-notify
trust on 3.4.25/3.5.8. Anchored prefix rewrite via the single
documented formula. Lease-backed keys detected and counted
(kv.Lease != 0).

Tests: unit coverage per module plus an embedded-etcd integration
suite covering scan/tail convergence, fence overlap and epoch/role
rejection, compaction during scan and drain, forced-resync
mark-and-sweep, checkpoint resume, prune, and error classification.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Xavier Lange <xrlange@gmail.com>
A single post-heal put can race the recovery attempt's restart backoff
and land below the scan's R0: the scan applies it, the live tail has
nothing to deliver, and no live apply ever clears the latch. Write a
fresh key per poll tick so one is guaranteed to arrive via the tail.

Signed-off-by: Xavier Lange <xrlange@gmail.com>
Deadline-driven pass on the steady-state tail loop: consume arms a timer
from a Run-goroutine-owned deadline and runs the diff-and-repair pass
inline, so it can never overlap the genesis scan, a forced-resync sweep
(which re-arms the deadline — it just produced the same signal), a drain
(whose own verification supersedes it), or itself. The drain repair
branch now records its drift, recordKeyCounts stamps the pass-completion
time on every counts-producing pass — including the count-only drain
verify — keeping the always-on counts contract behind InvariantsHeld,
and Validate rejects negative intervals.

Signed-off-by: Xavier Lange <xrlange@gmail.com>
- publish counts, pass timestamp, and drift in one Snapshot update inside
  reconcilePass; drop the torn caller-side recordDrift
- cancel the live source watch once a periodic/drain repair pass queues
  more than a few seconds of MaxOpsPerSecond pacing, bounding clientv3's
  unbounded per-watcher buffer during a long healthy pass
- correct the ReconcileInterval re-arm, key-count population, and
  maybeReconcile error-flow comments to match actual behavior
- damage the target atomically (single Txn) in the drift test so a pass
  boundary cannot split the damages across two drifts

Signed-off-by: Xavier Lange <xrlange@gmail.com>
@xrl xrl force-pushed the pr/etcdmirror-reconciliation-pass branch from adcdde4 to 43e41de Compare July 10, 2026 06:09
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant