Repair Pulsar Subscription Cursor Replay Gaps

Repair Pulsar Subscription Cursor Replay Gaps — EVIDENCE-LIMITED REVIEW

Scenario received

A consumer group restarts from an unexpected position and appears to skip records.

Evidence inventory

• S1: Pulsar subscription cursor metrics — partial extract described by the requester; original full source not provided here.
• S2: consumer acknowledge logs — partial extract described by the requester; original full source not provided here.
• S3: topic retention policy — NOT PROVIDED; required to validate dependent findings.
• S4: redelivery settings — NOT PROVIDED; required to validate dependent findings.

Case-specific checks

Align mark delete position changes. INCOMPLETE until matching source anchors can be examined; document the observable facts, counterexamples and decision gate.
2.

Compare acknowledgment and processing durability. INCOMPLETE until matching source anchors can be examined; document the observable facts, counterexamples and decision gate.
3.

Check retention eligibility. INCOMPLETE until matching source anchors can be examined; document the observable facts, counterexamples and decision gate.
4.

Test bounded replay and dedupe behavior. INCOMPLETE until matching source anchors can be examined; document the observable facts, counterexamples and decision gate.

Deliverables

Cursor movement ledger. Create an evidence-linked section with observations, missing data, owner and next validation step.

Consumer checkpoint correction. Create an evidence-linked section with observations, missing data, owner and next validation step.

Loss and duplication test plan. Create an evidence-linked section with observations, missing data, owner and next validation step.

Acceptance gate

Proposed recovery explicitly accounts for retention expiry and durable deduplication limits. Do not label this gate passed until supplied source records actually demonstrate it.

Verdict and next action

INCOMPLETE — evidence required. Request S3 and S4 from the authorized owner, reconcile them with S1 and S2, then repeat the four checks and log remaining conflicts. No production change, approval, communication, or test run is claimed.