Struct CdcFatalDecision
The decision a CDC processor's consume loop must act on after each iteration: whether it may advance its durable checkpoint and whether it must stop.
public readonly record struct CdcFatalDecision : IEquatable<CdcFatalDecision>
- Implements
- Inherited Members
Remarks
Produced exclusively by Decide(Exception?, IMessageFailureClassifier?). The safety invariant — a fatal (or transient) fault never advances the durable checkpoint past the unprocessed change — is enforced by the idiom native to each provider class:
- Poll-batch providers (Cosmos, DynamoDb) gate the durable advance literally
on AdvanceCheckpoint (
if (decision.AdvanceCheckpoint) { …Confirm… }), reached on both the success and the captured-fault path, so theif(true)mutant advances past a failed change and turns the field-gate lock RED — structurally non-vacuous. - Streaming providers (Postgres, Mongo) enforce the same invariant by confirm-site placement at the transaction-commit / invalidation boundary: a mid-batch fault unwinds out of the consume method before the confirm site is reachable. A field-gate there would be vacuous (the confirm point never sees a Decide(Exception?, IMessageFailureClassifier?)-evaluated fault), so the placement is regression-locked by the non-skipped real-infra restart-redelivery test instead.
A "reconnect-on-transient" signal is intentionally NOT a field here: it is the implicit else of a non-fatal fault (Stop is false); a future reconnect-policy knob would be a tracked feature bead, not a silent unread field.
Constructors
CdcFatalDecision(bool, bool)
The decision a CDC processor's consume loop must act on after each iteration: whether it may advance its durable checkpoint and whether it must stop.
public CdcFatalDecision(bool AdvanceCheckpoint, bool Stop)
Parameters
AdvanceCheckpointbooltrue only on the clean-success path; false on ANY fault (fatal or transient) so the durable checkpoint is never advanced past an unprocessed change.
Stopbooltrue when the fault is fatal (non-retryable) — the processor stops loudly rather than silently reconnecting forever. When false on a fault, the loop reconnects and retries from the (un-advanced) checkpoint.
Remarks
Produced exclusively by Decide(Exception?, IMessageFailureClassifier?). The safety invariant — a fatal (or transient) fault never advances the durable checkpoint past the unprocessed change — is enforced by the idiom native to each provider class:
- Poll-batch providers (Cosmos, DynamoDb) gate the durable advance literally
on AdvanceCheckpoint (
if (decision.AdvanceCheckpoint) { …Confirm… }), reached on both the success and the captured-fault path, so theif(true)mutant advances past a failed change and turns the field-gate lock RED — structurally non-vacuous. - Streaming providers (Postgres, Mongo) enforce the same invariant by confirm-site placement at the transaction-commit / invalidation boundary: a mid-batch fault unwinds out of the consume method before the confirm site is reachable. A field-gate there would be vacuous (the confirm point never sees a Decide(Exception?, IMessageFailureClassifier?)-evaluated fault), so the placement is regression-locked by the non-skipped real-infra restart-redelivery test instead.
A "reconnect-on-transient" signal is intentionally NOT a field here: it is the implicit else of a non-fatal fault (Stop is false); a future reconnect-policy knob would be a tracked feature bead, not a silent unread field.
Properties
AdvanceCheckpoint
true only on the clean-success path; false on ANY fault (fatal or transient) so the durable checkpoint is never advanced past an unprocessed change.
public bool AdvanceCheckpoint { get; init; }
Property Value
Stop
true when the fault is fatal (non-retryable) — the processor stops loudly rather than silently reconnecting forever. When false on a fault, the loop reconnects and retries from the (un-advanced) checkpoint.
public bool Stop { get; init; }