RFC-0043 · Design Q&A
Versioned draft snapshot
RFCs, reference catalogs, and evidence reflect Spec 3b2e3e7a114fb826b1445d39a48756b8fff629cd. Counts and pending/status notes belong to that snapshot. Read the dated project update for later progress.
Source repositories are private during staging. Links marked “repository access required” need access to the linked repository.
Q: How are external adapter calls and emitted events reconciled with transaction rollback?
Status: Open
Source: RFC-0043 §§1, 5.1–5.3, 9.2
Issue: The pipeline promises rollback of every effect on failure, but adapter.call can have an external side effect and event.emit can reach subscribers before the graph transaction commits. Neither can generally be rolled back with a database transaction.
Recommendation: Define adapter purity and side-effect declarations. Persist outgoing events and external commands in a transactional outbox, then deliver after commit with replay and compensation rules.
Q: How is the idempotency result committed atomically with a mutation?
Status: Open
Source: RFC-0043 §§6.2–6.3, 9.2 steps 6–12 Issue: The operation commits graph writes at step 11 but persists the idempotency result at step 12. A crash between them permits a replay that performs the mutation again. In-flight locking and its scope are also unspecified. Recommendation: Reserve the scoped key, payload fingerprint, status, and result in the same durable transaction as graph writes, or define an equivalent recovery protocol and concurrency guarantee.
Q: What happens if a caller reuses an idempotency key with different input?
Status: Open
Source: RFC-0043 §§6.1–6.3 Issue: A successful replay returns the original result by key alone, even if the later request supplies different input. This can silently return a result for an unrelated requested mutation. Recommendation: Bind a canonical request fingerprint to the scoped key and reject conflicting reuse with a stable error.
Q: When are guards evaluated relative to the transaction and mutable state?
Status: Open
Source: RFC-0043 §§7.1, 9.2 steps 6–8 Issue: §7.1 says guards run before opening the transaction, while the runtime sequence opens the transaction first. Guards can inspect runtime conditions that change before the first write; the two orders provide different isolation guarantees. Recommendation: Define which guards are pure input checks outside the transaction and which state-dependent guards run under the same isolation and locks as effects.
Q: How is a tenant boundary derived from input authorized before tenant context is set?
Status: Open
Source: RFC-0043 §§3.3, 4, 9.2 steps 2–4
Issue: tenantBoundaryRef may read an input field, and the runtime sets tenant context before authentication and authorization. The RFC does not require the supplied tenant to belong to the principal or constrain cross-tenant probing at this stage.
Recommendation: Authenticate before binding tenant context, validate membership or delegated access to the resolved tenant, and use the same bound tenant for every authorization and effect.
Q: What makes adapter.call safe in a query operation?
Status: Open
Source: RFC-0043 §§3.2, 5.1–5.2
Issue: Queries forbid mutating effects, but adapter.call is marked non-mutating only “per-adapter” without a capability or compile-time check. An adapter operation may perform external writes from a nominal query.
Recommendation: Require adapters to declare a verifiable read-only effect class and reject any query call without that declaration.
Q: Which resource carries workflow state and how is it locked?
Status: Open
Source: RFC-0043 §§8.1–8.3 Issue: The workflow looks up a designated state field on an operation’s subject Resource, but the declaration does not identify the subject or state field. The promised row lock therefore has no defined target, and effect steps may update the same state independently. Recommendation: Add explicit subject and state-field references, define transaction isolation and transition write ordering, and reject effects that bypass the workflow transition rule.