Optimistic-concurrency edit-safety on mutable entities

Ce contenu n’est pas encore disponible dans votre langue.

Status (2026-06-26)

Shipped via PR #167 (Phase 1.16 — edit-safety half) on 2026-06-27. Pattern applied to PATCH /assets/{id}, PATCH /collections/{id}, PATCH /posts/{id}. Frontend modals updated to capture baseline + surface 409 with explicit “overwrite” action. Search-half of Phase 1.16 (/search activation + facets + saved searches) remains for follow-up issue.

Context

The RS-gap audit (2026-06-22) flagged “resource locking” as a P1 fundamental — concurrent edits on the same asset silently lose data today (last-write-wins). RS’s solution (resource.lock_user + save_resource_data guard) is a coarse pessimistic lock taken when a user opens an edit modal and released on save/cancel/timeout. The pattern works but has real downsides:

Stranded locks when a user closes the tab without saving — needs a TTL + admin force-release UI, which RS implements with operator burden.
Federation-hostile — a federated update arriving from a peer would have to wait on or break a local user’s edit lock; the semantics don’t compose with the ArchivePub model where any peer can mutate.
UI brittleness — the “X is editing this” banner needs polling or a websocket; without it the lock surfaces as “save denied” with no explanation.

A purely optimistic approach using HTTP conditional-request headers (RFC 7232) carries the same safety guarantee with no server-side lock state, composes with federation, and matches modern API conventions (GitHub, Linear, Notion all use this shape).

Decision

Every mutable entity carries updated_at TIMESTAMPTZ NOT NULL (already present on asset / collection / post per ADR 0044’s audit-ledger pattern). PATCH endpoints require an If-Unmodified-Since header carrying the client’s baseline updated_at; if the row’s current updated_at is later than the header value, the server returns 409 Conflict with a body shape:

{
  "error": "stale_write",
  "current_updated_at": "2026-06-26T14:22:31.485Z",
  "your_baseline": "2026-06-26T14:21:08.112Z",
  "current_state": { /* the full current entity, so client can diff */ }
}

The client can then either: (a) discard the user’s edits, (b) re-prompt with the diff and let the user merge, or (c) explicitly overwrite by re-submitting with the new If-Unmodified-Since value.

Header choice

If-Unmodified-Since over If-Match because:

updated_at is already authoritative and human-readable in audit logs.
If-Match requires generating + tracking an opaque ETag; needless complexity when the timestamp already serves.
HTTP date parsing handles the format; no custom token vocabulary needed.

Frontend pattern

Edit modals capture the baseline updated_at at modal-open time, send it on save, and surface the 409 with a single explicit action: “This was changed by someone else. View their changes / Overwrite with mine.” No silent re-prompting; no automatic merge — the user always sees what changed.

Federation interaction

A federated update arriving via inbox is treated exactly like a local update for the purposes of updated_at advancement. A local user mid-edit who tries to save after a federated update lands gets the 409 with the federated state visible — same shape as any other concurrent edit. No special federation casing.

Scope (1.16 edit-safety half)

Applied to PATCH on:

asset (per-field updates, sensitivity changes, tag set changes)
collection (membership changes, metadata, theme)
post (body, asset list reordering, state transitions)

NOT applied (out of scope for 1.16):

Comment edits — comment threading already uses append-only semantics; no replace surface to protect
Notification mark-as-read — single-bit toggle; race is benign
User profile edits — single-user surface; concurrent-edit scenarios negligible

What this is NOT

Not a resource lock. No server-side state tracks “who’s editing what.”
Not a merge engine. The 409 hands the user the conflict; the human decides.
Not federation-specific. Same pattern protects single-instance concurrent edits AND inbound federation updates.
Not a replacement for batch-edit isolation. Batch metadata edit (separate 1.16 sub-phase) needs per-asset atomicity but is orthogonal to the optimistic-concurrency surface.

Consequences

Positive:

Zero server-side lock state to maintain, time out, or force-release.
Federation composes naturally — peer updates and local edits share one ordering primitive (updated_at).
Familiar pattern; future API consumers (mobile, third-party tools, MCP clients) hit a standard HTTP convention.
409 response carries the full current state so the client can diff intelligently without a second round-trip.

Negative:

Bandwidth cost on conflicts (full entity in 409 body). Mitigated by conflicts being rare in practice.
Client must remember the baseline timestamp per edit session. Frontend convention handles this; third-party API users need to follow it explicitly.
Sub-second concurrent edits hitting the same exact millisecond can both succeed if updated_at precision is too coarse. Postgres TIMESTAMPTZ resolves to microseconds — adequate.

Alternatives considered

RS-style pessimistic lock (lock_user + TTL). Rejected for the federation-composition and stranded-lock reasons above.
Opaque ETag (If-Match). Rejected — updated_at already serves as the version token without additional bookkeeping.
Last-write-wins with audit-log diffing. Rejected — surfaces silent data loss to the user only after the fact; the optimistic check catches it at the moment that matters.
CRDT-based merge. Rejected — overkill for human-edit cadences on these entities; would require materially different data shape.

Reference

RFC 7232 (HTTP Conditional Requests) — the If-Unmodified-Since header semantics
ADR 0044 — Activities ledger / CQRS-lite; updated_at advancement is the projection-side equivalent of the ledger’s event timestamp
Phase 1.16 — Search + edit-safety (this ADR covers the edit-safety half; search activation is a separate follow-up)
RS-gap audit 2026-06-22 — original “resource locking” P1 finding