Skip to content

design-notes.md

Latest commit

 

History

History
319 lines (246 loc) · 9.96 KB

design-notes.md

File metadata and controls

319 lines (246 loc) · 9.96 KB
Raw

Treeify — AI test case design copilot

Structured, traceable, high-coverage test design — faster.
treeifyai.com

Try Treeify Free

Design Notes

“Refunds must be predictable, auditable, and exactly-once.”
These notes capture architecture, contracts, failure modes, and test hooks for Refund Workflow v1. Pair with the Brief for outcomes and acceptance.

1) Goals & non-goals

Goals

  • Deterministic, auditable state machine for refunds (full/partial).
  • Exactly-once provider submission and ledger posting.
  • Strong observability and reconciliation with payment provider.
  • Safe under retries, outages, and abuse attempts.

Non-goals (v1)

  • Split-tender, multi-currency conversions, RMA logistics.
  • Chargeback dispute automation (separate process).
  • Subscription proration.

2) Architecture (logical)

[UI / Agent Console]
        |
        v
[Refund API]  --(RPC)-->  [Policy Service]
     |                    [Ledger Service]
     |                    [Provider Adapter(s)]
     |                    [Attachments Store]
     |                           |
     |                           +-- PSP(s) / Bank rails
     |
     +--(DB)-----> Refunds (state), Idempotency store, Outbox/Inbox
     +--(Cache)--> Short-lived reads (refund status)
     +--(Events)--> Event Bus (refund.*)

Patterns

  • Outbox: queue provider submissions transactionally with refund state.
  • Inbox: dedupe provider webhooks by provider_refund_id + hash.
  • Saga: refund approval → submission → pending → completion/failure.

3) State machine (detailed)

requested → approved → submitting → provider_pending → completed | failed | canceled

Transitions

  • requested → approved (auto or agent decision).
  • approved → submitting (enqueue job; mark attempt).
  • submitting → provider_pending (accepted by provider, async outcome).
  • provider_pending → completed/failed (webhook or poll).
  • requested|approved → canceled (if allowed before submit).

Rules

  • Only approved refunds may be submitted.
  • Amount cannot exceed remaining refundable.
  • State transitions are idempotent; repeated webhooks are safe.

4) Data model (persistence)

CREATE TABLE refunds (
  refund_id TEXT PRIMARY KEY,
  order_id TEXT NOT NULL,
  user_id TEXT NOT NULL,
  amount_minor INTEGER NOT NULL,
  currency TEXT NOT NULL,
  reason TEXT NOT NULL,                         -- taxonomy in brief
  state TEXT NOT NULL CHECK (state IN (
    'requested','approved','submitting','provider_pending',
    'completed','failed','canceled'
  )),
  provider_refund_id TEXT,
  provider_attempts INTEGER NOT NULL DEFAULT 0,
  last_error_code TEXT,
  last_error_at TIMESTAMPTZ,
  idempotency_key TEXT,                         -- for create
  created_at TIMESTAMPTZ NOT NULL,
  updated_at TIMESTAMPTZ NOT NULL
);
CREATE UNIQUE INDEX refunds_order_remaining_guard
  ON refunds(order_id, refund_id);              -- combined with ledger guard

CREATE TABLE idempotency_keys (
  key TEXT PRIMARY KEY,
  request_hash TEXT NOT NULL,
  response_json JSONB NOT NULL,
  created_at TIMESTAMPTZ NOT NULL
);

CREATE TABLE inbox_webhooks (
  provider TEXT NOT NULL,
  event_id TEXT NOT NULL,
  hash TEXT NOT NULL,
  received_at TIMESTAMPTZ NOT NULL,
  PRIMARY KEY(provider, event_id)
);

Ledger

  • Post REFUND_PENDING on approval; REFUND_SETTLED on completion.
  • Guard remaining balance with sum of completed + approved ≤ captured.

5) Contracts (API)

Create refund (idempotent)

POST /v1/orders/{order_id}/refunds

Headers: Idempotency-Key, X-Correlation-Id
Body (simplified):

{ "amount_minor": 2500, "currency": "USD", "reason": "not_received", "evidence": [] }

Success 202:

{ "refund_id":"rf_123","state":"approved","message_id":"refund.request.accepted" }

Errors:

  • 400 ERR.VALIDATION.amount.range
  • 400 ERR.BUSINESS.refund.exceeds_remaining
  • 402 ERR.BUSINESS.refund.not_captured
  • 403 ERR.AUTHZ.scope
  • 409 ERR.CONFLICT.idempotency

Agent decision

POST /v1/refunds/{refund_id}/decision { "decision":"approve|deny", "note":"..." }

Submit refund (system)

POST /v1/refunds/{refund_id}/submit → enqueues provider job; returns 202.

Read/list

GET /v1/refunds/{refund_id}
GET /v1/orders/{order_id}/refunds

Webhooks (provider)

POST /webhooks/payments with refund.succeeded|refund.failed (signed + timestamp).

6) Idempotency & exactly-once

  • Create: store request hash + response; on replay with same key/identical hash → return stored response; with different hash → 409.
  • Submit: generate provider idempotency key per provider contract; store provider_refund_id once; prevent duplicate posts.
  • Webhooks: dedupe by (provider,event_id); if already applied → no-op.

7) Provider adapters

  • One adapter per PSP with uniform interface:
    • submit_refund(order_ref, amount_minor, currency, key) -> { provider_refund_id, state }
    • verify_webhook(headers, body) -> { event_id, type, provider_refund_id, status }
  • Map PSP-specific states to our state machine.
  • Handle clock skew and signature validation.

8) Reconciliation

  • Daily job joins ledger entries with provider refunds by provider_refund_id.
  • Emit refund_mismatch_rate_pct.
  • Export CSV of mismatches with reasons (missing webhook, amount mismatch, currency mismatch).
  • Alert when mismatch exceeds threshold.

9) Security, privacy, abuse

  • RBAC: only agents with scope can approve/deny; dual-control for goodwill > threshold.
  • Rate-limit refund create & decision endpoints.
  • No PAN or sensitive PII in logs; attachment storage encrypted; signed URLs with short TTL.
  • Prevent over-refund by transactional check against captured minus prior refunds.

10) Observability (signals)

Logs (msgid)

  • MSG.refund.requested
  • MSG.refund.approved
  • MSG.refund.submitted
  • MSG.refund.completed
  • MSG.refund.failed

Error codes

  • ERR.BUSINESS.refund.not_captured
  • ERR.BUSINESS.refund.exceeds_remaining
  • ERR.CONFLICT.idempotency
  • ERR.DEPENDENCY.timeout

Metrics

  • refund_create_latency_ms (histogram)
  • refund_outcome_total{state}
  • refund_mismatch_rate_pct
  • webhook_replay_dedup_total

Traces

  • Spans: policy.evaluate, ledger.post, provider.submit, webhook.process.
  • Attach exemplars to histograms.

11) Failure modes & mitigations

  • Provider timeout/5xx → mark provider_pending; retry with backoff + jitter; surface ETA to user.
  • Webhook lost → poll status via adapter; reconcile job detects gap.
  • Idempotency key collision → 409 with prior response; never overwrite.
  • Ledger post failure → outbox ensures retry; state blocked from completed until ledger write succeeds.
  • Partial provider outage → circuit breaker opens; queue submissions; show “processing” state.

12) Performance & capacity

  • Budgets: create p95 ≤ 250 ms; status read p95 ≤ 150 ms.
  • Expected throughput aligned with order volume; queue sized for burst 3×.
  • Storage: attachments limited by type/size/count; virus scan pipeline.

13) UX & a11y hooks

  • Message IDs for each state; short actionable copy.
  • ARIA live updates on status changes; focus returns to actionable control.
  • Emails/notifications localized; timelines vary by rail (card vs bank).
  • Dark mode/RTL verified; long text expansion ×1.3.

14) Testing strategy

Unit: amounts, remaining logic, state transitions, idempotency hashing.
Contract: OpenAPI schemas; reject unknown fields.
Integration: provider simulators (succeeded/failed/timeout); signature checks.
E2E: MAE scenarios from the brief with evidence bundle.
Chaos: provider timeouts and webhook delays.
Telemetry assertions: msgid/err.code in logs; metrics changes; traces present.

15) Open questions

  • Dual-control trigger thresholds by currency—static or FX-adjusted?
  • Allow refund to different instrument (out of scope v1 but design migration path)?
  • SLA visualization to customer—show expected date by rail?

16) API schema stub (OpenAPI fragment)

post:
  /v1/orders/{order_id}/refunds:
    headers:
      Idempotency-Key: string
      X-Correlation-Id: string
    body:
      type: object
      required: [amount_minor, currency, reason]
      properties:
        amount_minor: { type: integer, minimum: 1 }
        currency: { type: string, pattern: "^[A-Z]{3}$" }
        reason: { $ref: "#/components/schemas/RefundReason" }
        evidence: { type: array, items: { $ref: "#/components/schemas/Attachment" } }
    responses:
      "202": { $ref: "#/components/schemas/Refund" }
      "400": { $ref: "#/components/schemas/Error" }
      "402": { $ref: "#/components/schemas/Error" }
      "403": { $ref: "#/components/schemas/Error" }
      "409": { $ref: "#/components/schemas/Error" }

17) Links

  • Brief → ./brief.md
  • API Coverage → ../../60-checklists/api-coverage.md
  • Performance → ../../60-checklists/performance-review.md
  • Security → ../../60-checklists/security-review.md
  • SRE (SLOs & breakers) → ../../57-cross-discipline-bridges/for-sres.md
  • Traceability → ../../65-review-gates-metrics-traceability/traceability.md

Built with Treeify — AI test case design copilot.
Structured, traceable, high-coverage test design — faster.