max / makenotwork
| 1 | -- Payment finalize idempotency (ultra-fuzz Payments A+, finding M-Pay1). |
| 2 | -- |
| 3 | -- Webhook secondary effects (license-key mint, revenue splits, etc.) run after |
| 4 | -- the transaction is committed. On crash-recovery redelivery the finalizer |
| 5 | -- re-runs, so the side-effecting writes must be structurally safe to repeat. |
| 6 | -- These two indexes are the backstop that makes a double-mint / double-split |
| 7 | -- impossible even if application-level pre-checks are bypassed. |
| 8 | -- |
| 9 | -- NOTE: both indexes are created NON-CONCURRENTLY (sqlx runs each migration |
| 10 | -- inside its own transaction, and CREATE INDEX CONCURRENTLY cannot run there). |
| 11 | -- Acceptable at current table sizes; revisit if license_keys / revenue_splits |
| 12 | -- grow large enough that the brief write lock matters. |
| 13 | |
| 14 | -- One auto-minted license key per purchase. The partial predicate excludes the |
| 15 | -- manually-created keys (transaction_id IS NULL) that creators issue by hand, |
| 16 | -- so it constrains only purchase-minted keys. |
| 17 | NOT EXISTS license_keys_transaction_id_key |
| 18 | ON license_keys (transaction_id) |
| 19 | WHERE transaction_id IS NOT NULL; |
| 20 | |
| 21 | -- One revenue split per recipient per transaction. Lets create_transaction_splits |
| 22 | -- use ON CONFLICT DO NOTHING so a finalize re-run is a no-op. |
| 23 | NOT EXISTS revenue_splits_tx_recipient_key |
| 24 | ON revenue_splits (transaction_id, recipient_id); |
| 25 |