Beyond the controller.

Wahda Bank's digital platform started life as a well-behaved Laravel application. By the time it was carrying a meaningful share of the country's retail payments it was mature in the way that word really means: hundreds of endpoints, Eloquent models with real money in them, integrations with a core-banking system and instant payment rails — OnePay, LYPAY — a mobile app in the field that people would not update, and a business that could not tolerate a bad Tuesday.

The problem arrived on salary days. Every transfer was processed inline in the request: validate, call core banking, write the ledger, notify, log. Under peak load the connection pool filled, lock waits on account rows cascaded, and a notification provider having a slow hour could make a payment feel slow.

The obvious answer — “rewrite the payments layer in something fast” — was also the most dangerous one. A rewrite is a promise to rediscover every edge case a mature system has already paid for.

So I didn't rewrite. I investigated the existing behaviour first, drew a line around the one thing that actually needed to be fast — the transfer hot path — and treated everything else as a scheduling problem.

• Everything that wasn't the money left the request. Notifications, statements, analytics and fraud scoring became Redis-backed queued jobs (Horizon) with explicit retry and idempotency semantics. Events carried the rest. The request that remained did one thing.
• The hot path moved into a small Go service behind the exact contract Laravel had always exposed. Laravel stayed the system of record and the orchestrator; Go became a worker that is very good at one narrow job. They speak through a versioned interface with contract tests, so the mobile app in the field never knew anything changed.
• Correctness first, speed second. Idempotency keys on every money-moving call; a reconciliation job comparing the new path against the old on shadow traffic before a single real transfer used it; per-cohort feature flags so rollout was a dial, not a switch.

Performance you can't afford to keep is technical debt with better PR. So each shortcut had an expiry: the parallel code path was written up in an ADR with a delete-by release; the dual-write comparison came out as soon as it had been quiet for two cycles; any query I touched left with an index and a test.

I let the old code stay ugly where it was safe, and made it clean only where it was hot — after I understood why it looked that way.

The next salary day was boring, which was the point. Peak-hour latency stopped tracking queue depth, a slow vendor stopped being able to slow a payment, and the mature codebase got a little smaller instead of a little scarier.

That platform now moves roughly 156 billion LYD a year across 143 million sessions, 77 million transactions, and 601 thousand users — and it still runs on Laravel where Laravel is right, and Go where Go is right.