Techbanq

Restaurants · peak reliability

Friday at eight is a load test the whole city runs on you

Food ordering is not steady traffic with occasional spikes; it is a platform that earns most of its living inside a few evening hours, and in Ramadan inside minutes. Engineering for that shape — and rehearsing it before the season arrives — is the difference between a rush and an outage.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Structured network cabling running into a switch cabinet

Information security for software delivery

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The shape of demand

A whole city gets hungry inside the same two hours

We do not have an office in Saudi Arabia

We serve organisations there from our Riyadh office and our engineering hubs in Lahore and Karachi, travelling for the parts of an engagement that need to happen in person.

Most web platforms meet their users spread across a day. Food ordering does not. Dinner arrives for everyone at once, weekends multiply it, and the order that was going to be placed at some point is placed now, together with everyone else's. The platform's year is decided in these windows: a system that hums along the daytime trough and folds at the evening crest has failed at its actual job, however good its monthly averages look.

Ramadan sharpens the curve into a spike. The iftar order is not spread across an evening; it clusters around a single minute that the whole market observes together, and the platform meets in one short window the demand another month spreads across hours. A Saudi ordering platform that has not planned for Ramadan has not planned for its own busiest month — the season is in the calendar, which makes being surprised by it a choice.

The failure, when it comes, is also self-deepening. A slow checkout makes hungry guests refresh and resubmit, which multiplies the load that made it slow; client apps retry on their own schedules; and a platform at the edge of saturation is pushed over it by its own users' impatience. This is why peak engineering is less about raw capacity than about behaviour at the boundary — what the system does in the minute it stops keeping up decides whether the evening is a slowdown or a story.

The usual suspects

The parts of an ordering platform that give way first

Peak failures repeat across ordering platforms with remarkable consistency. These are the four places we look before anywhere else.

01

Payment callbacks under load

Checkout waits on a payment provider's response, and at peak those responses slow just as their volume multiplies. Connections held open pile up until the whole platform queues behind its slowest partner — unless timeouts, retries and an idempotent confirmation path are designed for exactly this hour.

02

Menu media weight

A menu of full-resolution photography served to thousands of phones on cellular connections is a bandwidth bill and a slow first screen at the moment patience is shortest. Sized variants, aggressive caching and a CDN doing the carrying turn the menu from a peak liability into the one part that never struggles.

03

Database saturation on hot rows

Every order at peak contends for the same few records — stock counters, branch state, the orders table — and locks that never met each other on Tuesday queue into deadlock on Friday. Read replicas, queue-based writes and contention-aware design are decided in architecture, not discovered in the incident.

04

The retry storm

The guests, their apps and your own services all retry when responses slow — and untamed, those retries triple the load at the worst possible minute. Backoff on clients, idempotency on order submission and a front door that sheds excess politely keep a wobble from amplifying itself into a fall.

The rehearsal

Rehearse Ramadan before Ramadan rehearses you

Peak readiness is a rehearsal, not a purchase. The sequence below is run weeks before the season, so its findings can be fixed rather than merely confirmed.

STAGE 01

Model the peak from your own history

Last year's order curves — the evening window, the weekend multiple, the iftar minute — plus this year's growth define the target load. A rehearsal against a guessed number rehearses the wrong evening. A target curve the whole team agrees to engineer for.

STAGE 02

Drive the real order path at a multiple

Synthetic load walks the genuine journey — browse, build a basket, pay against a sandbox, receive confirmation — because a test that skips payment skips the stage most likely to fold. The first bottleneck, found on a quiet morning instead of a live Friday.

STAGE 03

Break it on purpose and watch the order of failure

Load is pushed past the target until something gives, because what fails first, and how loudly, is the information the degradation ladder is built from. A ranked list of weak points with the fix for each.

STAGE 04

Agree the degradation ladder in writing

What sheds first when the platform strains — recommendations off, live tracking coarser, the cached menu carrying browsing, orders queuing with an honest confirmation — is decided now, so the worst evening runs a plan instead of an argument. A ladder the on-call engineer can climb without waking a committee.

STAGE 05

Staff the first night like a launch

The first Friday of the season gets eyes on the dashboards, a rollback for anything shipped that week, and a freeze on changes that can wait — because the cheapest incident is the one whose response began before it started. A season opened deliberately rather than survived.

Our role

Peak engineering we deliver and the evenings you keep serving

  1. 01

    The readiness review

    We examine the platform against its own demand curve — payment path, media delivery, database contention, retry behaviour — and return a ranked list of what gives way first, with the effort each fix costs.

  2. 02

    Degradation designed, not improvised

    We build the ladder into the platform — the cached menu that keeps browsing alive, order queuing with honest confirmation, feature shedding by switch — so strain produces a slower service instead of a blank screen.

  3. 03

    The rehearsal, run with your team

    We run the load rehearsal alongside your engineers and leave the harness, the scripts and the runbooks behind, so next season's rehearsal is yours to run without us.

How much capacity headroom to pay for is a commercial judgement made against your own peak, not a constant we impose; we make the curve, the weak points and the cost of each rung visible so the judgement is an informed one.

Common questions

What buyers ask

What should guests see while the ordering platform is struggling?+

A slower, smaller, honest service — never a spinner that leads nowhere. The menu should stay readable from cache even when ordering is strained, because a guest who can browse will often wait where a guest facing a blank page opens an aggregator app. If order submission is backed up, take the order into a queue and say so plainly, with a confirmation that arrives when it is genuinely accepted — the deadly pattern is accepting payment against an order the kitchen never receives. And if the platform truly cannot take orders, the honest closed sign, with a time to come back, preserves more goodwill than an error page pretending to be temporary. All of this has to be designed and wired in advance; a degradation story cannot be written at eight on Friday.

How early before Ramadan should load preparation start?+

Early enough that the findings can be fixed, which in practice means starting the readiness review about two months out and running the first full rehearsal no later than four to six weeks before the season. The rehearsal exists to find problems, and the problems it finds — a database that needs a replica, a payment path that needs idempotency, images that need a CDN — take weeks to fix properly, not days. A rehearsal run the week before Ramadan can only confirm what will happen; run six weeks before, it changes what will happen, and it leaves a second rehearsal slot to verify the fixes under load rather than trusting them on the first live night.

We are on the cloud already — why does the platform still fail at peak?+

Because the cloud rents you capacity, not architecture. Autoscaling multiplies web servers well, but the dinner peak rarely fails there: it fails in the parts that do not scale by adding copies — the database rows every order contends for, the payment provider whose callbacks slow under everyone's load at once, the retry behaviour that multiplies traffic precisely when capacity is shortest. Scaling can even sharpen the failure, as a larger web tier delivers more concurrent load onto the same saturated database. The peak work is therefore architectural: contention designed out, queues absorbing the spike, degradation shedding weight — with autoscaling as one useful tool inside that design rather than a substitute for it.

Rehearse your peak with us

A peak-readiness review that models your real demand curve and returns a prioritised list of what would fail first and what to rehearse before the next season.

Start the conversation