Techbanq

Cloud · peak load

The system that works all year fails on the day it exists for

Campaign day, the Ramadan rush, the televised launch — the moments the business built towards are the moments the application has never experienced. Everyday slowness is a different problem with a different fix; this one is about the cliff, and the only honest preparation is to hit the cliff in rehearsal first.

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

Structured network cabling running into a switch cabinet

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The anatomy of the day

Hour by hour, how the campaign takes its own system down

Peak-day failures follow a script. Reading it in advance is the point: every stage has an intervention that works before the day and almost none that work during it.

01

The campaign goes out

The SMS blast, the TV spot, the influencer post — traffic multiplies in minutes, not hours, in a shape no normal day produces.

The load model is built from this moment: the forecast peak, its arrival speed and the journey users take, agreed with marketing before the test is designed.

  • CDN
  • Landing pages
02

The first bottleneck saturates

One shared resource — usually the database connection pool — hits its ceiling. Requests queue behind it and response times climb for everyone at once.

The load test found this ceiling weeks earlier, and it was raised, pooled differently or shielded by a cache — because on the day itself it can only be watched.

  • Connection pool
  • Database
03

Retries amplify the damage

Timeouts trigger automatic retries and impatient refreshes, so the real load doubles exactly when capacity is exhausted — the queue becomes a stampede.

Backoff on retries, a request queue with a visible position, and rate limits at the edge are designed in beforehand, so overload degrades the queue rather than the core.

  • Edge
  • Queue
04

Everything fails together

Without a degradation plan, the checkout, the browse pages and the marketing site share one fate — the outage takes the revenue path down along with everything optional.

Graceful degradation is decided in a meeting, not in an incident: recommendations, search filters and personalisation switch off first, so the paying path keeps its capacity longest.

  • Feature flags
  • Checkout

The honest limits

Autoscaling multiplies the servers, not the database behind them

Autoscaling is real and worth having, and it solves exactly one shape of problem: stateless compute that runs out. Web and application tiers that do the same work for every user can be multiplied on demand, and for traffic that arrives gradually, that multiplication genuinely absorbs it. What autoscaling cannot do is multiply the things the new servers share — the database taking the writes, the connection pool in front of it, the payment gateway’s rate limit, the lock on the hot inventory row. Add servers against a saturated shared resource and the queue simply forms in more places; the extra capacity can even deepen the failure by hammering the bottleneck harder.

It also has a speed problem the campaign calendar exposes: scaling reacts to load it has already seen, and a peak that arrives in two minutes outruns instances that take five to boot. For a known event, the honest configuration is pre-scaling — capacity raised before the SMS goes out, on a schedule, with autoscaling kept as the safety margin above the plan rather than the plan itself. Chronic everyday slowness, by contrast, is not a capacity story at all; that diagnosis is a different discipline, covered on its own page.

The usual ceilings

What actually saturates first when the traffic multiplies

The load test decides which of these is yours; the list exists because the CPU graph everyone watches is rarely the answer. The first ceiling is usually a limit somebody configured years ago and forgot.

  1. 01

    The connection pool configured for a quieter era

    A fixed number of database connections shared by every server. Under peak load the pool empties, requests wait for a free connection, and the application looks down while the database sits half idle.

  2. 02

    Memory that survives hours but not minutes of peak

    Sessions, caches and in-flight requests all grow with concurrency. A process that fits comfortably at normal load exhausts memory at peak and is killed mid-request — and its traffic lands on its neighbours, which follow it down.

  3. 03

    The lock on the row everybody wants

    A flash sale concentrates thousands of writes on one inventory row or one counter. The database serialises them, each transaction waits for the last, and throughput collapses exactly where the campaign pointed everyone.

  4. 04

    The third party that did not plan your campaign

    The payment gateway, the SMS provider, the identity service — each has its own rate limit, sized for your normal day. At peak, their throttling reads as your outage, unless the integration queues and degrades deliberately.

The rehearsal

Finding the cliff in a test so the campaign never does

A capacity rehearsal is not a tool run against a URL; it is the event, staged early. The order below matters because each step gives the next one its numbers.

STAGE 01

Model the real event

The forecast peak, how fast it arrives, and the journeys users will actually take — browsing, carting, paying — agreed with the business, because a test of the wrong shape passes and proves nothing. A load profile the business recognises as its day.

STAGE 02

Push a production-like environment to failure

The test ramps past the forecast until something breaks — deliberately, because the goal is to learn where the cliff is and what falls off it first, while the audience is synthetic. The first bottleneck, named with evidence.

STAGE 03

Fix the ceiling, then find the next one

The pool is resized, the hot row redesigned, the cache put in front — and the test runs again, because removing one bottleneck promotes the next. The loop repeats until the failure point sits above the forecast with margin. A measured failure point above the forecast.

STAGE 04

Rehearse the day itself

Pre-scaling schedules, dashboards, the degradation switches and the on-call roster are exercised as a drill against the modelled load, so the team has operated the peak before the peak exists. A team that has already run this day once.

Our role

What we break in rehearsal so your event cannot break it first

  1. 01

    The test, built from your event

    We build the load model with your commercial team and run it against a production-like environment, ramping past the forecast on purpose — a test that never fails was not a test.

  2. 02

    The bottlenecks, removed in order

    We fix what each round of testing exposes — pools, memory ceilings, hot rows, third-party limits — and re-test after each fix, so the margin above your forecast is measured rather than hoped for.

  3. 03

    The degradation plan, signed before it is needed

    We wire the switches that shed optional load and get the business to decide, in advance, what turns off first — because that decision made during an outage is made badly.

How much margin to buy above the forecast is a commercial judgement you own; we make the cliff measurable, push it past your plan, and leave a team that has already survived the day once.

Common questions

What buyers ask

Our app is fine on normal days — why does it only die during campaigns?+

Because a peak is not a busier normal day; it is a different physical regime. Every system has fixed ceilings — a connection pool size, a memory limit, a third party’s rate limit, a lock on one hot row — that normal traffic never approaches, so nothing about ordinary operation reveals them. A campaign multiplies concurrency in minutes and drives the system into the first ceiling, where queues form, timeouts trigger retries, and the retries double the load at the worst moment. That is also why this problem is distinct from chronic everyday slowness: a slow application needs tracing and profiling of its usual behaviour, while a peak failure needs the ceilings found under deliberately excessive load. The system is not lying to you on normal days — it is simply never being asked the question the campaign asks.

Will autoscaling keep us up through the Ramadan peak?+

It will carry the part of the load it can multiply, and that is often not the part that fails. Autoscaling adds stateless servers, so if your ceiling is web-tier compute, it genuinely helps. But Ramadan-scale peaks usually break on what the servers share — the database, its connection pool, a payment gateway’s limit — and adding servers against a shared bottleneck adds queuing, not capacity. It is also reactive: scaling responds to load already arriving, and a post-iftar surge that builds in minutes can outrun instance boot times. For a peak you can see on the calendar, pre-scale on a schedule before it starts, keep autoscaling as the margin above that plan, and load test the shared tiers in advance — those are the components a bigger fleet cannot save.

When the system is overwhelmed, which features should switch off first?+

The ones furthest from revenue and identity, in an order the business has agreed before the day. In practice the shedding sequence runs: personalisation and recommendations first, since generic content serves fine; then search refinements, live counters and anything that fans out into multiple backend calls per page; then non-essential writes like activity feeds and analytics events, which can queue for later. What must keep capacity longest is the path that takes money and the ability to sign in. Two things make this real rather than aspirational: the switches must exist as flags that operations can flip in seconds, and the sequence must be signed off by commercial leadership in a calm meeting — because the same decision, improvised during an outage by whoever is on call, reliably sheds the wrong things.

Rehearse the peak with us

A capacity rehearsal scoped to your next real event — the load modelled, the first bottleneck found under test, and a degradation plan the business has signed.

Start the conversation