An enrolment platform that handles twenty families at a time can fall flat at two thousand, on the morning the window opens at nine and half the town signs in inside the same quarter of an hour. The pattern repeats when a ticket sale starts, after a radio mention, or during the first hour of a promotion. A load test is how that limit gets met in October, on a duplicate environment, rather than live in front of the people you meant to serve.
A load run earns its keep only when it looks like a genuine busy day. So the starting point is your analytics and your server logs, never a round number somebody suggested in a meeting.
How many of your visitors open a product page, how many reach for search, how many create an account, how many arrive at card payment through Stripe or PayPlug. Those proportions come out of your own measurements rather than our imagination.
Scenarios built in Gatling, JMeter or k6, with reading pauses, sign-in, varied content and sessions that last. A robot hammering one address proves nothing whatsoever.
Virtual users arrive in batches, and we record the batch at which response time crosses two seconds, then the batch at which timeouts and 500 errors begin.
Steady traffic held for several hours. Memory leaks, queues that back up and disks quietly filling with log files never reveal themselves inside the first five minutes.
Processor, memory, disk activity, database, outside services. More often than not the whole thing hangs on one query without an index, a listing page with no cache, or a connection pool set too narrow.
Once traffic falls away, does the system pick itself up, or must somebody restart a service by hand? Three minutes of downtime against two hours hangs on that answer.
Which changes belong in the code, which in the configuration and which in the hosting, in what order, and the gain expected from each. Renting a bigger machine rarely sits at the top of the list.
Writing the scripts occupies a handful of working days. Each run is short, and they get repeated after every repair. The load comes out of the cloud; your servers receive no installation of any kind.
Concurrent users on one side, an acceptable response time on the other. Lacking those two figures, a run can neither pass nor fail.
A duplicate of the system, carrying a database of roughly production size. Everything answers instantly on an empty database, and the figures then tell you nothing.
Stepped ramp-up, a measurement at every level, with your own server monitoring open alongside for the whole duration.
A report naming each bottleneck, remedies ordered by gain against effort, then a further run once the profitable ones have shipped.
Outside services usually give way before your own servers do. The payment platform, the carrier API, the mail delivery service and the product feed connector each enforce rate limits of their own. Two honest options exist: warn them and obtain written agreement, or swap them for stubs while the run lasts. Either way it belongs in the conclusions, stated plainly.
Better not. You risk blocking genuine visitors and spoiling a whole day of measurements. A duplicate environment stays the right answer. Failing that, the run happens at night, throughput capped, inside a window you have agreed in writing.
Your peak, never your average, plus a reserve factor of two or three. A monthly average says nothing useful: traffic shows up in waves, when a mailing goes out, when the press mentions you, when an online service window opens.
With caching and with database queries: the cheapest lever, and the one whose gains show up most plainly. Resizing the hosting comes second, reworking the architecture third, once the first two are spent.
Almost always. A surge out of nowhere from four or five IP addresses looks exactly like a denial of service attack, and filtering at the host or the CDN cuts the line midway. We help draft the request and book the slot with OVHcloud, Scaleway or whoever hosts you.
Give us the traffic you expect and the date it falls on. The load gets simulated and you see where the trouble begins.
Message received
An answer follows inside one working day. Report an outage that is stopping people working and it moves ahead of everything else.
Nothing here under that name. Check the spelling, or simply choose the nearest large city instead. Since every engagement runs remotely, whichever you pick changes nothing about what we do for you.
