Moving to the cloud came with a comforting promise: traffic goes up, capacity goes up, problem solved. Autoscaling is real and it works. But it only scales the things that scale, and most cloud outages are caused by the things that don't.
Elastic capacity grows your compute to meet demand. Fixed constraints are the limits that don't grow with it — the connection pool, the rate limit, the licence cap, the single-threaded dependency. Load testing in the cloud is really about finding those, because autoscaling will never save you from them.
More instances don't help when the bottleneck is somewhere a new instance can't reach:
Database connection limits that saturate long before compute does
Third-party API rate limits that throttle you regardless of your capacity
Cold starts that add latency exactly when a spike demands speed
Shared queues or locks that serialise work no matter how many workers you add
You can scale the front end infinitely and still be gated by a database that accepts a fixed number of connections.
Assumes elasticity
Ramp traffic, watch it scale
Measure average response
Trust the platform
Pass at target load
Finds the constraint
Push until something stops scaling
Measure where latency turns non-linear
Trust the test
Name the first fixed limit hit
A cloud load test that only proves autoscaling works has proven the easy part. The valuable output is the first fixed constraint the system hits and how far away it is.
Guide: IR's Cloud Load Testing guide covers testing for fixed constraints in elastic environments, downloadable as a PDF. → https://www.ir.com/guides/cloud-load-testing
IR Infrastructure, powered by Prognosis, gives you visibility into the layers autoscaling can't stretch, correlating application load with the databases, dependencies, and shared resources underneath it. When a test pushes the system toward a fixed limit, you see which constraint is about to bind before it becomes an outage, instead of discovering it live when the spike arrives.
The cloud didn't remove your failure modes. It moved them somewhere autoscaling can't see. Test for the fixed limits, not the elastic ones, and the spike that would have taken you down becomes a constraint you raised on purpose, on a Tuesday, with no one watching.