Loading…
A short voice session: one question, one follow-up, then instant feedback — up to 5 minutes.
The bottleneck may be shared, the new instances may not be ready, or the scaling signal may lag. More replicas can also overload a constrained dependency.
Choose signals tied to the workload, such as useful throughput, queue age, or saturation, rather than CPU by default. Account for startup time, node provisioning, limits, and connection capacity. Set stabilization and scale-down behavior to avoid oscillation. Load-test the complete path and define overload controls for demand that scaling cannot absorb.