load-testing

You can find the maximum sustainable queries per second for a Databricks App by running controlled Locust-based load tests that ramp concurrent users against your invocations endpoint, tracking QPS, latency, and failure rates to identify the throughput ceiling.

Yes, you can isolate infrastructure throughput from LLM latency during load testing by enabling optional mocked LLM streaming, which removes variable model response times to benchmark the raw infrastructure capacity of your Databricks App accurately.

Yes, you need OAuth client credentials using the M2M service-principal client_credentials flow to authenticate long-running load tests reliably, ensuring sustained benchmarking sessions against your Databricks App endpoints without token expiration interruptions.

Load testing streaming SSE endpoints collects queries per second, percentile latency, time to first token (TTFT), and failure-rate evidence, which are visualized in an interactive HTML dashboard showing the QPS ramp progression and saturation behavior.

Locust handles capacity planning for agent-style invocations by orchestrating ramp-step load tests with concurrent users, measuring peak QPS and saturation behavior to produce repeatable performance benchmarks across different compute sizes and worker counts.

Locust-based performance troubleshooting is limited to evaluating Databricks Apps implementing streaming Responses-style invocations, requiring OAuth M2M authentication and depending on urllib3 and requests libraries to execute the controlled load tests.