Load Testing Your Databricks App

Goal: Find the maximum QPS (queries per second) your Databricks App can support.

Assumes MLflow AgentServer. These templates serve the agent with the MLflow AgentServer — it exposes the POST /invocations endpoint (streaming @invoke/@stream, SSE terminated by [DONE]) the scripts below hit, and emits the per-request MLflow trace used in Step 6. Any app works if it exposes the same /invocations streaming contract.

Before You Start — Gather Parameters

Before beginning, use the AskUserQuestion tool to collect the following from the user:

1. Do they already have deployed apps to test, or do they need to set up new apps?
2. Do they want to mock LLM calls? Mocking isolates infrastructure throughput from LLM latency — useful for capacity planning. Testing without mocks measures end-to-end performance.
3. What compute sizes do they want to test? (Medium, Large, or both)
4. How many worker configurations do they want to test? (e.g., 2, 4, 6, 8 workers)
5. Do they have M2M OAuth credentials (service principal client_id/client_secret)? — Recommended for tests longer than ~30 minutes. If not, guide them to create one.
6. What is their DATABRICKS_HOST? (workspace URL)

Step 1: Set Up Load Testing Scripts

Create a load-test-scripts/ directory in the project with the following files. These scripts are framework-agnostic and work with any Databricks App.

Directory Structure

<project-root>/
  agent_server/                # Existing agent code
  load-test-scripts/           # Load testing scripts (create this)
    run_load_test.py           #   Main CLI — orchestrates Locust tests
    locustfile.py              #   Locust test definition (SSE streaming, TTFT tracking)
    dashboard_template.py      #   Generates interactive HTML dashboard from results
    .env.example               #   Template for env vars
  load-test-runs/              # Test results (auto-created per run)
    <run-name>/
      dashboard.html           #   Interactive dashboard
      test_config.json         #   Test parameters
      <label>/                 #   Per-config Locust CSV results

Required Files

locustfile.py — Locust load test that:

• Sends POST /invocations with {"input": [...], "stream": true} to the app
• Parses SSE stream (data: {json} lines) and counts chunks until data: [DONE]
• Tracks TTFT (time to first data: line) as a custom Locust metric
• Uses M2M OAuth token exchange (client_credentials grant to {host}/oidc/v1/token) with auto-refresh
• Implements StepRampShape — ramps users from step_size to max_users, holding each level for step_duration seconds

run_load_test.py — CLI orchestrator that:

• Accepts --app-url (repeatable), --client-id, --client-secret, --max-users, --step-size, --step-duration, --run-name, --dashboard, --compute-size, --label flags
• Tests each app URL sequentially (isolated metrics per config)
• Refreshes OAuth token before each app
• Runs healthcheck + warmup before each test
• Saves results to load-test-runs/<run-name>/<label>/
• Generates dashboard at the end if --dashboard is passed

dashboard_template.py — Generates a self-contained HTML dashboard with Chart.js:

• KPI cards (best config, peak QPS, lowest latency, total requests)
• Bar charts: QPS by config (median + peak), latency (p50 + p95), TTFT, total requests
• QPS Ramp Progression: line charts with QPS/Latency/Failures tabs and a max-users slider
• Grouped by compute size (medium/large side-by-side)
• Full results table with peak QPS, users at peak, latency percentiles, failure rate
• Can be run standalone: uv run dashboard_template.py ../load-test-runs/<run-name>/

Starter examples: a minimal examples/locustfile.py (streaming, TTFT, M2M OAuth, ramp shape) and examples/validate_with_mlflow.py (server-side reliability + latency and a span breakdown that pinpoints the bottleneck tool) are included to copy from.

Install Dependencies

The load testing scripts use their own pyproject.toml inside load-test-scripts/ to avoid polluting the agent's production dependencies.

# load-test-scripts/pyproject.toml
[project]
name = "load-test-scripts"
version = "0.1.0"
requires-python = ">=3.10"
dependencies = [
    "locust>=2.32,<2.40",
    "urllib3<2.3",
    "requests",
    "mlflow>=3.0",   # Step 6 — validate_with_mlflow.py (search_traces, span breakdown)
    "pandas",        # Step 6 — trace DataFrame reliability/latency stats
]

Then install from within the load-test-scripts/ directory:

cd load-test-scripts/
uv sync

Note: locust>=2.43 has a known RecursionError. Pin to <2.40 to avoid it.

Step 2 (Optional): Mock Your Agent for Load Testing

Mocking is optional — you can skip this step to test your real agent end-to-end (including LLM latency). However, mocking is useful for:

• Capacity planning — isolating Apps infrastructure throughput from LLM latency (which adds 1-30s per request)
• Cost savings — avoiding FMAPI/token usage during load tests
• Reproducibility — getting consistent measurements independent of LLM response variability

How to Mock

The mock timing is controlled by two environment variables (set in app.yaml or databricks.yml):

• MOCK_CHUNK_DELAY_MS — delay between text chunks in milliseconds (default: 10)
• MOCK_CHUNK_COUNT — number of text chunks per response (default: 80)

For OpenAI Agents SDK templates: Create a MockAsyncOpenAI client that replaces AsyncDatabricksOpenAI. It simulates tool call streaming (instant) and text response streaming (delayed chunks). A reference implementation is available at examples/mock_openai_client.py:

from agent_server.mock_openai_client import MockAsyncOpenAI
set_default_openai_client(MockAsyncOpenAI())
set_default_openai_api("chat_completions")

For LangGraph templates: Replace the ChatDatabricks model with a mock that returns pre-built AIMessage objects with tool calls and text content using configurable delays.

For custom agents: Wrap whatever external API calls you make (LLM, vector search, etc.) with mock implementations that return realistic response shapes.

Step 3: Deploy Load Testing Apps

Deploy multiple Databricks Apps with varying compute sizes and worker counts.

Recommended Test Matrix

Configuring Compute Size

Databricks CLI:

databricks apps create <app-name> --compute-size MEDIUM
databricks apps update <app-name> --compute-size LARGE

Databricks UI: Go to Compute > Apps > your app > Edit > Configure > Compute.

Configuring Worker Count

start-server (via AgentServer.run()) accepts a --workers flag directly. Pass the worker count in the command array using a DAB variable — no wrapper script needed:

variables:
  app_name:
    default: "my-agent-medium-w2"
  workers:
    default: "2"

resources:
  apps:
    load_test_app:
      name: ${var.app_name}
      source_code_path: .
      config:
        command: ["uv", "run", "start-server", "--workers", "${var.workers}"]
        env:
          - name: MOCK_CHUNK_DELAY_MS
            value: "10"
          - name: MOCK_CHUNK_COUNT
            value: "80"

targets:
  medium-w2:
    default: true
    variables:
      app_name: "my-agent-medium-w2"
      workers: "2"
  large-w8:
    variables:
      app_name: "my-agent-large-w8"
      workers: "8"

Deploying

databricks bundle deploy --target medium-w2
databricks bundle run load_test_app --target medium-w2

Verify apps are ACTIVE before proceeding:

databricks apps get <app-name> --output json | jq '{app_status, compute_status, url}'

Step 4: Run Load Tests

Authentication — M2M OAuth (Required for Long Tests)

Load tests can run for hours. U2M OAuth tokens expire and break your test mid-run. Use M2M (machine-to-machine) OAuth with a service principal instead.

export DATABRICKS_HOST=https://your-workspace.cloud.databricks.com
export DATABRICKS_CLIENT_ID=<your-client-id>
export DATABRICKS_CLIENT_SECRET=<your-client-secret>

Parameters Reference

Example Commands

cd load-test-scripts/

# Quick single-app test:
uv run run_load_test.py \
    --app-url https://my-app.aws.databricksapps.com \
    --client-id <ID> --client-secret <SECRET> \
    --dashboard --run-name quick-test

# Full matrix — 8 apps, overnight:
uv run run_load_test.py \
    --app-url https://my-app-medium-w2.aws.databricksapps.com \
    --app-url https://my-app-medium-w4.aws.databricksapps.com \
    --app-url https://my-app-large-w8.aws.databricksapps.com \
    --app-url https://my-app-large-w10.aws.databricksapps.com \
    --compute-size medium --compute-size medium \
    --compute-size large --compute-size large \
    --max-users 1000 --step-size 20 --step-duration 10 \
    --dashboard --run-name overnight-sweep

# Multiple runs for statistical consistency:
for RUN in r1 r2 r3 r4 r5; do
  uv run run_load_test.py \
      --app-url ... \
      --client-id <ID> --client-secret <SECRET> \
      --max-users 1000 --step-size 20 --step-duration 10 \
      --run-name my_test_${RUN} --dashboard || break
done

What Happens During a Run

1. Healthcheck — verifies the app streams correctly (receives [DONE])
2. Warmup — sends sequential requests to warm up the app
3. Ramp-to-saturation — steps up concurrent users every step_duration seconds
4. When QPS plateaus despite adding users, you've found the saturation point

Estimated Duration

• (max_users / step_size) * step_duration seconds per app
• With defaults: (300 / 20) * 30 = 15 steps * 30s = ~7.5 min per app
• For 4 apps: ~30 min per run

Step 5: View Results Dashboard

Opening the Dashboard

open load-test-runs/<run-name>/dashboard.html

Regenerating the Dashboard

cd load-test-scripts/
uv run dashboard_template.py ../load-test-runs/<run-name>/

What the Dashboard Shows

• KPI Cards — Best config (peak QPS), overall peak QPS, lowest latency, total requests
• QPS by Config — Grouped bars showing median QPS and peak QPS side-by-side
• Latency by Config — Grouped bars showing p50 and p95 latency
• TTFT by Config — Time to first token (p50 and p95)
• Total Requests Served — How many requests each config handled
• QPS Ramp Progression — Line charts with tabs for QPS, QPS (excl. failures), Latency, and Failures. Includes a max-users slider to zoom into lower concurrency ranges. Charts are grouped by compute size (medium/large).
• Full Results Table — All configs with peak QPS, users at peak, latency percentiles, and failure rate
• Load Test Parameters — Summary of test configuration for reproducibility

Interpreting Results

• Peak QPS — Maximum QPS at any ramp step. This is the throughput ceiling.
• Users at Peak — Concurrent users when peak QPS was achieved. More users beyond this doesn't help.
• Failure Rate — Should be 0% or very low. High rates mean the app is overloaded.
• QPS Ramp Chart — Look for where the line flattens. That's the saturation point.

Step 6: Validate & Diagnose with MLflow Tracing (server-side truth)

Load-testing an agent needs two planes: Locust drives concurrent load and measures the system from the outside (throughput, latency, TTFT, failures), while the agent's MLflow traces show it from the inside (per-request reliability, and where the time goes). On Databricks Apps the templates trace automatically — you get the inside view for free, which turns a QPS number into an explanation. Use Locust as ground truth for wall-clock, traces as ground truth for reliability and internal breakdown.

Reliability (source of truth). Each request is one trace with a state (OK/ERROR) and execution_duration. Cross-check against Locust: Locust can report success for a request that streamed an error (e.g. an FM 429), and can count infra failures (cold starts) that never created a trace.

Getting started — pull the server-side numbers for a run:

import mlflow
mlflow.set_tracking_uri("databricks")
exp = mlflow.get_experiment_by_name("/Shared/<your-experiment>")
df = mlflow.search_traces(locations=[exp.experiment_id], max_results=2000)
ok = df[df["state"].astype(str).str.contains("OK")]
print("error rate:", round(1 - len(ok) / len(df), 3))     # reliability = trace state
print("median ms:", ok["execution_duration"].median())    # latency over successful only
# note: request_time is epoch ms — filter df to your run's time window

Report latency over successful traces and error rate from trace state, then confirm they agree with Locust. If they diverge, trust the server for reliability and Locust for wall-clock latency.

Find bottlenecks (not just totals). Each trace is a span waterfall — child spans nested by parent_id, each recording start_time_ns/end_time_ns — so you can see where a request spends time, which a QPS/latency number can't. Open a trace on the experiment's Traces page for the waterfall, or filter/aggregate spans by type and name (span.type, span.name) to quantify the split: compare CHAT_MODEL (model) vs TOOL span time to see whether the model or the tools dominate, and group TOOL spans by name for per-tool p95 to find the slowest tool. A bottleneck that only appears at high concurrency — a tool's p95 climbing, or the model-vs-tool split shifting — is what saturation looks like inside the agent. examples/validate_with_mlflow.py does exactly this: reliability, successful-request latency, and the model-vs-tool split with per-tool p95.

Troubleshooting

Resources

• Locust — docs: https://docs.locust.io/ · custom load shapes (LoadTestShape): https://docs.locust.io/en/stable/custom-load-shape.html · request events (events.request.fire): https://docs.locust.io/en/stable/api.html
• MLflow Tracing — https://mlflow.org/docs/latest/genai/tracing/ · search traces: https://mlflow.org/docs/latest/genai/tracing/search-traces/ · Span/TraceData shapes: https://mlflow.org/docs/latest/api_reference/python_api/mlflow.entities.html
• MLflow AgentServer — https://mlflow.org/docs/latest/genai/serving/agent-server/
• Databricks — load test an agent app: https://docs.databricks.com/aws/en/agents/agent-framework/load-test-agent-app