	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Agentic AI > ADK Agent Testing > ADK Agent Stress Testing Under Sustained Load

ADK Agent Stress Testing Under Sustained Load

Author: Venkata Sudhakar

ShopMax India's ADK agents handle thousands of concurrent customer requests during peak sale events in Mumbai, Delhi, and Bangalore. Sustained load testing simulates hours of high-traffic pressure rather than a single burst, revealing latency degradation and error rate increases that only appear after the agent has been running under load for several minutes.

The test launches multiple asyncio workers that call the agent concurrently through a semaphore, collecting per-request latency and error flags over a timed window. After the run, p95 latency and error rate are calculated from the collected results. A p95 above 1.5x the baseline or an error rate above 1% flags a failing test.

The example below runs a 3-second sustained stress test with 3 concurrent workers against a mock agent and asserts that p95 latency and error rate stay within acceptable bounds.

import asyncio
import time
import pytest

BASELINE_LATENCY_MS = 150.0

async def mock_agent_call(query):
    await asyncio.sleep(0.05)
    return "Response for: " + query

async def stress_test_worker(sem, results, stop_event, queries):
    idx = 0
    while not stop_event.is_set():
        async with sem:
            start = time.monotonic()
            try:
                await mock_agent_call(queries[idx % len(queries)])
                latency_ms = (time.monotonic() - start) * 1000
                results.append({"latency_ms": latency_ms, "error": False})
            except Exception:
                results.append({"latency_ms": None, "error": True})
            idx += 1

async def run_sustained_stress_test(duration_seconds=5, concurrency=5):
    queries = [
        "Track order ORD-7821 Mumbai",
        "Stock check Samsung TV Delhi",
        "Return policy Hyderabad",
        "Cancel order ORD-9012 Bangalore",
        "Phones under Rs 20000 Chennai",
    ]
    results = []
    stop_event = asyncio.Event()
    sem = asyncio.Semaphore(concurrency)
    workers = [
        asyncio.create_task(stress_test_worker(sem, results, stop_event, queries))
        for _ in range(concurrency)
    ]
    await asyncio.sleep(duration_seconds)
    stop_event.set()
    for w in workers:
        w.cancel()
    await asyncio.gather(*workers, return_exceptions=True)
    return results

def p95_latency(results):
    latencies = sorted(r["latency_ms"] for r in results if not r["error"] and r["latency_ms"] is not None)
    if not latencies:
        return 0.0
    idx = max(0, int(len(latencies) * 0.95) - 1)
    return latencies[idx]

def error_rate(results):
    if not results:
        return 0.0
    return sum(1 for r in results if r["error"]) / len(results)

def test_no_degradation_under_sustained_load():
    results = asyncio.run(run_sustained_stress_test(duration_seconds=3, concurrency=3))
    p95 = p95_latency(results)
    err_rate = error_rate(results)
    print("Total requests: " + str(len(results)))
    print("p95 latency: " + str(round(p95, 1)) + "ms")
    print("Error rate: " + str(round(err_rate * 100, 2)) + "%")
    assert p95 <= BASELINE_LATENCY_MS * 1.5, (
        "p95 " + str(round(p95, 1)) + "ms exceeds 225ms threshold"
    )
    assert err_rate <= 0.01, (
        "Error rate " + str(round(err_rate * 100, 2)) + "% exceeds 1% threshold"
    )

It gives the following output,

Total requests: 178
p95 latency: 52.3ms
Error rate: 0.0%
. (1 passed in 3.08s)

In production, replace mock_agent_call with a real ADK runner call and set duration_seconds to at least 60 for meaningful results. Tune BASELINE_LATENCY_MS using values observed during normal traffic in your staging environment. Run the sustained stress test as a nightly CI job so that gradual latency regressions introduced by model updates or new tool calls are caught before they affect customers.

Send your comments, suggestions or queries regarding this site to [email protected].