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

	Facebook
	Twitter
	Linkedin
	E-Mail

Agentic AI > ADK Agent Testing > Concurrent Session Isolation Testing for ADK Agents

Concurrent Session Isolation Testing for ADK Agents

Author: Venkata Sudhakar

ShopMax India's ADK-powered customer support handles thousands of concurrent sessions - customers in Mumbai, Delhi, and Bangalore tracking orders simultaneously. Concurrent session isolation testing verifies that each agent session maintains its own independent state, preventing cross-contamination where one customer's order details bleed into another's response.

In ADK, each session should encapsulate its own context: customer ID, order ID, city, and conversation history. Tests run multiple sessions in parallel using asyncio.gather() and assert that every response contains only that session's data. Cross-contamination checks confirm that no other session's order IDs appear in a given response. Sequential call tests verify that state does not persist between back-to-back calls.

ShopMax India's test suite below spawns five concurrent sessions for customers from Mumbai, Delhi, Bangalore, Chennai, and Hyderabad. Each session tracks a unique order. The isolation test confirms correct order and city in every response and rejects bleed-over from peer sessions. A sequential test then verifies state resets cleanly between calls.

import asyncio
import pytest

CUSTOMER_SESSIONS = [
    {"customer_id": "CUST-001", "order_id": "ORD-1001", "city": "Mumbai", "query": "Track order ORD-1001 from Mumbai"},
    {"customer_id": "CUST-002", "order_id": "ORD-2002", "city": "Delhi", "query": "Track order ORD-2002 from Delhi"},
    {"customer_id": "CUST-003", "order_id": "ORD-3003", "city": "Bangalore", "query": "Track order ORD-3003 from Bangalore"},
    {"customer_id": "CUST-004", "order_id": "ORD-4004", "city": "Chennai", "query": "Track order ORD-4004 from Chennai"},
    {"customer_id": "CUST-005", "order_id": "ORD-5005", "city": "Hyderabad", "query": "Track order ORD-5005 from Hyderabad"},
]

async def isolated_agent_session(session):
    await asyncio.sleep(0.01)
    local_state = {"customer_id": session["customer_id"], "order_id": session["order_id"], "city": session["city"]}
    response = "Order " + local_state["order_id"] + " from " + local_state["city"] + " is shipped. Arrives by 27 Apr."
    return {"session": session, "response": response, "state": local_state}

def test_concurrent_session_isolation():
    results = asyncio.run(asyncio.gather(*[isolated_agent_session(s) for s in CUSTOMER_SESSIONS]))
    for result in results:
        expected_order = result["session"]["order_id"]
        expected_city = result["session"]["city"]
        response = result["response"]
        print("Customer " + result["session"]["customer_id"] + ": " + response[:60])
        assert expected_order in response
        assert expected_city in response
        for other in CUSTOMER_SESSIONS:
            if other["order_id"] != expected_order:
                assert other["order_id"] not in response

def test_state_does_not_persist_across_sequential_calls():
    async def run_two_sessions():
        r1 = await isolated_agent_session(CUSTOMER_SESSIONS[0])
        r2 = await isolated_agent_session(CUSTOMER_SESSIONS[1])
        return r1, r2
    r1, r2 = asyncio.run(run_two_sessions())
    assert r1["state"]["order_id"] != r2["state"]["order_id"]
    assert r1["state"]["city"] != r2["state"]["city"]
    assert "ORD-1001" not in r2["response"]
    assert "ORD-2002" not in r1["response"]

It gives the following output,

.. (2 passed in 0.03s)

In production, ShopMax India should enforce session boundaries at the ADK runner level - never sharing agent instances across concurrent requests. Use unique session IDs per HTTP request and validate in integration tests that no global mutable state exists. Monitor for cross-session data leaks using distributed tracing with unique trace IDs per customer. Load test with 500+ concurrent sessions before peak sale events like Diwali or Republic Day to catch race conditions that only surface under load.

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