Parallel Execution, Isolation, Concurrency, and Test Sharding: Diagnostics, Failure Modes, and Production Practices

1. Broken example: several threads share one WebDriver

This is intentionally wrong. Both tasks mutate the same session. One can navigate while the other locates, one can close a window the other owns, and teardown can race with commands.

from concurrent.futures import ThreadPoolExecutor
from selenium import webdriver

shared_driver = webdriver.Chrome()  # intentionally broken shared mutable session


def broken(case_id):
    shared_driver.get("http://127.0.0.1:8765/work?case=" + case_id)
    return shared_driver.title

try:
    with ThreadPoolExecutor(max_workers=2) as pool:
        print(list(pool.map(broken, ["a", "b"])))
finally:
    shared_driver.quit()

2. Separate browsers can still collide through shared AUT data

Use the Chapter 21 fixture’s /claim endpoint. Two fresh browser sessions request the same synthetic identity at nearly the same time. One claim succeeds; the other receives a deterministic collision state. This proves that one-driver-per-test is necessary but not sufficient.

from concurrent.futures import ThreadPoolExecutor
from threading import Barrier
from selenium import webdriver
from selenium.webdriver.common.by import By

barrier = Barrier(2)


def claim_same_identity(label):
    driver = webdriver.Chrome()
    try:
        barrier.wait(timeout=10)
        driver.get("http://127.0.0.1:8765/claim?identity=shared@example.test&hold_ms=450")
        return label, driver.find_element(By.CSS_SELECTOR, "[data-testid='status']").text
    finally:
        driver.quit()

with ThreadPoolExecutor(max_workers=2) as pool:
    print(list(pool.map(claim_same_identity, ["A", "B"])))

Expected result: one claimed and one collision. Repair the data contract by giving each test a unique identity such as claim-A@example.test and claim-B@example.test; do not serialize the whole suite unless the resource is truly global.

3. Port and artifact collisions are ordinary shared-resource races

Two workers binding hard-coded port 8765 can produce “address already in use.” Two workers writing failure.png can overwrite one another. Use one fixture server owned by the suite, or let the OS choose an ephemeral port and publish it through fixture configuration. Namespace evidence by run/test/attempt/session.

from pathlib import Path


def evidence_path(root: Path, run_id: str, test_id: str, attempt: int, name: str) -> Path:
    safe_test = test_id.replace("/", "_").replace("::", "__")
    path = root / run_id / safe_test / f"attempt-{attempt}"
    path.mkdir(parents=True, exist_ok=True)
    return path / name

4. Hidden test order becomes nondeterministic under parallel scheduling

If Test B expects Test A to create data first, parallelism merely exposes an existing dependency. Preserve the failing order/evidence, then move setup into B’s own fixture or a lower-layer API. Do not force alphabetical execution to hide the dependency.

5. More workers can create infrastructure incidents

Runner workers above Grid slots create queue pressure. Sessions above safe CPU/RAM create browser instability. Requests above AUT limits create 429/5xx or slow responses. Retries multiply these loads. A two-attempt retry on 20 simultaneously failing tests can turn 20 failing attempts into roughly 40 browser attempts and worsen the incident.

6. Concurrency diagnostic sequence

1. Preserve first-failure evidence and the original worker/shard count.
2. Confirm Selenium/binding/browser/driver/Grid versions.
3. Confirm target environment, shard manifest, synthetic data IDs, and evidence namespace.
4. Inspect session IDs and prove whether each worker owns a distinct session.
5. Inspect locator/wait/browser context only after ownership is proven.
6. Inspect AUT conflicts, rate limits, server logs, and shared records.
7. If remote, inspect Grid queue, slots, Node health, CPU/RAM and CI executor load.
8. Apply the least destructive correction: unique state, lower concurrency, or measured capacity change.
9. Rerun the smallest controlled scenario before reopening the full matrix.

7. Shortcuts that hide the layer

The following table organizes the key choices and evidence for Shortcuts that hide the layer. Use it together with the surrounding prose so the rows serve as a comparison aid rather than standalone rules.

8. Performance attribution

Separate runner overhead, browser/session startup, Grid queue time, AUT/network latency, artifact IO, and retry cost. “Parallel run took 30 seconds” is not enough. Compare per-test durations and session creation timing, then identify where concurrency adds waiting or saturation.

Knowledge check

Official references and current-version notes

• Selenium 4.47 release notes
• Selenium downloads — current stable client and Grid versions
• Avoid sharing state — Selenium test practices
• Test independency — Selenium test practices
• Getting started with Selenium Grid — capacity guidance
• Grid CLI options — max sessions and queue controls
• Python concurrent.futures — ThreadPoolExecutor