Enable multiple k6-browser instances to run concurrent tests against a single browser

[inancgumus]

Problem

When we run multi-VU/instance tests, one of the test runs can close a page of its own. Since each test run attaches to all the browser pages (this is another area we might want to discuss), this sometimes creates a race between the test runs. Here is a step-by-step explanation.

There are two instances and different sessions. However, both instances attach/detach from the same pages since we run in the same browser. instance2 panics while trying to send CDP messages to a closed web socket connection since it detaches from it (because instance1 closes it) and tries to attach to it concurrently (race).

Instance	Notes	Instance Attachments	Browser pages
🔴 i1	newPage		page1
🔴 i1	attach to page 1	page1	page1
🟢 i2	newPage	page2	page2, page1
🔴 i1	attach to page 2	page2, page1 (!)	page2, page1
🟢 i2	attach to page 1 🚨 ...note that this operation continues concurrently...	page2, page1...	page2, page1
🔴 i1	page1.Close()	page2, page1	page2
🔴 i1	detach from page 1 closes page1 session for instance1. these sessions are not shared with other instances. they are specific to this instance. however, the pages on the browser are shared between instances.	page2	page2
🟢 i2	detach from page 1 closes the page1 (targetID) session for instance2.	page2	page2
🟢 i2	attach to page 1 still continues... 🚨panic: session does not exist. There was a panic because instance2.attach to page 1 operation was going on before instance1 closed page1.	page2, page1...	page2, page1

sequenceDiagram
    participant instance1
    participant instance2
    participant Browser pages
    instance1->>Browser pages: newPage
    instance1->>page1: attach to page 1
    instance2->>Browser pages: newPage
    instance1->>page2: attach to page 2
    instance2->>page1: attach to page 1: did not end yet.
    instance1->>page1: page1.Close()
    instance1->>+page1: detach from page 1
    instance2->>+page1: detach from page 1
    instance2->>+page1: attach to page 1 🚨 panic

Loading

Fix

This check allows the test run to continue instead of panicking because of another test run's page detachments. The fix allows running the k6 browser in high concurrency with multiple VUs and instances. Note that it is still possible to get errors since we need to correctly handle the CDP messages while sending and receiving them (the order of them). This can cause timeout and other sorts of errors.

Testing

Here's the test script I used to test this:

import { check } from 'k6';
import { chromium } from 'k6/x/browser';

export default async function () {
  const URL = "ws://127.0.0.1:9222/devtools/browser/035490d7-d3d2-4426-9e38-450adcc7cd74";
  
  console.log('Connecting to browser... VU:', __VU, 'ITER:', __ITER);  
  const browser = chromium.connect(URL);

  console.log('Creating new context... VU:', __VU, 'ITER:', __ITER);
  const context = browser.newContext();

  console.log('Opening new page... VU:', __VU, 'ITER:', __ITER);
  const page = context.newPage();

  console.log('Navigating to website... VU:', __VU, 'ITER:', __ITER);
  await page.goto('https://test.k6.io/', { waitUntil: 'networkidle' });

  console.log('Getting page title... VU:', __VU, 'ITER:', __ITER);
  console.log('page title:'+ page.title(), 'VU:', __VU, 'ITER:', __ITER);
  check(page, {
     'title': p => p.title() == 'Demo website for load testing',
  });
  
  console.log('Closing page... VU:', __VU, 'ITER:', __ITER);
  page.close()
  console.log('Closing context... VU:', __VU, 'ITER:', __ITER);
  context.close()
  console.log('Disconnecting from browser... VU:', __VU, 'ITER:', __ITER);
  browser.close()

  console.log('❌ connected : ', isConnected(browser));
  console.log('Test completed.');
}

function isConnected(browser) {
  return browser.isConnected() ? '✅' : '❌';
}

Here's the bash script (multik6b.sh) that can run multiple tests (a courtesy of @ankur22 🙇):

#!/bin/bash

NUM_K6_INSTANCES=$1
TEST_ITERATION=$2
TEST_FILE=$3
LOG_TRACE='info'

if [[ ! -f .last_run ]];
then
    echo "Last run    : never"
else
    echo "Last run    : $(<.last_run)"
fi
echo "Current time: $(date '+%Y-%m-%d %H:%M:%S')"
echo

# # Delete log files from previous runs
rm -f log_*.log

# Check if any files have changed since the last run
if [[ ! -f .last_run || $(fd --type f --changed-within "$(<.last_run)" --exclude 'k6' | wc -l) -gt 0 ]]; then
    echo "✅ Files have changed since the last run, rebuilding k6-browser"
    echo "------------------------------------------------------"
    # Rebuild k6-browser if any files have changed
    xk6 build --with github.com/grafana/xk6-browser=.
fi

date '+%Y-%m-%d %H:%M:%S' > .last_run

echo
echo "------------------------------------------------------"
echo

run_test(){
    index=$1
    if XK6_BROWSER_LOG=$LOG_TRACE ./k6 run -q --vus 1 -i $TEST_ITERATION $TEST_FILE > log_$index.log 2>&1; then
        echo "✅ test run $index succeeded"
    else
        echo "❌ test run $index failed"
    fi
}

i=1
while [[ $i -le $NUM_K6_INSTANCES ]]
do
    run_test $i &
    ((i = i + 1))
done

wait
echo .
echo "All instances exited"

Here's an example command for testing:

$ ./multik6b.sh 10 2 script.js
Last run    : 2023-04-05 10:59:40
Current time: 2023-04-05 11:00:02

------------------------------------------------------

✅ test run 5 succeeded
✅ test run 6 succeeded
✅ test run 7 succeeded
✅ test run 2 succeeded
✅ test run 4 succeeded
✅ test run 3 succeeded
✅ test run 1 succeeded
✅ test run 10 succeeded
✅ test run 8 succeeded
✅ test run 9 succeeded

The PR also refactors the page attachment logic for better maintenance and readability—also fixes the linter warnings. This helped me find the problem since it made it easier to understand the code. I didn't prefer to put it in another PR but rather as a commit here. I believe "make it better than you found it" is a nice approach for reducing technical debt :)

[ankur22]

Great bit of debugging to catch this and resolve it! Nice work 👏

I'm wondering if you could help me a bit by splitting the last commit into two, one for the new changes to onAttachedToTarget and another for the splitting/refactoring the method into smaller methods?

[ankur22]

👏 Thanks for splitting the changes into smaller commits, it was a lot easier to follow. I only have some very minor suggestions.

LGTM 🎉

[ka3de]

Great work @inancgumus ! 🎉
LGTM.