Define behaviour of commands in terms of unexpected navigations, content process crashes etc.

[gsnedders]

Currently, various implementations have pretty useless behaviour in face of some errors (e.g., both GeckoDriver and ChromeDriver seem to return null for execute script if the content process crashes). We should better define what implementations should do, and probably define something distinguishable from potentially correct behaviour (given null from that and null from return null; are the same!).

There's a few cases here:

• An implementation like ChromeDriver where the WebDriver end point communicates over another IPC channel to the browser itself (over CDP in that case), where there are multiple possible failure modes:

  • the CDP connection being closed (and being unable to reconnect)
  • the browser stopping responding to CDP requests (and being unable to reconnect and remake the request, if it's idempotent!)
  • the browser responding to CDP requests but the content process not responding to the browser ("Aw, snap!")

• An implementation where the browser is itself the end point (I'm not sure anyone does this any more?):

  • the end point stops responding because the browser itself has crashed
  • the content process isn't responding

• An intermediary node losing its connection to the end point.

These are all general connectivity problems where the local end can still communicate with the first remote end, so we should define something here, as we still have a working WebDriver connection.

[AutomatedTester]

So there are a few issues we need to be aware of, which can be overcome by the remote end.

1. What happens if the "comms" thread crashes but the browser is still alive?
2. What happens when the main thread dies taking out the entire browser?

For the most part this leads to us making sure the process the end point is managing is alive or dead and giving a meaningful message. Ideally we don't want to define process management because that feels wrong (though I can't articulate why clearly so can be swayed).

This work would also probably need to be done for each of the commands as a crash can happen at any point.

Is this something that we could leave undefined for webdriver but focus on the work in WPT et al. for now and see if there is commonality after that work?

[gsnedders]

Ideally we don't want to define process management because that feels wrong (though I can't articulate why clearly so can be swayed).

Strongly agree. I don't think we want to tie anything to actual processes or anything. The failure modes I think we really want to expose, at most, are:

• This window is unresponsive/not alive.
• This session is unresponsive/not alive.

I think the specification should be clear about behaviour in all of the cases in the original post, but probably not by actually exhaustively listing all of them.

[jugglinmike]

Geckodriver 0.24.0 now returns "unknown error" if the crash occurs during
Execute Async Script. Chromedriver 73.0.3683.68 still returns "success" with
data null. I'll include a Bash script below in case that's helpful for
researching this.

Is this something that we could leave undefined for webdriver but focus on
the work in WPT et al. for now and see if there is commonality after that
work?

Here's a workaround for one specific case of this in WPT's tooling:

web-platform-tests/wpt#16157

`webdriver-crash.sh`

#!/bin/bash

set -e

port=2323
webdriver_bin=$1

if [ ! -x "${webdriver_bin}" ]; then
  echo Usage: $0 PATH_TO_WEBDRIVER_EXECUTABLE >&2
  exit 1
fi

$webdriver_bin --port=${port} > /dev/null &

webdriver_pid=$!

function cleanup {
  if [ -n "${session_id}" ]; then
    curl \
      -X DELETE \
      --silent \
      http://localhost:${port}/session/${session_id} > /dev/null
  fi

  if [ -n "${webdriver_pid}" ]; then
    kill -9 ${webdriver_pid}
  fi
}

trap cleanup EXIT

sleep 1

session_id=$(
  curl \
    --data '{"desiredCapabilities":{}}' \
    --silent \
    http://localhost:${port}/session | \
  python -m json.tool | \
  grep sessionId | \
  sed -e 's/^.*://' -e 's/.*"\([^"]\+\)".*/\1/'
)

curl \
  --verbose \
  --data '{"script": "", "args": []}' \
  http://localhost:${port}/session/${session_id}/execute/async &

browser_pid=$(ps aux | \
  grep -E 'chrom.*enable-automatio[n]|firefox.*marionett[e]' | \
  grep -v renderer | \
  awk '{print $2}')

echo Browser process running with identifier $browser_pid.

if kill -9 $browser_pid; then
  echo Process successfully terminated.
else
  echo Unable to terminate process.
fi