Marcin,

It just dawned on me that the main reason 008's start_grid takes so long is that the docker images are fresh pulled every time. Several hundred MB, every time (ugh, sorry). We can and should cache them. What do you think about trying this before doing anything else? [it would also be a good time to update from actinium to the latest, iron.]

@Barak Korren you once mentioned to me we should cache these if they are ok to cache (they are). How do we do that?

docker.io/selenium/node-chrome-debug   3.9.1-actinium      327adc897d23        13 months ago       904 MB
docker.io/selenium/node-firefox-debug   3.9.1-actinium      88649b420bd5        13 months ago       814 MB

Greg


On Tue, Mar 5, 2019 at 6:15 AM Greg Sheremeta <gshereme@redhat.com> wrote:

On Tue, Mar 5, 2019 at 4:55 AM Marcin Sobczyk <msobczyk@redhat.com> wrote:

Hi,

On 3/4/19 7:07 PM, Greg Sheremeta wrote:
Hi,

Thanks for trying to improve the tests!

I'm reluctant to give up Firefox sanity tests on every commit, though. In fact, I wanted to add Edge and Safari, because those are also supported browsers. Just today a Firefox only issue was reported, so they are valuable.

Was the Firefox-only issue detected by basic suite or some other tests?

It was reported by a developer. Because GWT compiles permutations per browser, and each browser therefore loads completely separate JavaScript payloads, it's just too easy for it to break in one browser and be fine in the other, so I'm really not ok to remove Firefox.

If Admin Portal was React where there is a single JavaScript payload that's shared among all browsers, then I'd consider it.


Did you consider either leaving a grid up permanently or perhaps using a third party like saucelabs?
I did consider simply having our own grid for the OST.
There's even a thread somewhere on ovirt-devel, where someone found OST trying to connect to one of my VMs in Tel Aviv, where my own grid was running :D
I couldn't make a public demo though - OST executors couldn't see my VM in tlv.

This approach has 2 big flaws:

  • it requires quite a lot of resources for the grid to always be there for us
What about Saucelabs or another third party free tool?
 
  • it makes OST running times somehow undeterministic - situations, where WebDriver has to wait for Selenium hub/nodes to be free, will probably take place
The way I see basic suite's UI sanity tests, is that they're exactly what they're called - sanity tests.
We do trivial checks like "can we log in to the webadmin site", "can we go to 'virtual machines' sub-page".
I'm not in favor of dropping these completely - I think they make sense, but I also think we can live with a trimmed-down version that saves a lot of time.
As I said - AFAIK QE have their own Selenium grid, where they run more complex tests on the UI.

Yes, OST basic_ui_sanity tests aren't "compatibility" tests. We're not checking pixels or look. They are super simple "does the app load" tests, are very valuable, and we're not dropping them.

Greg

Regards, Marcin



Best wishes, 
Greg 

On Mon, Mar 4, 2019, 11:39 AM Marcin Sobczyk <msobczyk@redhat.com> wrote:

Hi,

TL; DR Let's cut the running time of '008_basic_ui_sanity.py' by more than 3 minutes by sacrificing Firefox and Chrome screenshots in favor of Chromium.

During the OST hackathon in Brno this year, I saw an opportunity to optimize basic UI sanity tests from basic suite.
The way we currently run them, is by setting up a Selenium grid using 3 docker containers, with a dedicated network... that's insanity! (pun intended).
Let's a look at the running time of '008_basic_ui_sanity.py' scenario (https://jenkins.ovirt.org/view/oVirt%20system%20tests/job/ovirt-system-tests_manual/4197/):

01:31:50 @ Run test: 008_basic_ui_sanity.py:
01:31:50 nose.config: INFO: Ignoring files matching ['^\\.', '^_', '^setup\\.py$']
01:31:50   # init:
01:31:50   # init: Success (in 0:00:00)
01:31:50   # start_grid:
01:34:05   # start_grid: Success (in 0:02:15)
01:34:05   # initialize_chrome:
01:34:18   # initialize_chrome: Success (in 0:00:13)
01:34:18   # login:
01:34:27   # login: Success (in 0:00:08)
01:34:27   # left_nav:
01:34:45   # left_nav: Success (in 0:00:18)
01:34:45   # close_driver:
01:34:46   # close_driver: Success (in 0:00:00)
01:34:46   # initialize_firefox:
01:35:02   # initialize_firefox: Success (in 0:00:16)
01:35:02   # login:
01:35:11   # login: Success (in 0:00:08)
01:35:11   # left_nav:
01:35:29   # left_nav: Success (in 0:00:18)
01:35:29   # cleanup:
01:35:36   # cleanup: Success (in 0:00:06)
01:35:36   # Results located at /dev/shm/ost/deployment-basic-suite-master/008_basic_ui_sanity.py.junit.xml
01:35:36 @ Run test: 008_basic_ui_sanity.py: Success (in 0:03:45)

Starting the Selenium grid takes 2:15 out of 3:35 of total running time!

I've investigated a lot of approaches and came up with something like this:

  • install 'chromium-headless' package on engine VM
  • download 'chromedriver' and 'selenium hub' jar and deploy them in '/var/opt/' on engine's VM
  • run 'selenium.jar' on engine VM from '008_basic_ui_sanity.py' by using Lago's ssh
  • connect to the Selenium instance running on the engine in '008_basic_ui_sanity.py'
  • make screenshots
This series of patches represent the changes: https://gerrit.ovirt.org/#/q/topic:selenium-on-engine+(status:open+OR+status:merged).
This is the new running time (https://jenkins.ovirt.org/view/oVirt system tests/job/ovirt-system-tests_manual/4195/):

20:13:26 @ Run test: 008_basic_ui_sanity.py:
20:13:26 nose.config: INFO: Ignoring files matching ['^\\.', '^_', '^setup\\.py$']
20:13:26   # init:
20:13:26   # init: Success (in 0:00:00)
20:13:26   # make_screenshots:
20:13:27     * Retrying (Retry(total=2, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('<urllib3.connection.HTTPConnection object at 0x7fdb6004f8d0>: Failed to establish a new connection: [Errno 111] Connection refused',)': /wd/hub
20:13:27     * Retrying (Retry(total=1, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('<urllib3.connection.HTTPConnection object at 0x7fdb6004fa10>: Failed to establish a new connection: [Errno 111] Connection refused',)': /wd/hub
20:13:27     * Retrying (Retry(total=0, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('<urllib3.connection.HTTPConnection object at 0x7fdb6004fb50>: Failed to establish a new connection: [Errno 111] Connection refused',)': /wd/hub
20:13:28     * Redirecting http://192.168.201.4:4444/wd/hub -> http://192.168.201.4:4444/wd/hub/static/resource/hub.html
20:14:02   # make_screenshots: Success (in 0:00:35)
20:14:02   # Results located at /dev/shm/ost/deployment-basic-suite-master/008_basic_ui_sanity.py.junit.xml
20:14:02 @ Run test: 008_basic_ui_sanity.py: Success (in 0:00:35)

(The 'NewConnectionErrors' is waiting for Selenium hub to be up and running, I can silence these later).
And the screenshots are here: https://jenkins.ovirt.org/view/oVirt%20system%20tests/job/ovirt-system-tests_manual/4195/artifact/exported-artifacts/screenshots/

The pros:
  • we cut the running time by more than 3 minutes
The cons:
  • we don't get Firefox or Chrome screenshots - we get Chromium screenshots (although AFAIK, QE has much more Selenium tests which cover both Firefox and Chrome)
  • we polute the engine VM with 'chromium-headless' package and deps (in total: 'chromium-headless', 'chromium-common', 'flac-libs' and 'minizip'), although we can remove these after the tests

Some design choices explained:

Q: Why engine VM?

A: Because the engine VM already has 'X11' libs. We could install 'chromium-headless' (and even other browsers) on our Jenkins executors, but that would mess them up a lot.

Q: Why Chromium?

A: Because it has a separate 'headless' package.

Q: Why not use 'chromedriver' RPM in favor of https://chromedriver.storage.googleapis.com Chromedriver builds?

A: Because the RPM version pulls a lot of extra dependencies even on the engine VM ('gtk3', 'cairo' etc.). Builds from the URL are the offical Google Chromedriver builds, they contain a single binary, and they work for us.

What still needs to be polished with the patches:

  • Currently 'setup_engine_selenium.sh' script downloads each time 'selenium.jar' and 'chromedriver.zip' (even with these downloads we get much faster set-up times) - we should bake these into the engine VM image template.
  • 'selenium_hub_running' function in 'selenium_on_engine.py' is hackish - an ability to run an ssh command with a context manager (and auto-terminate on it exits) should be part of Lago. Can be refactored.

Questions, comments, reviews are welcome.

Regards, Marcin



_______________________________________________
Devel mailing list -- devel@ovirt.org
To unsubscribe send an email to devel-leave@ovirt.org
Privacy Statement: https://www.ovirt.org/site/privacy-policy/
oVirt Code of Conduct: https://www.ovirt.org/community/about/community-guidelines/
List Archives: https://lists.ovirt.org/archives/list/devel@ovirt.org/message/RLB2KSNJS4YKVMCDUUHOZJWBQDGJCXGZ/


--

GREG SHEREMETA

SENIOR SOFTWARE ENGINEER - TEAM LEAD - RHV UX

Red Hat NA

gshereme@redhat.com    IRC: gshereme



--

GREG SHEREMETA

SENIOR SOFTWARE ENGINEER - TEAM LEAD - RHV UX

Red Hat NA

gshereme@redhat.com    IRC: gshereme