Auto-tools/Projects/Stockwell/disable-recommended

From MozillaWiki
< Auto-tools‎ | Projects‎ | Stockwell
Revision as of 16:30, 9 February 2018 by Jmaher (talk | contribs) (→‎Requesting review: - minor updates)
Jump to navigation Jump to search

This page is under construction!

[stockwell disable-recommended]

When an intermittent-failure bug has recorded 150 or more failures in the last 21 days, the Orange Factor robot will change the bug's stockwell whiteboard tag to "[stockwell disable-recommended]". These bugs should be reviewed at least twice per week (current schedule is Monday and Thursday).

This page describes how to review [stockwell disable-recommended] bugs.

Finding bugs to disable

  • Use bugzilla to find all bugs with whiteboard tag "[stockwell disable-recommended]".
  • For each disable-recommended bug, determine if the bug can be addressed by disabling tests.
    • Ignore "meta" bugs or any bugs not related to a particular test, some examples:
    • Eliminate bugs that have been recently fixed.
      • Check 7-day history on OrangeFactor.
      • bug 1420472
      • bug 1421775
      • bug 1423386 (fixed 2 days ago- recent comment and proved on graph)
      • If there are no recent failures, perhaps the bug can be marked [stockwell fixed], [stockwell disabled] or [stockwell unknown].
    • Eliminate bugs with a patch under review, or coming very soon.
      • Review recent bug history.
      • If there is a patch under review or comments indicate that a fix is coming soon, consider postponing action on this bug.

Finding the manifest

For each test to be disabled, find the associated test manifest:

  • Search for the test name in searchfox or dxr, or,

Searchfox manifest.jpg

  • Run 'mach test-info --show-info <test-name>' in a mozilla-central check-out.
$ ./mach test-info --show-info browser/base/content/test/general/browser_ctrlTab.js
===== browser/base/content/test/general/browser_ctrlTab.js =====
Found browser\base\content\test\general\browser_ctrlTab.js in source control.
Build configuration changed. Regenerating backend.
browser\base\content\test\general\browser_ctrlTab.js found in manifest browser/base/content/test/general/browser.ini
  flavor: browser-chrome

In the above cases you would want to edit browser/base/content/test/general/browser.ini

$ pwd
$ /home/mozilla/mozilla-central
$ nano browser/base/content/test/general/browser.ini

Determining which platforms are affected

On OrangeFactor, check the detail view for the bug for the last 30 days. Look down the list of bugs for affected Platforms and Build Types. Does this test need to be disabled on all platforms, or only on some? For opt or debug builds, or both?

There are many reasons why a test might only fail on certain configs:

  • we don't run that test on a specific platform (for example devtools on android)
  • the test is skipped already on a configuration or platform

We typically have 4 platforms:

  • Linux (opt|debug|pgo|asan|ccov) - also bits (32|64)
  • OSX - (opt|debug) - only run on 10.10
  • Windows - (opt|debug|pgo|ccov) - also Windows 7|10
  • Android - (opt|debug) - x86|arm
    • no browser-chrome, devtools, web-platform-tests, and a few others

If the test is failing at least 5 times in the last 7 days on any given config, lets skip it. We cannot skip on pgo specifically, so that is opt (!debug). Often I try to make the skip syntax as simple as possible for simplicity sake and future editing. If we have:

[browser_sanity.js]
skip-if = (os == 'linux' && debug && bits == 64) || (os == 'linux' && !debug && bits == 64) || (os == 'osx' && !debug)

I would prefer to see:

[browser_sanity.js]
skip-if = (os == 'linux' && bits == 64) || (os == 'osx' && !debug)

If the test already had a skip if:

[browser_sanity.js]
skip-if = (os == 'win') # Bug 3141592 - timeout

we would end up with:

[browser_sanity.js]
skip-if = true # Bug 3141592, 271289

The reason for this is we would be skipping on linux/windows/osx, but not linux32 or android. Since this test doesn't run on Android, we would only be skipping on linux32- and a test that only runs on 1 platform doesn't provide a lot of value unless it is platform specific.

Updating a manifest to disable a test

There are 3 different types of manifest files, each with a different format:

  • "ini" manifests like mochitest.ini, chrome.ini, browser.ini, xpcshell.ini
  • reftest or "list" manifests like reftest.list, crashtest.list
  • web-platform manifests in testing/web-platform/meta

ini manifests

  • In this type of manifest, add a "skip-if" statement under the test name -- on the next line.
[some_test.html]
skip-if = ...
  • Some examples of "skip-if":
skip-if = true                             # skip this test everywhere / always (test never runs)
skip-if = os == "android"                  # skip on Android only (continues to run on Linux, Windows, etc.)
                                           # other os strings: "win", "linux", "mac"
skip-if = debug                            # skip on all debug builds
skip-if = !debug                           # skip on all non-debug builds (opt, pgo)
skip-if = os == "android" || os == "linux" # skip on Android and skip on Linux
skip-if = os == "android" && debug         # skip on Android/Debug only (continues to run on Android/Opt)
  • For any test, there can only be one "skip-if" line. Often it is necessary to use || to expand the scope of a skip-if. For example, if a manifest already has:
[some_test.html]
skip-if = os == "android" || debug

but some_test.html is still failing on Windows/opt, it can be changed to:

[some_test.html]
skip-if = os == "android" || debug || os == "win"

reftest manifests

 skip-if(...) == some_test.html some_ref.html
  • Some examples of "skip-if":
skip-if(gtkWidget)                  # skip on Linux
skip-if(cocoaWidget)                # skip on Mac
skip-if(winWidget)                  # skip on Windows
skip-if(Android)                    # skip on Android
skip-if(isDebugBuild)               # skip on all Debug builds
skip-if(!isDebugBuild)              # skip on all non-Debug builds
skip-if(Android&&isDebugBuild)      # skip on Android/Debug

web-platform test manifests

(we haven't talked about these yet)

Making a patch

To actually disable a test, we need to update the manifest file in mozilla-central. Before doing that, create a mercurial patch containing the change, and upload the patch to bugzilla for review.

Some useful mercurial commands for creating a patch with "mq":

hg qseries       # see what patches are in your queue / what is applied
hg qdiff         # see the file changes in the current patch
hg qnew          # create a new patch
hg qrefresh      # update your patch with local file changes
hg qrefresh -e   # edit the commit message associated with your patch


I typically follow this workflow:

hg qnew bug1234567.patch
<edit> <manifest>.ini
hg qrefresh
hg qdiff # self review
hg qrefresh -e
<inside editor add comment |Bug 1234567 - disable <test> for frequent failures. r=gbrown| and save>
cp .hg/patches/bug1234567.patch ~/
hg qpop

NOTE: the self review is a great time to make sure:

  • there are not extra spaces on a blank line or at the end of a line, if there are they will be highlighted in red
  • you are only editing one line in a file, not multiple lines on accident
  • a chance to review your syntax and any comments

NOTE: the comment for the commit needs to follow the general pattern above.

Requesting review

In bugzilla, in the bug with [stockwell disable-recommended], use "Attach File" to attach your patch. Browse to find the patch in <your repo directory>/.hg/patches. Make sure you select "patch" for the Content Type. Then request review, "r?", and select a reviewer. In most cases, ask :gbrown or :jmaher for review.

In the above section it was suggested to copy your patch to ~/ via

cp .hg/patches/bug1234567.patch ~/

In doing this, you could easily find your patch when in bugzilla and looking to attach a file since it will be ~/bug1234567.patch.

A few extra notes here:

  • when uploading a patch the checkbox for "assign this bug to myself" is checked by default, uncheck this as you are just disabling a test
  • use a description, not the patch name. Something like: |disable test on <platforms>|


There are additional, generic instructions for getting Mozilla reviews at:

https://developer.mozilla.org/en-US/docs/Mozilla/Developer_guide/How_to_Submit_a_Patch

After the test is disabled

  • Check OrangeFactor to see if the patch was effective!

Notes from email

Before disabling or when you don't disable:

  • you have to read the history and see if anyone is working on it or trying to figure out the bug
* review recent bug history
** if there is a patch under review or comments indicate that a fix

is coming soon, consider postponing disabling

* look at the last 7 days in orange factor
** if no longer failing, mark [stockwell fixed]

if there are tests to disable we need to edit the manifests:

  • What type of test is this and where is the manifest? ./mach test-info
  • Which platforms are affected? opt/debug/both, os, other
  • One time mercurial setup:

https://developer.mozilla.org/en-US/docs/Mozilla/Developer_guide/Source_Code/Mercurial

  • Creating a patch and requesting review

We'll need separate guides to disabling syntax for:

* manifest parser
** mochitest.ini, browser.ini, chrome.ini, xpcshell.ini, a11y.ini
* reftest manifests
** reftest.list, crashtests.list
** use fuzzing if possible
* web-platform tests
** testing/web-platform
  • autophone (easily confused with above .ini/.list manifests)