LKML Archive on lore.kernel.org
 help / color / Atom feed
From: Guillaume Tucker <guillaume.tucker@collabora.com>
To: Thorsten Leemhuis <linux@leemhuis.info>
Cc: "Rafael J. Wysocki" <rafael@kernel.org>,
	regressions@lists.linux.dev,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	Konstantin Ryabitsev <konstantin@linuxfoundation.org>,
	Pablo Neira Ayuso <pablo@netfilter.org>,
	ksummit@lists.linux.dev, workflows@vger.kernel.org,
	Greg KH <gregkh@linuxfoundation.org>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	"kernelci@groups.io" <kernelci@groups.io>,
	automated-testing@lists.yoctoproject.org
Subject: Re: RFC: building a regression tracking bot for Linux kernel development
Date: Fri, 23 Apr 2021 21:45:08 +0100
Message-ID: <de923dec-b0e8-788d-f73f-ee7a1c6af47b@collabora.com> (raw)
In-Reply-To: <268a3049-7c0b-8a33-1ff6-5a2d35fcba16@leemhuis.info>

+kernelci +automated-testing

Hi Thorsten,

On 22/04/2021 08:16, Thorsten Leemhuis wrote:
> Lo! As mentioned a few times recently I'm staring to build a bot for
> semi-automatic Linux kernel regressions tracking. Find below a rough
> description of how I imagine it's going to work. That way I want to give
> everyone a chance to influence things before I'm starting to code for
> real. Early feedback will help to build something that's acceptable for
> the Linux kernel developer community and used in practice in the long
> run, and that's what I aim for.
> 
> I know, I know, "Talk is cheap. Show me the code.". But I had to think
> things through and write some of it down anyway, so no harm done in
> posting it as RFC. I CCed ksummit, as many maintainers hang out there
> and because this is a follow up to my former regression tracking work we
> discussed on both kernel and maintainers summit 2017; it fact it
> hopefully might be something for this year as well, we'll see, too early
> to tell.

This sounds great, with a simple email-based interface and a well
defined scope for the bot's role.

There are a few things that are worth mentioning from a KernelCI
point of view though, to ensure both tools work together and not
against each other (see https://kernelci.org).

The first thing is that KernelCI detects and tracks a fair amount
of regressions all the time, then runs automated bisections to
find the breaking commits.  This currently leads to between 1 and
10 unique "bug reports" per week.  So including "regzbot" in
those reports would initially seem simple and very effective.

Then another aspect to take into account is the proliferation of
tools.  KernelCI's mission is not only to run tests but also to
gather results from other test systems into a common database
called KCIDB.  The main goal is to provide a full picture of all
the issues in one place, with unified email reports and a web
dashboard.

Tracking regressions is on the roadmap for KCIDB, although it's
not yet entirely decided how it will actually work.  Ideally it
would simply let systems submit their own regression data to
KCIDB, which sounds very similar to what regzbot would be doing.

I can think of several ways to orchestrate these things together.
In a nutshell, this is what I believe to be the best way around:

* automated test systems submit regressions to KCIDB, just like
  some of them already do with build and test data (syzbot, Red
  Hat's CKI, linux.kernelci.org, tuxsuite...)

* regzbot provides a way for reporting regressions by hand via
  email, and forwards them automatically to KCIDB too

Essentially, regzbot would remain autonomous but also act as an
email-based interface to submit data to KCIDB.  This gives you a
web dashboard and a common place to gather all regressions "for
free" among other things on kcidb.kernelci.org (still early
days).  You may also generate some regzbot dedicated web pages
elsewhere if needed in practice, both could co-exist.

How does that sound?

Another hypothetical scenario would be if regzbot was the
unifying tool, and all automated test systems sent their
regression data to it.  But then, KCIDB would become either
redundant or starved of the regression data it needs to
complement its test results.  So this doesn't seem so good.


I know it's important to have tools that work "by hand" for
developers, but automated testing leads to a better life!

Best wishes,
Guillaume

> So how will the "regzbot" work? The ideal case is simple:
> 
> Someone reports a regression to the recently created regressions mailing
> list(regressions@lists.linux.dev). There the user includes a tag like this:
>> #regzb introduced: 94a632d91ad1 ("usc: xhbi-foo: check bar_params earlier")
> 
> That will make regzbot add the report to its list of regressions it
> tracks, which among other will make it store the mail's message-id
> (let's assume it's `xt6uzpqtaqru6pmh@earth.solsystem`). Ideally some
> developer within a few days will fix the regression with a patch. When
> doing so, they already often include a tag linking to the report:
>> Link: https://lore.kernel.org/r/xt6uzpqtaqru6pmh@earth.solsystem
> 
> 
> Regzbot will notice this tag refers to the regression it tracks and
> automatically close the entry for it.
> 
> That's it already. The regression was tracked with:
> 
>  * minimal overhead for the reporter
>  * no additional overhead for the developers – only something they ought
> to do already became more important
> 
> Invisible ideally
> -----------------
> 
> In the ideal case regzbot thus seems to be of no use. But obviously
> things will be anything else than ideal quite often – for example when
> nobody fixes the reported regression.
> 
> The webpages that Regzbot will generate (see below) will show this. They
> among others are meant for Linus or Greg to check how things stand, so
> they can simply fix a regression by reverting the causing commit if they
> want to; in other situations they might decide to delay a release to get
> crucial regressions solved.
> 
> And that's what regression tracking is about: providing a view into the
> state of things with regards to regressions, as that's the important
> thing missing in Linux kernel development right now.
> 
> 
> That can't be all
> -----------------
> 
> Of course the world is more complicated than the simple example scenario
> above, as the devil is always in the details. The three most obvious
> problems the initial ideal scenario left aside:
> 
> * The reporter doesn't specify the #regzb tag at all. Regzbot can't do
> anything about it, it sadly won't have visionary power and a AI engine
> any time soon. Some human (for a while that often will be me) thus needs
> to reply with the tag with a proper reply-to to the report to make
> regboz track it.
> 
> * The commit causing the regression is unknown to the reporter. In that
> case the tag should mention the span when the regression was introduced:
>> #regzb introduced: v5.7..v5.8-rc1
> 
> * The developer who fixes the issue forgets to place the "Link:" tag,
> which can't be added once committed. In that case some human needs to
> reply to the thread with the initial report with a tag like this:
>> #regzb Fixed-by: c39667ddcfd5 
> 
> 
> How will it look
> ----------------
> 
> Here is a mockup on the website for the regzbotproject:
> https://linux-regtracking.leemhuis.info/images/regzbot-mockup.png
> 
> You'll notice a few things:
> 
>  * regressions for mainline kernel will be shown on a different page
> than those in stable and longterm kernels, as they are handled by
> different people.
> 
>  * regressions where the culprit is known get the top spot, as the
> change causing them can sometimes simply be reverted to fix the regression.
> 
>  * the second spot is for regressions in the current cycle, as contrary
> to those in previous release there is still time to fix those before the
> next release.
> 
>  * Regzbot will try to monitor the process between reporting and fixing
> and provide links to lookup details. Regzbot will thus watch the thread
> where the regression was reported and show when it noticed the last
> activity; it will also look out for `#regszb Link:` and `Link:` tags in
> patch submissions and linux-next. That way release managers can
> immediately see if things stalled after the regression was reported; it
> also allows them to see if developers are working on a fix and how far
> it got in the machinery. If the causing commit is known, the webview
> obviously will link to it as well.
> 
>  * regressions where nothing happened for a while will be moved to the
> "dormant" page, to prevent the status page from getting filled by
> reports that obviously nobody cares about anymore. Reporters will be
> told about this by mail to give them a chance to provide a fresh status
> update to get things rolling again.
> 
> 
> Even more problems in the details
> ---------------------------------
> 
> Regzbot on purpose will lack many features found in traditional bug
> trackers: it's meant to be a simple tool acting in the background
> without much overhead, as it doesn't want to become yet another bug
> tracker. Nevertheless, it will need a few features they typically offer.
> Those will be usable via tags that need to be dropped into mails send in
> direct or indirect reply to the mail with the report:
> 
> * Mark a report as a duplicate of another or revert such a marking:
>> #regzb dup: https://lore.kernel.org/r/yt6uzpqtaqru6pmh@mars.solsystem
> 
>> #regzb undup
> 
> * Mark a report as invalid.
>> #regzb invalid: Turned out it never worked
> 
> 
> * generate a new title
>> #regzb new-title: Insert better description of the regression
> 
> 
> * the initially mentioned tag can be used in replies to the report to
> specify the commit causing the regression:
>> #regzb introduced: v5.7..v5.8-rc1
> 
> 
> * Tell regzbot that a discussion is related to a tracked regression:
>> #regszb Link: https://lore.kernel.org/r/yt6uzpqtaqru6pmh@mars.solsystem
> 
>   In the long run this is supposed to work in both directions, so you
> can use it in a thread started by a regression report to link to some
> other discussion or vice versa.
> 
> 
> Implications and hidden aspects
> -------------------------------
> 
> There are a few things of note:
> 
>  * The plan for now is to not have a tag like `#regzb unfix`: in case it
> turns out a commit did not fix a regression it's likely better to start
> with a fresh report anyway. That forces someone to explain the current
> state of things including the history clearly and straight forward; that
> makes things a lot easier to follow for others in these situations and
> thus is a good thing.
> 
>  * regzbot works without a public unique-id, as it uses the URL of the
> report instead and keeps any eye on is using the mail's message-id (say
> 20210406135151.xt6uzpqtaqru6pmh@earth.solsystem).
> 
>  * regzbot won't be able to handle regressions reported to a mailing
> list thread that is already tracked by regzbot, as it will assume all
> mails in a thread are related to the earlier report. In that case the
> reporter must be asked to start a new mailing list thread for the second
> regression. But that's quite normal, as a similar approach is needed
> when somebody reports an issue deep in a bug tracker ticket that was
> crated for a totally different issue.
> 
>  * Initially it won't be possible to track reports that are filed in bug
> trackers; but this use-case will be kept in mind during the design to
> make sure such a functionality can be added later easily.
> 
>  * developer when fixing a regression with a bisected "#regzb
> introduced:" tag can simply do `s/#regzb introduced:/Fixes:/` to get a
> tag they are supposed to add.
> 
>  * regression in stable and longterm kernels sometimes affect multiple
> versions, for example if a patch that works fine in mainline was
> backported to the longterm kernel 5.10 and 5.4 – but causes problems in
> both, as something required by the patch is missing in those lines. How
> this will be solved exactly remains to be seen, maybe like this:
>> #regzb Introduced: c39667ddcfd6 e39667ddcfd1 ("usc: xhbi-foo: check bar_params a little later again")
> 
>  Then regzbot can look those commits up and from that determine the
> affected versions. Obviously the reporter will likely not be aware of
> it, hence it's likely that the stable maintainer or the developer need
> to send a mail to make regzbot aware that this regression affects
> multiple versions.
> 
>  * Regzbot will need to be able to work with mails where mailers placed
> a linebreak into the text that follows the #regzb tag. This will be
> tricky, but is doable.
> 
>  * to keep things simple there are neither authentication nor
> restrictions for now, so anyone could mess things up by sending mails to
> an open list and using those tags. If that against expectations turns
> out to become a problem some restrictions will need to be put in place,
> for example to allow changes only from email addresses that (1) are on
> an allow list, (2) participated in the discussion or (3) have commits in
> the kernel. People could still forge complete mails including "From",
> but that's quite some work for not much to gain (except for messing
> regression tracking up).
> 
> 
> Implementation
> --------------
> 
> The rough initial idea had been to reuse parts of the syzbot golang
> source code, which already has an email interface similar to the one
> regzbot needs. But the closer I looked, the more I came to the
> conclusion that writing something in python is easier and better (even
> if that means I need to bring my really rusty python skills up to
> speed). That also has the benefit that python afaics is preferred by the
> kernel.org admins, which would make it more attractive for them to host
> the bot later.
> 
> The focus will be to properly establishing regression tracking with
> regszbot first. All features not strictly needed will thus be left out
> first to focus on what's most important. I'll also provide documentation
> and will use the bot myself to track regressions as I did a few years
> ago. Just like any other tracking solution it will always need some
> hand-holding...
> 
> = EOF =
> 
> That's it. FWIW, this mail is slightly modified version of a text I
> posted on the website for the regzbot project:
> https://linux-regtracking.leemhuis.info/post/regzbot-approach/
> 
> Side note: that project and my work is funded by NGI pointer for one
> year (see the website's about page for details). Follow-up funding won't
> be possible from there, but hopefully by then I can find some other way
> to keep things running and me in a position to look after regression
> tracking.
> 
> Ciao, Thorsten
> 


  parent reply index

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-04-22  7:16 Thorsten Leemhuis
2021-04-22 14:51 ` Mark Brown
2021-04-23  6:34   ` Thorsten Leemhuis
2021-04-23 10:11 ` Greg KH
2021-04-23 10:49   ` Thorsten Leemhuis
2021-04-23 11:01     ` Takashi Iwai
2021-04-23 11:11       ` Thorsten Leemhuis
2021-04-23 20:45 ` Guillaume Tucker [this message]
2021-04-24  9:43   ` Thorsten Leemhuis

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=de923dec-b0e8-788d-f73f-ee7a1c6af47b@collabora.com \
    --to=guillaume.tucker@collabora.com \
    --cc=automated-testing@lists.yoctoproject.org \
    --cc=gregkh@linuxfoundation.org \
    --cc=kernelci@groups.io \
    --cc=konstantin@linuxfoundation.org \
    --cc=ksummit@lists.linux.dev \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux@leemhuis.info \
    --cc=pablo@netfilter.org \
    --cc=rafael@kernel.org \
    --cc=regressions@lists.linux.dev \
    --cc=torvalds@linux-foundation.org \
    --cc=workflows@vger.kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

LKML Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/lkml/0 lkml/git/0.git
	git clone --mirror https://lore.kernel.org/lkml/1 lkml/git/1.git
	git clone --mirror https://lore.kernel.org/lkml/2 lkml/git/2.git
	git clone --mirror https://lore.kernel.org/lkml/3 lkml/git/3.git
	git clone --mirror https://lore.kernel.org/lkml/4 lkml/git/4.git
	git clone --mirror https://lore.kernel.org/lkml/5 lkml/git/5.git
	git clone --mirror https://lore.kernel.org/lkml/6 lkml/git/6.git
	git clone --mirror https://lore.kernel.org/lkml/7 lkml/git/7.git
	git clone --mirror https://lore.kernel.org/lkml/8 lkml/git/8.git
	git clone --mirror https://lore.kernel.org/lkml/9 lkml/git/9.git
	git clone --mirror https://lore.kernel.org/lkml/10 lkml/git/10.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 lkml lkml/ https://lore.kernel.org/lkml \
		linux-kernel@vger.kernel.org
	public-inbox-index lkml

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-kernel


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git