[lustre-devel] Should we have fewer releases?
Christopher J. Morrone
morrone2 at llnl.gov
Thu Nov 5 13:45:02 PST 2015
Hi,
I think that Cory meant to send his message to this list. Please read
his comment at the end before reading my reply here.
Peter Jones is summarized in those notes as saying that how long
releases take seems to depend on how much change was introduce into the
tree. I agree; this is a causal relationship.
I believe that if our six months releases are often late and take in the
7-9 month range, then I think that planned nine month releases will in
actuality take 12+ months.
It may not be the current advocate's reason for suggesting the longer
release cycle, but one argument I have heard many times is that a longer
cycle will reduce the amount of manpower needed to create releases. I
don't think that is substantially true. While there are some fixed
costs in creating a release, there is no real reason that those fixed
costs need be a dominant factor for manpower demands. On the other
hand, required manpower is almost always going to be strongly
proportional to, and dominated by, the amount of change we introduce.
If we perform excellent, in-depth reviews on all code changes and we
also perform strong testing throughout the development cycle, then the
manpower centered around "release time" need not be very high. But
right now our peer reviews aren't quite as in depth as they could be,
and community testing, while improving of late, is unpredictably applied
and concentrated near the end of the cycle. This guarantees a large and
unpredictable amount of development effort shortly before the release
date, often resulting in a missed release target.
So lets think about what happens if we extend the development cycle,
including extending freeze dates. Assuming only minor, gradual
improvements in code reviews and continuous testing (a very safe
assumption, I think), the amount of change introduced into the release
will be proportionally higher the longer we leave the landing window
open. The greater the change, the larger the amount of effort needed to
stabilize the code after the fact.
Furthermore, I would speculate that extending the release cycle and
putting off the testing and stabilization effort will actually require a
super linear increase in the time for that effort.
Consider for instance that the longer we make the release cycle, the
more likely that bug authors have moved on to another task or project.
Since this is an open source project we don't have any way to order the
bug author back to work on her code. Even if the original author is
available to work on the bug, she may need significant time to shift
gears and remember how the code she touched works before she can make
significant progress. If the original author is not available, then
someone else needs to learn that portion of code and that has even more
obvious impact on time to solution and release.
I think there are also other effects that will conspire (e.g. unexpected
change interactions) to make the testing and stabilization period grow
super-linearly with the increase in the landing window.
Therefore, I would argue that lengthening the release cycle will neither
reduce our manpower needs nor result in more predictable release dates.
On the contrary, we need to go in the opposite direction to achieve
those goals. We need to shorten the release cycle and have more
frequent releases. I would recommend that we move to to a roughly three
month release cycle. Some of the benefits might be:
* Less change and accumulate before the release
* The penalty for missing a release landing window is reduced when
releases are more often
* Code reviewers have less pressure to land unfinished and/or
insufficiently reviewed and tested code when the penalty is reduced
* Less change means less to test and fix at release time
* Bug authors are more likely to still remember what they did and
participate in cleanup.
* Less time before bugs that slip through the cracks appear in a major
release
* Reduces developer frustration with long freeze windows
* Encourages developers to rally more frequently around the landing
windows instead of falling into a long period of silence and then trying
to shove a bunch of code in just before freeze. (They'll still try to
ram things in just before freeze, but with more frequent landing windows
the amount will be smaller and more manageable.)
It was also mentioned in the LWG email that vendors believe that the
open source releases need to adhere to an advertised schedule. Having
shorter release cycles with smaller and more manageable change will
directly contribute to Lustre releases happening on a more regular schedule.
Those same vendors tend to be concerned that they will not be able to
productise every single release if they happen on a three month
schedule. It is important to recognize that a vendor's product schedule
need not be directly in sync with every community release. It is
actually quite common in the open source world for vendors to select a
version to productise, and skip over some community releases to find the
next version which they will productise. Consider, for instance, the
Linux kernel. RedHat selects a version of the kernel to include in RHEL
and then sticks with the base of code fore many years. They will
backport changes as they see fit, but their base on that release remains
the same. The next kernel that they decide to package in their product
will skip over many of the upstream Linux releases.
Some Lustre vendors already operate this way, and the ones that do not
need to adapt to this common, successful open source model.
Shortening the release cycle will help encourage and sustain an active
open source community of Lustre developers from a diverse set of
organizations.
Conversely, lengthening the release cycle will result in less Lustre
stability and encourage stagnation. It will make us less nimble, less
likely to meet the needs of our current user base, and slower to expand
into new markets.
Lets start working through what process changes we will need to make to
shorten the development cycles and make lustre releases more often.
Thanks,
Chris
On 11/04/2015 01:16 PM, Cory Spitz wrote:
> Hello, Lustre developers.
>
> On today¹s OpenSFS LWG teleconference call (notes at
> http://wiki.opensfs.org/LWG_Minutes_2015-11-04) I proposed that we change
> the Lustre release cadence from six months to nine months. Chris M.
> responded (below) that any discussion about development changes should
> happen here on lustre-devel. I agree, developers need to be on-board.
>
> So what do you think about release changes? What requirements do you
> have? What issues would you have if OpenSFS changed the major release
> cadence to nine months?
>
> Thanks,
>
> -Cory
>
> On 11/4/15, 1:58 PM, "lwg on behalf of Christopher J. Morrone"
> <lwg-bounces at lists.opensfs.org on behalf of morrone2 at llnl.gov> wrote:
>
>> On 11/04/2015 10:28 AM, Cory Spitz wrote:
>>
>>> Lustre release cadence
>>> We haven¹t been good about hitting our 6 month schedules
>>> Cory proposed a 9 month cadence just to recognize reality. Certainly
>>> pros/cons to any scheme. Should be up for discussion. How/when to
>>> decide?
>>
>> Any development change like that needs to be discussed on lustre-devel.
>>
>> Chris
>>
>> _______________________________________________
>> lwg mailing list
>> lwg at lists.opensfs.org
>> http://lists.opensfs.org/listinfo.cgi/lwg-opensfs.org
>
> _______________________________________________
> lustre-devel mailing list
> lustre-devel at lists.opensfs.org
> http://lists.opensfs.org/listinfo.cgi/lustre-devel-opensfs.org
>
More information about the lustre-devel
mailing list