12 Jun
2012
12 Jun
'12
8:21 a.m.
On Jun 9, 2012, at 9:37 PM, Caleb James DeLisle wrote:
Yes, I understand, I had the same issue. However:
1) it's not like we have all the time in the world. This is a balance between
testing/quality and developing new features/fixing bugs. So if you push one to take the
other one.
2) we only include blocker bugs in RC and all devs have their eyes on commits (ie they do
reviews) in RC phase
3) Although it can happen, it didn't happen very often.
4) It's not because we test that we find bugs
5) I remember one instance when our final had a big issue in a given domain. So we
released a point release a few days after (2-3 days after).
All in all since it doesn't happen often and since we can release a point release
quickly I think we have a relatively good balance.
Sure, I'm all for doing experiments. That would be:
* 2 weeks milestone 1 (note that the point of M1 is to have a milestone to check our
progress)
* 2 weeks milestone 2
* 1 week RC1
* 1 week Final
Now we need to adapt the workload (ie roadmap ) accordingly of course.
Well, I'd like to rephrase what I said because I know from experience that it leads to
problems. When developers code they're always 80% finished and the hard part if
finishing the remaining 20%. If you don't have a deadline you'll find yourself
tuning and improving stuff close to infinitely and it'll take whatever time you have
available. So we do need to put some deadlines.
What we need to do is split large features into smaller ones and assign them to releases.
We still need that plan.
Thanks
-Vincent
On 06/09/2012 03:39 AM, Vincent Massol wrote:
I have checked and there are quite a few.
Here are some not reported by us in relatively recent releases:
- http://jira.xwiki.org/jira/browse/XWIKI-7035
- http://jira.xwiki.org/jira/browse/XWIKI-7645
- http://jira.xwiki.org/jira/browse/XWIKI-7645
- http://jira.xwiki.org/jira/browse/XWIKI-7753
- http://jira.xwiki.org/jira/browse/XWIKI-7773
- http://jira.xwiki.org/jira/browse/XWIKI-7774
- http://jira.xwiki.org/jira/browse/XWIKI-7763
On Jun 9, 2012, at 6:28 AM, Sergiu Dumitriu wrote:
I've heard you mention this but I've not noticed any example of a user finding an
issue in an RC and
reporting it, causing it to be fixed for the final. I've only been around for 2 years
though. Do you
have such an example? On 06/08/2012 05:07 AM, Caleb James DeLisle
wrote:
I have one big problem with this strategy that I've already explained several past
emails.
The current way of doing staging with RC is very powerful because it's a real release
and as such is advertised everywhere: on xwiki.org in download, in blog post, on twitter,
on wikipedia, freshmeat, etc. Which means we have a lot of users who can download it and
try it out. This will not happen with staging. It won't be advertised significantly
(it cannot since it's not a real release) thus leading to a less good solution than
what we have.
It seems that if we want to shorten a release cycle, we have 2 options:
#1 No RC release
#2 No staging
I think it would be a shame to scrap staging over this, especially since
my understanding is we want to move toward frequent releases with no
release candidates or milestones so staging would be a requirement.
This is an interesting topic which can be discussed.
Right now for 4.1 I think we need a coherent proposal rather than an adhoc chop-and-slice
of the agreed upon schedule and dates.
Finally, I'm concerned about changing our release process in the middle
of a release which is behind it's normal schedule. I'm not completely
opposed to a change but IMO if we want to change it we need to proceed with extra
caution.
The purpose of the RC releases was as a kind of staging, without actually using real
staging. So I'd say that for the next releases we should only have milestones, and
staged final releases.
However, for 4.1 we should keep the rc-1 release in the plan, since our staging process
is new and not thoroughly tested yet.
The way I simulated "staging releases" when I had the RM hat was to leave a
window between uploading the artifacts in our maven repository and pushing them further,
and in this window I did quick smoke tests to see that everything works (with the help of
Sorin and the rest of the QA team at XWiki SAS).
So, I propose this as the strategy for our next releases:
- Each release goes to a staging repository first
- Milestone releases spend a little time in staging, enough for a few quick smoke tests;
depending on the time of day, this could last from an hour to at most a day
- There are no RC releases, there are only RC staging repositories. Currently the staging
repositories have random names, but the RM should rename each repository with the proper
RC name: 4.2-rc-1, 4.2-rc-2 if the first build was busted, and so on. The version written
inside the packages is the final one, like 4.2, without a rc-x suffix. While in staging, a
RC release goes through more intensive tests, including smoke testing, the full MTR, and
developers that contributed to the release should test their committed features. Maybe
even a Jenkins job could be executed for that release, with all the automated tests.
Timing? Since we don't know how many times we have to stage releases, I'd say
that we can start building RCs one week before the planned final release date, and if bugs
are found, a new RC is build the next day (to let issues accumulate), hoping that there
will be a stable successful build before the final release date, without a 72h mandatory
waiting period. Thus the only
valid solution for me to reduce release cycles is to do 2 things at the same time:
- work on the quality of our releases (i.e. increase our automated test suite and test
coverage)
This is good no matter how we release.
- do short release cycle so that if a user finds
a problem in say version 4.45 then we release 4.46 the week after or 2 days after if
it's critical (this is basically what Jenkins is doing)
Right now our cycle is 3 months for each release which means that if some dev cannot make
the date for ex he has to wait 3 months for his stuff to be available to end users, which
means devs are pushing stuff even if they are half baked (missing tests for example).
Thus, with the current situation, I'd very much prefer to not do staging (in the
sense of Nexus staging) and instead continue to use milestones and RCs. Now we can still
do staging as a quick check for the RM before pushing to the main repo but it
shouldn't last more than 1 day, as otherwise it adds up with the RC time and thus
leads to even longer release cycles.
This would make sense if "final" was a clone of rc-X.
What we do now is build an RC, fix whatever issues we have with it then build
"final" and release it without manual testing.
I do not like releasing final without testing, especially since we use the name rc-1,
leading users to believe that we have a candidate is tested then selected for release.
<brainstorming>
Moving forward:
what we could do would be to reduce the whole cycle from 3 months to 3 weeks:
* 1 week milestone 1
* 1 week milestone 2
* 1 week final (we need less time for stabilization over what we have now since
there's less work done in m1/m2 than when they last 4 weeks each)
I like the general thinking but from 12 weeks to 3 weeks is a huge step, I think we
should at least try 6 week cycle and then do post-release brainstorming to see what issues
come up with shortened release times. If a developer
doesn't have time to push his stuff in m1 or m2 then he'll have to wait only 1
week.
Of course we should have Roadmap that last several releases so that we don't do
roadmap preparation all the time and so that we can put large features in the roadmap that
span more than 1 release. Actually the roadmap would list priority stuff we want to work
on but will not say in what exact version we'll have it.
This is good, currently we encourage developers to delay the release. The advantage
of this is we'll get even more testing/feedback from community since we'll have
real releases and users will get features/improvements/bugfixes pushed to them faster. The
challenge is to keep good quality but I think we can do that and in any case there's
only one way to know if it works: to try it ;)
If we do manual testing on final before publicizing it then we're already ahead of
the current process. Thanks,
Caleb
>
> I think that without Sergiu's automated script for releases doing a release was
too costly to do this. This means doing 3 releases in 3 weeks (ie. one per week). Is the
automated script good enough to allow for this.
>
> Also we're going to have automated upgrades very soon now so it seems like
upgrading versions is going to be less an issue for users (note that they don't have
to upgrade at each version and they can skip releases).
>
> WDYT? Do you think the pace is sustainable for developers and that we won't
spend too much time in "bureaucracies"?
>
> IMO this would help us automate our release process even further so that it becomes a
nobrainer and will help us improve ourselves wrt quality/automated tests even further.
>
> There's one problem maybe with APIs. They need time to mature but with the
young-api strategy that could be ok.
> <brainstorming>
>
> Thanks
> -Vincent
>
>>> Thanks,
>>>
>>> Caleb
>>>
>>>
>>> On 06/08/2012 03:15 AM, Vincent Massol wrote:
>>>> Hi guys,
>>>>
>>>> I'd like to bring an issue with this VOTE below.
>>>>
>>>> When I initially read it I didn't realize that this was about doing
double-staging:
>>>> * once with nexus staging
>>>> * another one with the RC release
>>>>
>>>> So it increases the time we spend for doing releases instead of reducing
it which is the direction we would like to go.
>>>>
>>>> The increase is bad because we're already spending too much time just
on the release itself while we should reduce it to a minimum so that we can focus on
developing new features/improvements/fixing bugs.
>>>>
>>>> So IMO if we really want to go with staging we need to remove the RC
phase and go from M2 to Final directly. However if we were to do this we would need to
find a way to advertise it as a release on all channels because this is the time when we
need to most testers. Right now it seems to me that an official RC is much more powerful
than staging
>>>>
>>>> Thus I'd like to retract my vote on this (if it's not possible
I'll send a new vote to not do double staging).
>>>>
>>>> Thanks
>>>> -Vincent
>>>>
>>>> PS: Sorry for not realizing this earlier...
>>>>
>>>> On May 22, 2012, at 10:55 AM, Caleb James DeLisle wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> I'd like to add staging to our official release process.
>>>>> For milestone releases, I propose the staging cycle be for "0
time" (this may be revisited later).
>>>>> For RC or finals, we place the release in staging and immediately
call a VOTE to publish the release, this gives our testing team (everybody!) 72 hours to
raise a potential issue.
>>>>>
>>>>> Why:
>>>>>
>>>>> #1. After some chat on IRC I decided that it is advantageous to move
toward a faster release cycle and begin moving away from milestone releases in favor of
staging. This will set the stage for the release method we will need.
>>>>>
>>>>> #2. Staging is easy, I've modified the release script to include
staging and with the script, it is a simple matter of about 5 clicks on nexus to
"login", "close repository", "release repository".
>>>>>
>>>>> #3. Staging is safe, the RM need not worry about fat fingers breaking
the release, all it costs is time.
>>>>>
>>>>> #4. The release process should be as close to the same as possible
for milestone and RC/final releases. This simplifies scripting of the process, decreases
the amount the RM must remember and makes every milestone release a rehearsal.
>>>>>
>>>>> #5. Everybody else is doing it (is that even a reason?!)
>>>>>
>>>>>
>>>>> Mandatory?
>>>>> I would rather impress the RM with how easy and helpful staging can
be than bind him with rules.
>>>>> If I had followed the existing process to the letter, I would not
have had any experience with staging to begin with.
>>>>> In the interest of continuous improvement I would like to make this a
strong recommendation, not a strict rule.
>>>>>
>>>>>
>>>>> Here's my +1
>>>>>
>>>>> Caleb