I'm interested in using chef+gerrit+jenkins to test new cookbooks in
our environment. I'm thinking of a workflow like this:
git checkin of a cookbook.
The cookbook will have an associated unit test suite that jenkins will
run by deploying to a VM (probably an amazon EC2 t1.micro unless a
particular architecture is needed). The cookbook will be run and if
it checks out, the version will be recorded as a candidate to become
elevated to the next environment. I'm thinking of:
sandbox (least reliable)
dev (someone's at least looking at it, may get incorporated in
integration tests)
staging (will be used for pre-producting integration testing)
production (the real thing)
My only question is how to inject version numbers into the metadata in
a sane manner. I can create/manipulate metadata.rb, but I can't see
an easy and clean way to put basic assurances with the X.Y.Z
versioning.
The closest I've gotten is something like this:
checkin foo with metadata.rb's version == 0.1.1. All X.Y.Z releases
where (X%2 == false) are submitted by the dev/user, so can never be
part of dev, staging, or production.
When it's committed and pushed to the git/gerrit server a jenkins job
is kicked off which uploads the new cookbook and launches a server
with this cookbook version explicitly in the run list to verify that
the job passes. If it passes, then the jenkins server increments the
metadata so that version = 0.2.1 (matching Z values) and commits that
to a branch that doesn't require review. That branch then gets
uploaded to the chef server again, but now with the new version
number.
When moving from unit test to integration tests (e.g. with
cluster_chef/metachef, doing an entire environment or facet build)
then the Y value will get incremented to 3, 4, or 5 as testing
proceeds.
The thing I'm trying to solve is being able to use the same chef
account for the life of the cookbook so that version X.Y.Z will always
be the same if called in any environment. The issue I'm having is
that fencing off particular minor (Y) revision numbers seems to invite
mis-understandings and casual mistakes.
Has anyone else dealt with this issue, or does anyone have a more
enlightened way of looking at the problem?
-Peter
On Thu, Jan 5, 2012 at 2:15 PM, James Litton james@pagerduty.com wrote:
The git repo clone on the host is not necessary. It could certainly be
done from a dev box or build server. Doing that from a build server
would make a lot of sense, actually. The reason has just been a matter
of convenience. We don't have to run multi-knife or a similar system
because dev boxes always push to dev, test to test and so on. This
also prevents an accidental push to prod in the case that someone
forgets to branch after push, etc.
versioning of cookbooks has been entirely haphazard(at best) with no
ill effects so far, but this approach has only been in place a month
or so. none of our cookbook depends specify version numbers. The
biggest problem with this approach has been that it makes it more
difficult to commit back to public cookbooks.
No, that situation has not come up. If it were to come up, we would
roll back to that point in the git repository not just the cookbook.
James
On Thu, Jan 5, 2012 at 1:12 AM, Torben Knerr ukio@gmx.de wrote:
Hi James,
thanks for sharing! Interesting to see the different approaches you can go
for.
Few questions about your approach:
-
as you describe you have a git repo clone of your chef repository on each
of your chef servers, then check out locally on the chef server and then
knife upload to localhost if i understood correctly. Why do you need a git
repo clone of your chef repository on the chef server? Wouldn't it be enough
to knife upload from an external box (e.g. developer's box, build server) to
either of your chef servers?
-
as far as i understand your cookbook dependencies are versionless during
delevopment/upload/test/fix cycles and you always release the whole chef
repo instead of single cookbooks. As soon as you do a release, does the
script you mention update the Version number of all your cookbooks? Does it
at this point in time also add this version number to the cookbook depends
statements (and remove it afterwards for the next dev/upload/test/fix
cycle)?
-
do you ever have the situation that you need to set up a node with an
older cookbook version (e.g. for regression testing)?
Cheers,
Torben
On Thu, Jan 5, 2012 at 4:24 AM, James Litton james@pagerduty.com wrote:
Torben,
Thank you for bringing this up. I have also been puzzling about this
recently.
Right now the process that I am using involves 3 chef servers.
Development:
- Branch git
- Test/Deploy on Chef Solo (dedicated dev box with separate role in
prod chef server)
Test:
- Checkout branch on test chef server
- Push entire configuration (via shell script that I wrote)
- Force a chef-client run on a hosts in test by environment (Load
Test, staging, preview)
- Complete testing
Code Review
- Review the changes in the branch as well as the test cases
Prod (hosts developer environment and production hosts):
- Merge branch to master
- git pull on prod
- Push config via script
- Force chef-client run here only if required due to external requirement
The only issue really is the Test Chef server sharing, but testing can
generally be done quickly by maintaining smaller updates.
cookbook versions number changes aren't as important due to re-upload
of the cookbook with each push and splitting environments across chef
servers
The release process is not currently automated other than the script
that updates all roles, environments, data bags and cookbooks with
each release.
Generally the entire chef repository is treated as a version/release
in order to avoid version dependencies within each cycle.
I'd also love to hear what other people are doing.
James
PagerDuty
On Wed, Jan 4, 2012 at 11:58 AM, Torben Knerr ukio@gmx.de wrote:
Hi everybody,
Happy New Year! 
This is my first post to this mailing list. I'm trying to implement a
proper
workflow for developing cookbooks within a small team. Coming from a
slightly different background (Java, Maven and the like) I'm looking for
some experience on how to implement a typical develop/test/release cycle
in
a Chef environment.
So, assuming you are developing cookbooks in a team, how would you go
for
adding a new feature to a cookbook:
(a) develop -> test (using Vagrant/Chef Solo) -> repeat
develop/fix/test
cycle until done -> upload to team's chef server
(b) develop -> test (upload to a private chef server) -> repeat
develop/fix/test cycle until done -> upload to team's chef server
(c) develop -> test (upload directly to team's chef server) -> repeat
develop/fix/test cycle until done
I'm especially interested in:
-
at which point in time do you share your sources (git push)?
- are you working on feature branches, then merge + push when done?
- when do you increase the cookbook version? in the commit where you
merge back to master?
- when do you create the tag? everytime you commit something on
master?
-
at which point in time do you upload the cookbook to the team's chef
server (knife upload)?
- whenever you want to test something?
- strictly only after creating a tag (i.e. releases only)?
- do you always freeze uploaded cookbooks?
- do you allow to overwrite cookbooks at all?
-
how do you handle releases?
- is every upload to the chef repository automatically a (public?)
release for you?
- do you have the release process (roughly: "run tests -> create tag
->
upload cookbook") automated?
- how do you handle cookbook dependencies when you release? do you
allow
for versionless cookbook depends?
Eager to hear your experiences. Any help is appreciated!
Cheers,
Torben
