Hacking on OpenShift
====================
## Building a Release
To build an OpenShift release you run the `hack/build-release.sh` script on a
system with Docker, which will create a build environment image and then
execute a cross platform Go build within it. The build output will be copied
to `_output/releases` as a set of tars containing each version. It will also
build the `openshift/origin-base` image which is the common parent image for all
OpenShift Docker images.
$ make release
NOTE: Only committed code is built.
To build the base and release images, run:
$ hack/build-base-images.sh
Once a release has been created, it can be pushed:
$ hack/push-release.sh
To cut an official tag release, we generally use the images built by
[ci.openshift.redhat.com](https://ci.openshift.redhat.com)
under the devenv_ami job.
1. Create a new git tag `git tag vX.X.X -a -m "vX.X.X" HEAD`
2. Push the tag to GitHub `git push origin --tags` where `origin` is
`github.com/openshift/origin.git`
3. Run the "devenv_ami" job
4. Once the images are pushed to the repository, run `OS_PUSH_TAG="vX.X.X"
hack/push-release.sh`. Your tag must match the Git tag.
5. Upload the binary artifacts generated by that build to GitHub release page
6. Send an email to the dev list, including the important changes prior to the
release.
We generally cut a release before disruptive changes land.
## Test Suites
OpenShift uses three levels of testing - unit tests, integration test, and
end-to-end tests (much like Kubernetes).
### Unit tests
Unit tests follow standard Go conventions and are intended to test the behavior
and output of a single package in isolation. All code is expected to be easily
testable with mock interfaces and stubs, and when they are not it usually means
that there's a missing interface or abstraction in the code. A unit test should
focus on testing that branches and error conditions are properly returned and
that the interface and code flows work as described. Unit tests can depend
on other packages but should not depend on other components (an API test should
not be writing to etcd).
The unit tests for an entire package should not take more than 0.5s to run, and
if they do, are probably not really unit tests or need to be rewritten to avoid
sleeps or pauses. Coverage on a unit test should be above 70% unless the units
are a special case.
See `pkg/template/generator` for examples of unit tests. Unit tests should
follow Go conventions.
Run the unit tests with:
$ hack/test-go.sh
or an individual package using its relative path with:
$ hack/test-go.sh pkg/build
or an individual package and all packages nested under it:
$ hack/test-go.sh pkg/build/...
To run only a certain regex of tests in a package, use:
$ hack/test-go.sh pkg/build -test.run=SynchronizeBuildRunning
To get verbose output for the above example:
$ hack/test-go.sh pkg/build -test.run=SynchronizeBuildRunning -v
To run all tests with verbose output:
$ hack/test-go.sh -v
To change the timeout for individual unit tests, which defaults to one minute,
use:
$ TIMEOUT=<timeout> hack/test-go.sh
To enable running the kubernetes unit tests:
$ TEST_KUBE=true hack/test-go.sh
To run unit test for an individual kubernetes package:
$ hack/test-go.sh vendor/k8s.io/kubernetes/examples
To change the coverage mode, which is `-cover -covermode=atomic` by default,
use:
$ COVERAGE_SPEC="<some coverage specification>" hack/test-go.sh
To turn off coverage calculation, which is on by default, use:
$ COVERAGE_SPEC= hack/test-go.sh
To run tests without the go race detector, which is on by default, use:
$ DETECT_RACES= hack/test-go.sh
To create a line coverage report, set `OUTPUT_COVERAGE` to a path where the
report should be stored. For example:
$ COVERAGE_OUTPUT_DIR='/path/to/dir' hack/test-go.sh
After that you can open `/path/to/dir/coverage.html` in the browser.
To generate a jUnit XML report from the output of the tests, and see a summary
of the test output instead of the full test output, use:
$ JUNIT_REPORT=true hack/test-go.sh
`hack/test-go.sh` cannot generate jUnit XML and a coverage report for all
packages at once. If you require both, you must call `hack/test-go.sh` twice.
### Integration tests
Integration tests cover multiple components acting together (generally, 2 or
3). These tests should focus on ensuring that naturally related components work
correctly. They should not be extensively testing branches or error conditions
inside packages (that's what unit tests do), but they should validate that
important success and error paths work across layers (especially when errors
are being converted from lower level errors). Integration tests should not be
testing details of the inter-component connections - API tests should not test
that the JSON serialized to the wire is correctly converted back and forth (unit test
responsibility), but they should test that those connections have the expected
outcomes. The underlying goal of integration tests is to wire together the most
important components in isolation. Integration tests should be as fast as possible
in order to enable them to be run repeatedly during testing. Integration tests
that take longer than 0.5s are probably trying to test too much together and
should be reorganized into separate tests. Integration tests should generally
be written so that they are starting from a clean slate, but if that involves
costly setup those components should be tested in isolation.
We break integration tests into two categories, those that use Docker and those
that do not. In general, high-level components that depend on the behavior of code
running inside a Docker container should have at least one or two integration tests
that test all the way down to Docker, but those should be part of their own
test suite. Testing the API and high level API functions should generally
not depend on calling into Docker. They are denoted by special test tags and
should be in their own files so we can selectively build them.
All integration tests are located under `test/integration/*`. All integration
tests must set the `integration` build tag at the top of their source file,
and also declare whether they need etcd with the `etcd` build tag and whether
they need Docker with the `docker` build tag. For special function sets please
create sub directories like `test/integration/deployimages`.
Run the integration tests with:
$ hack/test-integration.sh
The script launches an instance of etcd and then invokes the integration tests.
If you need to execute a subset of integration tests, run:
$ hack/test-integration.sh <regex>
Where `<regex>` is some regular expression that matches the names of all
of the tests you want to run. The regular expression is passed into `grep -E`,
so ensure that the syntax or features you use are supported. The default
regular expression used
is `Test`, which matches all tests.
Each integration function is executed in its own process so that it cleanly
shuts down any background
goroutines. You will not be able to run more than a single test within a single
process.
There is a CLI integration test suite which covers general non-Docker
functionality of the CLI tool
working against the API. Run it with:
$ hack/test-cmd.sh
This suite comprises many smaller suites, which are found under `test/cmd` and
can be run individually by specifying a regex filter, passed through `grep -E`
like with integration tests above:
$ hack/test-cmd.sh <regex>
During development, you can run a file `test/cmd/*.sh` directly to test against
a running server. This can speed up the feedback loop considerably. All
`test/cmd/*` tests are expected
to be executable repeatedly - please file bugs if a test needs cleanup before
running.
For example, start the OpenShift server, create a "test" project, and then run
`oc new-app` tests against the server:
$ oc new-project test
$ test/cmd/newapp.sh
In order to run the suite, generate a jUnit XML report, and see a summary of
the test suite, use:
$ JUNIT_REPORT='true' hack/test-cmd.sh
### End-to-End (e2e) and Extended Tests
The final test category is end to end tests (e2e) which should verify a long
set of flows in the product as a user would see them. Two e2e tests should not
overlap more than 10% of function, and are not intended to test error conditions
in detail. The project
examples should be driven by e2e tests. e2e tests can also test external
components working together.
The end-to-end suite is currently implemented primarily in Bash, but will be
folded into the extended suite (located in test/extended) over time.
The extended suite is closer to the upstream Kubernetes e2e suite and
tests the full behavior of a running system.
Run the end to end tests with:
$ hack/test-end-to-end.sh
Run the extended tests with:
$ test/extended/core.sh
This suite comprises many smaller suites, which are found under `test/extended`
and can be run individually by specifying `--ginkgo.focus` and a regex filter:
$ test/extended/core.sh --ginkgo.focus=<regex>
In addition, the extended tests can be ran against an existing OpenShift
cluster:
$ KUBECONFIG=/path/to/admin.kubeconfig TEST_ONLY=true test/extended/core.sh --ginkgo.focus=<regex>
Extended tests should be Go tests in the `test/extended` directory that use
the Ginkgo library. They must be able to be run remotely, and cannot depend on
any local interaction with the filesystem or Docker.
More information about running extended tests can be found in
[test/extended/README](https://github.com/openshift/origin/blob/master/test/extended/README.md).
## Installing Godep
OpenShift and Kubernetes use [Godep](https://github.com/tools/godep) for
dependency management. Godep allows versions of dependent packages to be
locked at a specific commit by *vendoring* them (checking a copy of them into
`vendor/`).
This means that everything you need for OpenShift is checked into this
repository.
To install `godep` locally run:
$ go get github.com/tools/godep
If you are not updating packages you should not need godep installed.
## Cherry-picking an upstream commit into Origin: Why, how, and when.
Origin carries patches inside of vendor/ on top of each rebase.
Thus, origin carries upstream patches in two ways.
1. *periodic rebases* against a Kubernetes commit.
Eventually, any code you have in upstream kubernetes will land in Openshift
via this mechanism.
2. Cherry-picked patches for important *bug fixes*. We really try to
limit feature back-porting entirely.
### Manually
You can manually try to cherry pick a commit (by using git apply). This can
easily be done in a couple of steps.
- wget the patch, i.e. `wget -O /tmp/mypatch
https://github.com/kubernetes/kubernetes/pull/34624.patch`
- PATCH=/tmp/mypatch git apply --directory vendor/k8s.io/kubernetes $PATCH
If this fails, then it's possible you may need to pick multiple commits.
### For Openshift newcomers: Pick my kubernetes fix into Openshift vs. wait for
the next rebase?
Assuming you read the bullets above... If your patch is really far behind, for
example, if there have been 5 commits modifying the directory you care about,
cherry picking will be increasingly difficult and you should consider waiting
for the next rebase, which will likely include the commit you care about, or at
least decrease the amount of cherry picks you need to do to merge.
To really know the answer, you need to know *how many commits behind you are in
a particular directory*, often.
To do this, just use git log, like so (using pkg/scheduler/ as an example).
```
MYDIR=pkg/scheduler/algorithm git log --oneline --
vendor/k8s.io/kubernetes/${MYDIR} | grep UPSTREAM | cut -d' ' -f 4-10 | head -1
```
The commit message printed above will tell you:
- what the LAST commit in Kubernetes was (which effected
"/pkg/scheduler/algorithm")
- directory, which will give you an intuition about how "hot" the code you are
cherry picking is. If it has changed a lot, recently, then that means you
probably will want to wait for a rebase to land.
### Using hack/cherry-pick
For convenience, you can use `hack/cherry-pick.sh` to generate patches for
Origin from upstream commits.
The purpose of this command is to allow you to pull individual commits from a
local kubernetes repository into origin's vendored kuberenetes in a fully
automated manner.
To use this command, be sure to setup remote Pull Request branches in the
kubernetes repository you are using (i.e. like https://gist.github.com/piscisaureus/3342247).
Specifically, you will be doing this, to the git config you probably already
have for kubernetes:
```
[remote "origin"]
url = https://github.com/kubernetes/kubernetes
fetch = +refs/heads/*:refs/remotes/origin/*
### Add this line
fetch = +refs/pull/*/head:refs/remotes/origin/pr/*
```
so that `git show origin/pr/<number>` displays information about your branch
after a `git fetch`.
You must also have the Kubernetes repository checked out in your GOPATH
(visible as `../../../k8s.io/kubernetes`),
with openshift/kubernetes as a remote and fetched:
$ pushd $GOPATH/src/k8s.io/kubernetes
$ git remote add openshift https://github.com/openshift/kubernetes.git
$ git fetch openshift
$ popd
There must be no modified or uncommitted files in either repository.
To pull an upstream commit, run:
$ hack/cherry-pick.sh <pr_number>
This will attempt to create a patch from the current Kube rebase version in
Origin that contains the commits added in the PR. If the PR has already been
merged to the Kube version, you'll get an error. If there are conflicts, you'll
have to resolve them in the upstream repo, then hit ENTER to continue. The end
result will be a single commit in your Origin repo that contains the changes.
If you want to run without a rebase option, set `NO_REBASE=1` before the
command is run. You can also specify a commit range directly with:
$ hack/cherry-pick.sh origin/master...<some_branch>
All upstream commits should have a commit message where the first line is:
UPSTREAM: <PR number|drop|carry>: <short description>
`drop` indicates the commit should be removed during the next rebase. `carry`
means that the change cannot go into upstream, and we should continue to use it
during the next rebase.
You can also target repositories other than Kube by setting `UPSTREAM_REPO` and
`UPSTREAM_PACKAGE` env vars. `UPSTREAM_REPO` should be the full name of the Git
repo as Go sees it, i.e. `github.com/coreos/etcd`, and `UPSTREAM_PACKAGE` must be
a package inside that repo that is currently part of the Godeps.json file. Example:
$ UPSTREAM_REPO=github.com/coreos/etcd UPSTREAM_PACKAGE=store
hack/cherry-pick.sh <pr_number>
By default `hack/cherry-pick.sh` uses git remote named `origin` to fetch
kubernetes repository, if your git configuration is different, you can pass the git
remote name by setting `UPSTREAM_REMOTE` env var:
$ UPSTREAM_REMOTE=upstream hack/cherry-pick.sh <pr_number>
## Moving a commit you developed in Origin to an upstream
The `hack/move-upstream.sh` script takes the current feature branch, finds any
changes to the
requested upstream project (as defined by `UPSTREAM_REPO` and
`UPSTREAM_PACKAGE`) that differ from `origin/master`, and then creates a new
commit in that upstream project on a branch with the same name as your current
branch.
For example, to upstream a commit to OpenShift source-to-image while working
from Origin:
$ git checkout my_feature_branch_in_origin
$ git log --oneline
70ffe7e Docker and STI builder support binary extraction
75a22de UPSTREAM: <sti>: Allow prepared directories to be passed to STI
86eefdd UPSTREAM: 14618: Refactor exec to allow reuse from server
# we want to move our STI changes to upstream
$ UPSTREAM_REPO=github.com/openshift/source-to-image
UPSTREAM_PACKAGE=pkg/api hack/move-upstream.sh
...
# All changes to source-to-image in Godeps/. are now in a commit UPSTREAMED
in s2i repo
$ cd ../source-to-image
$ git log --oneline
c0029f6 UPSTREAMED
... # older commits
The default is to work against Kube.
go
## Updating Kubernetes from upstream
There are a few steps involved in rebasing Origin to a new version of
Kubernetes. We need to make sure
that not only the Kubernetes packages were updated correctly into `Godeps`, but
also that *all tests are
still running without errors* and *code changes, refactorings or the
inclusion/removal of attributes
were properly reflected* in the Origin codebase.
### 1. Preparation
Before you begin, make sure you have both
[openshift/origin](https://github.com/openshift/origin) and
[kubernetes/kubernetes](https://github.com/kubernetes/kubernetes) in your
$GOPATH. You may want to work
on a separate $GOPATH just for the rebase:
```
$ go get github.com/openshift/origin
$ go get k8s.io/kubernetes
```
You must add the Origin GitHub fork as a remote in your k8s.io/kubernetes repo:
```
$ cd $GOPATH/src/k8s.io/kubernetes
$ git remote add openshift git@github.com:openshift/kubernetes.git
$ git fetch openshift
```
Check out the version of Kubernetes you want to rebase as a branch or tag named
`stable_proposed` in
[kubernetes/kubernetes](https://github.com/kubernetes/kubernetes). For example,
if you are going to rebase the latest `master` of Kubernetes:
```
$ cd $GOPATH/src/k8s.io/kubernetes
$ git checkout master
$ git pull
$ git checkout -b stable_proposed
```
### 2. Rebase Origin to the new Kubernetes version
#### 2.1. First option (preferred): using the rebase-kube.sh script
If all requirements described in *Preparation* were correctly attended, you
should not have any trouble with rebasing the Kubernetes code using the script
that automates this process.
```
$ cd $GOPATH/src/github.com/openshift/origin
$ hack/rebase-kube.sh
```
Read over the changes with `git status` and make sure it looks reasonable.
Check specially the `Godeps/Godeps.json` file to make sure no dependency
is unintentionally missing.
Commit using the message `bump(k8s.io/kubernetes):<commit SHA>`, where
`<commit SHA>` is the commit id for the Kubernetes version we are including in
our Godeps. It can be found in our `Godeps/Godeps.json` in the declaration of
any Kubernetes package.
#### 2.2. Second option: manually
If for any reason you had trouble rebasing using the script, you may need to to
do it manually.
After following all requirements described in the *Preparation* topic, you will
need to run
`godep restore` from both the Origin and the Kubernetes directories and then
`godep save ./...`
from the Origin directory. Follow these steps:
1. `$ cd $GOPATH/src/github.com/openshift/origin`
2. `make clean ; godep restore` will restore the package versions specified in
the `Godeps/Godeps.json` of Origin to your GOPATH.
2. `$ cd $GOPATH/src/k8s.io/kubernetes`
3. `$ git checkout stable_proposed` will checkout the desired version of
Kubernetes as branched in *Preparation*.
4. `$ godep restore` will restore the package versions specified in the
`Godeps/Godeps.json` of Kubernetes to your GOPATH.
5. `$ cd $GOPATH/src/github.com/openshift/origin`.
6. `$ make clean ; godep save ./...` will save a list of the checked-out
dependencies to the file `Godeps/Godeps.json`, and copy their source code into `vendor`.
7. If in the previous step godep complaints about the checked out revision of a
package being different than the wanted revision, this probably means there are
new packages in Kubernetes that we need to add. Do a `godep save <pkgname>` with
the package specified by the error message and then `$ godep save ./...`
again.
8. Read over the changes with `git status` and make sure it looks reasonable.
Check specially the `Godeps/Godeps.json` file to make sure no dependency is
unintentionally missing. The whole Godeps directory will be added to version control, including
`_workspace`.
9. Commit using the message `bump(k8s.io/kubernetes):<commit SHA>`, where
`<commit SHA>` is the commit id for the Kubernetes version we are including in
our Godeps. It can be found in our `Godeps/Godeps.json` in the declaration of
any Kubernetes package.
If in the process of rebasing manually you found any corner case not attended
by the `hack/rebase-kube.sh` script, make sure you update it accordingly to help future rebases.
### 3. cherry-pick upstream changes pushed to the Origin repo
Eventually during the development cycle we introduce changes to dependencies
right in the Origin
repository. This is not a largely recommended practice, but it's useful if we
need something that,
for example, is in the Kubernetes repository but we are not doing a rebase yet.
So, when doing the next
rebase, we need to make sure we get all these changes otherwise they will be
overridden by `godep save`.
1. Check the `Godeps` directory [commits
history](https://github.com/openshift/origin/commits/master/Godeps)
for commits tagged with the *UPSTREAM* keyword. We will need to cherry-pick
*all UPSTREAM commits since
the last Kubernetes rebase* (remember you can find the last rebase commit
looking for a message like
`bump(k8s.io/kubernetes):...`).
2. For every commit tagged UPSTREAM, do `git cherry-pick <commit SHA>`.
3. Notice that eventually the cherry-pick will be empty. This probably means
the given change were
already merged in Kubernetes and we don't need to specifically add it to our
Godeps. Nice!
4. Read over the commit history and make sure you have every UPSTREAM commit
since the last rebase
(except only for the empty ones).
### 4. Refactor Origin to be compliant with upstream changes
After making sure we have all the dependencies in place and up-to-date, we need
to work in the Origin
codebase to make sure the compilation is not broken, all tests pass and it's
compliant with any refactorings, architectural changes or behavior changes
introduced in Kubernetes. Make sure:
1. `make clean ; hack/build-go.sh` compiles without errors and the standalone
server starts correctly.
1. all of our generated code is up to date by running all `hack/update-*`
scripts.
1. `hack/verify-open-ports.sh` runs without errors.
1. `hack/copy-kube-artifacts.sh` so Kubernetes tests can be fully functional.
The diff resulting from this script should be squashed into the Kube bump
commit.
2. `TEST_KUBE=1 hack/test-go.sh` runs without errors.
3. `hack/test-cmd.sh` runs without errors.
3. `hack/test-integration.sh` runs without errors.
3. `hack/test-end-to-end.sh` runs without errors.
See *Building a Release* above for setting up the environment for the
*test-end-to-end.sh* tests.
It is helpful to look at the Kubernetes commit history to be aware of the major
topics. Although it
can potentially break or change any part of Origin, the most affected parts are
usually:
1. https://github.com/openshift/origin/blob/master/pkg/cmd/server/start
2.
https://github.com/openshift/origin/blob/master/pkg/cmd/server/kubernetes/master
.go
3.
https://github.com/openshift/origin/blob/master/pkg/cmd/server/origin/master.go
4.
https://github.com/openshift/origin/blob/master/pkg/cmd/util/clientcmd/factory.g
o
5. https://github.com/openshift/origin/blob/master/pkg/cmd/cli/cli.go
6. https://github.com/openshift/origin/blob/master/pkg/api/meta/meta.go
Place all your changes in a commit called "Refactor to match changes upstream".
### 5. Pull request
A typical pull request for your Kubernetes rebase will contain:
1. One commit for the Kuberentes Godeps bump (`bump(k8s.io/kubernetes):<commit
SHA>`).
2. Zero, one, or more bump commits for any **shared** dependencies between
Origin and Kubernetes that have been bumped. Any transitive dependencies coming
from Kubernetes should be squashed in the Kube bump commit.
3. Zero, one, or more cherry-picked commits tagged UPSTREAM.
4. One commit "Boring refactor to match changes upstream" that includes boring
changes like imports rewriting, etc.
5. One commit "Interesting refactor to match changes upstream" that includes
interesting changes like new plugins or controller changes.
## Updating other Godeps from upstream
To update to a new version of a dependency that's not already included in
Kubernetes, checkout the correct version in your GOPATH and then run
`godep save <pkgname>`. This should create a new version of `Godeps/Godeps.json`,
and update `vendor`. Create a commit that includes both of these changes with message
`bump(<pkgname>): <pkgcommit>`.
## Updating external examples
`hack/update-external-example.sh` will pull down example files from external
repositories and deposit them under the `examples` directory.
Run this script if you need to refresh an example file, or add a new one. See
the script and `examples/quickstarts/README.md` for more details.
## Troubleshooting
If you run into difficulties running OpenShift, start by reading through the
[troubleshooting
guide](https://github.com/openshift/origin/blob/master/docs/debugging-openshift.
md).
## RPM Packaging
A specfile is included in this repo which can be used to produce RPMs including
the openshift binary. While the specfile will be kept up to date with build
requirements the version is not updated. You will need to either update the
Version, %commit, and %ldflags values on your own or you may use
[tito](https://github.com/dgoodwin/tito) to build
and tag releases.
## GSSAPI-enabled builds
When built with GSSAPI support, the `oc` client supports logging in with
Kerberos credentials on Linux and OS X.
GSSAPI-enabled builds of `oc` cannot be cross-compiled, but must be built on
the target platform with the GSSAPI header files available.
On Linux, ensure the `krb5-devel` package is installed:
$ sudo yum install -y krb5-devel
On OS X, you can obtain header files via Homebrew:
$ brew install homebrew/dupes/heimdal --without-x11
Once dependencies are in place, build with the `gssapi` tag:
$ hack/build-go.sh cmd/oc -tags=gssapi
Verify that the GSSAPI feature is enabled with `oc version`:
$ oc version
...
features: Basic-Auth GSSAPI Kerberos SPNEGO
## Swagger API Documentation
OpenShift and Kubernetes integrate with the [Swagger 2.0 API
framework](http://swagger.io) which aims to make it easier to document and
write clients for RESTful APIs. When you start OpenShift, the Swagger API
endpoint is exposed at `https://localhost:8443/swaggerapi`. The Swagger UI
makes it easy to view your documentation - to view the docs for your local
version of OpenShift start the server with CORS enabled:
$ openshift start --cors-allowed-origins=.*
and then browse to http://openshift3swagger-claytondev.rhcloud.com (which runs
a copy of the Swagger UI that points to localhost:8080 by default). Expand the
operations available on v1 to see the schemas (and to try the API directly).
Additionally, you can download swagger-ui from http://swagger.io/swagger-ui/
and use it to point to your local swagger API endpoint.
Note: Hosted API documentation can be found
[here](http://docs.openshift.org/latest/rest_api/openshift_v1.html).
## Performance debugging
OpenShift integrates the go `pprof` tooling to make it easy to capture CPU and
heap dumps for running systems. The following modes are available for the
`openshift` binary (including all the CLI variants):
* `OPENSHIFT_PROFILE` environment variable:
* `cpu` - will start a CPU profile on startup and write `./cpu.pprof`.
Contains samples for the entire run at the native sampling resolution (100hz).
Note: CPU profiling for Go does not currently work on Mac OS X - the stats are
not correctly sampled
* `mem` - generate a running heap dump that tracks allocations to
`./mem.pprof`
* `block` - will start a block wait time analysis and write `./block.pprof`
* `web` - start the pprof webserver in process at http://127.0.0.1:6060/debug/pprof
(you can open this in a browser). This supports `OPENSHIFT_PROFILE_HOST=`
and `OPENSHIFT_PROFILE_PORT=` to change default ip `127.0.0.1` and default port `6060`.
In order to start the server in CPU profiling mode, run:
$ OPENSHIFT_PROFILE=cpu sudo ./_output/local/bin/linux/amd64/openshift start
Or, if running OpenShift under systemd, append this to
`/etc/sysconfig/atomic-openshift-{master,node}`
OPENSHIFT_PROFILE=cpu
To view profiles, you use
[pprof](http://goog-perftools.sourceforge.net/doc/cpu_profiler.html) which is
part of `go tool`. You must pass the binary you are debugging (for symbols)
and a captured pprof. For instance, to view a `cpu` profile from above, you
would run OpenShift to completion, and then run:
$ go tool pprof ./_output/local/bin/linux/amd64/openshift cpu.pprof
or
$ go tool pprof $(which openshift) /var/lib/origin/cpu.pprof
This will open the `pprof` shell, and you can then run:
# see the top 20 results
(pprof) top20
# see the top 50 results
(pprof) top50
# show the top20 sorted by cumulative time
(pprof) cum=true
(pprof) top20
to see the top20 CPU consuming fields or
(pprof) web
to launch a web browser window showing you where CPU time is going.
`pprof` supports CLI arguments for looking at profiles in different ways -
memory profiles by default show allocated space:
$ go tool pprof ./_output/local/bin/linux/amd64/openshift mem.pprof
but you can also see the allocated object counts:
$ go tool pprof --alloc_objects ./_output/local/bin/linux/amd64/openshift
mem.pprof
Finally, when using the `web` profile mode, you can have the go tool directly
fetch your profiles via HTTP:
# for a 30s CPU trace
$ go tool pprof ./_output/local/bin/linux/amd64/openshift
http://127.0.0.1:6060/debug/pprof/profile
# for a snapshot heap dump at the current time, showing total allocations
$ go tool pprof --alloc_space ./_output/local/bin/linux/amd64/openshift
http://127.0.0.1:6060/debug/pprof/heap
See [debugging Go programs](https://golang.org/pkg/net/http/pprof/) for more
info. `pprof` has many modes and is very powerful (try `tree`) - you can pass
a regex to many arguments to limit your results to only those samples that
match the regex (basically the function name or the call stack).