Skip to content

Latest commit

 

History

History
340 lines (253 loc) · 13.2 KB

development.md

File metadata and controls

340 lines (253 loc) · 13.2 KB

Developer's Guide

The agent is built from a single multi-stage Dockerfile. This requires Docker 17.06+. There is a dev image that can be built for more convenient local development. To use it first do the following:

$ make dev-image
$ # The build will take a while...
$ mkdir local-etc && cp packaging/etc/agent.yaml local-etc/  # Get a basic operating config file in place
$ make run-dev-image

If all went well, you will now be attached to a shell inside of the dev image at the Go package dir where you can compile and run the agent.

Inside this image the agent "bundle" (which basically means collectd and all of its dependencies) is at /bundle, and the rest of the dev image contains useful tools for development, such as a golang build environment.

Within this image, you can build the agent with make signalfx-agent and then run the agent with ./signalfx-agent. The code directory will be mounted in the container at the right place in the Go path so that it can be built with no extra setup. There is also an environment variable SIGNALFX_BUNDLE_DIR set to /bundle so that the agent knows where to find the bundle when run. The agent binary itself is statically compiled, so it has no external library dependencies once built.

You can put agent config in the local-etc dir of this repo and it will be shared into the container at the default place that the agent looks for config (/etc/signalfx). The local-etc dir is ignored by git.

Development Practices

  • Try and do logging at the highest level of the component you are work on as possible. In the lower-level components, return errors (appropriately wrapped with fmt.Errorf("context of error: %v", err) to provide context) and only at the highest level where it doesn't make sense to return errors any more, should you log the error. E.g. for a monitor, the most approrpriate place to do logging is in the function that gets called on an interval.

  • Try and minimize memory allocations as much as possible. Allocations result in higher garbage collection CPU usage. Don't go crazy on trying to avoid allocations in every case where it's possible, but be aware of it. If your code is signficantly harder to understand then it probably isn't worth doing unless profiling shows a large benefit.

Profiling the agent

You can profile the agent with the pprof tool from Go. To enable a profiling HTTP endpoint in the agent, set profiling: true in the agent config. Then you can hit various endpoints on http://localhost:6060/debug/pprof/* (where * is various profiles documented here).

Improve build times on Mac

When developing on Mac and building in a Docker Linux container the source directory is shared using Docker volumes. It is relatively slow and increases build times. A quicker method (2-3x faster) is to do syncing of files to the Docker VM so that file access is in the same host as the Linux container. docker-sync will do this automatically once setup.

Once installed (see below) you run:

$ docker-sync start

Then use docker-compose to run the container in the background:

$ docker-compose up -d shell

You can now run any number of shells inside the container:

$ docker-compose exec shell bash

When you want want to rebase against a new dev-image run make dev-image then update:

$ docker-compose up -d shell

To reset containers:

$ docker-compose down

To reset everything including volumes:

$ docker-compose down --volumes

Installing docker-sync

Install the docker-sync gem:

$ gem install --user-install docker-sync

Make sure the user gems bin directory is in your PATH, for example:

export PATH="${PATH}:$(ruby -r rubygems -e 'puts Gem.user_dir')/bin"

Making the Final Docker Image

To make the final Docker image without all of the development tools, just run make image (either in or outside of the dev image) and it will make a new agent image with the name quay.io/signalfx/signalfx-agent-dev:<agent version>. The agent version will be automatically determined from the git repo status, but can be overridden with the AGENT_VERSION envvar. The image name itself can be overridden with the AGENT_IMAGE_NAME envvar.

Making the Standalone Bundle

To make the standalone .tar.gz bundle, simply run make bundle (either in or outside of the dev image). It will dump a file with the name signalfx-agent-<agent version>.tar.gz in the current directory. You can override the agent version used with the AGENT_VERSION envvar; otherwise, the version will be automatically inferred from the git repo.

Contributing

If you are a SignalFx employee, you should make commits to a branch off of the main code repository at https://github.com/signalfx/signalfx-agent and make a pull request back to the main branch. If you are not an employee, simply fork that repository and make pull requests back to our repo's main branch. We welcome any enhancements you might have, and will try to respond to all issues and pull requests quickly.

Trivial Commits

If you have a very simple commit that should not require a full CI run, just put the text [skip ci] in the commit message somewhere and CircleCI will not run for that commit.

Go Dependencies

We are using Go modules to manage dependencies.

If you add another Go package dependency, run go mod tidy and commit the updated go.mod and go.sum to the repository.

Development in Kubernetes (K8s)

You can use minikube when testing certain aspects of the K8s observers and monitors, but minikube is limited to a single node.

If you are a SignalFx employee, we have a private quay.io repository at quay.io/signalfx/signalfx-agent-dev where you can push test images to be deployed to K8s. If you are not an employee, quay.io offers free repositories as long as they are public, so you can make one.

Helm makes it easy to deploy the agent as well as services to monitor on K8s. There is a Helm values file for development in this repo that will use the quay.io private repo.

Developing on Mac OSX

The agent should compile fine on Mac, but collectd will not run. Also the Python subprocess based monitors will run but you have to customize the Python binary path.

Run the following to make the local-etc directory in your local checkout be linked to the main etc dir so that you don't have to pass any special flags to the agent when running it:

$ cd /path/to/signalfx-agent
$ mkdir local-etc
$ sudo ln -s /path/to/signalfx-agent/local-etc /etc/signalfx

The agent will look for local-etc/agent.yaml for its config.

Here is a sample config for use on Mac that will disable collectd and leave on a simple set of monitors:

---
# observers are what discover running services in the environment
observers:
  - type: docker
  - type: host

monitors:
 - type: host-metadata
 - type: memory
 - type: internal-metrics

collectd:
  disableCollectd: true

Running tests

The agent comes with a suite of unit and integration tests that exercise various components within the agent. All of these tests must pass for a branch to be merged into the mainline main branch. Our CircleCI configuration will automatically run them when a pull request is made, but you can run them manually as follows:

Go Unit Tests

Simply run make tests. You should add new unit tests for any new modules that have relatively self-contained functionality that is easy to isolate and test.

Integration Tests

These are all written using Python's pytest and are located in the tests directory. See there for more information.

Lint

We require 100% passing rate for the standard golint tool, which can be run with make lint.

Vet

We also require 100% passing for go vet for non-test code. Test code can fail if there is a good reason.

ARM targets

Building and testing for Linux targets with ARM processors can be done on any ARM environment. Amazon EC2 instances running Amazon Linux 2 for arm64 are known to support the build and test process.

Windows

We develop on a VirtualBox Windows Server 2008 Vagrant. You might want to develop on Windows Server 2012+ if you're using the evaluation boxes because the Windows Server 2008 evaluation only has a 10 day trial that can be renewed up to 5 times. To renew the Windows Server 2008 evaluation you must manually reset the activation period by using the slmgr.vbs script from the command prompt and restart the vm.

slmgr.vbs –rearm

Base Box

If you have a valid Windows Vagrant base box, set the box name in the windows Vagrant File.

If you do not have a base box, the makefile target win-vagrant-base-box will checkout the Windows Boxcutter Project and build the Windows Server base box image using the evaluation copy of Windows.

Please ensure that the requisites for Boxcutter are satisfied, including the installation of Packer, VirtualBox, and Vagrant.

Make File Targets

For convenience the Makefile in the scripts/windows/vagrant directory of this project has the following targets:

Target Description Example
win-vagrant-base-box Builds a base box using the Windows Boxcutter Project make win-vagrant-base-box
win-vagrant-up Alias for vagrant up that will start and provision the vagrant if necessary. You should be presented with virtualbox vm GUI window when complete. make win-vagrant-up
win-vagrant-destroy Alias for vagrant destroy that will destroy the vagrant make win-vagrant-destroy
win-vagrant-suspend Alias for vagrant suspend that will suspend the vagrant make win-vagrant-suspend
win-vagrant-provision Alias for vagrant provision that will suspend the vagrant make win-vagrant-provision

By default the makefile uses Windows Server 2008. If you want to override this, set the environment variable WIN_VER to choose a different version.

The following values are supported for WIN_VER

Value Windows Version Vagrant Base Box Name Virtual Box VM Name
server_2008 Windows Server 2008 r2 eval-win2008r2-standard-ssh Windows_Server_2008_SignalFx_Agent
server_2012 Windows Server 2012 r2 eval-win2012r2-standard-ssh Windows_Server_2012_SignalFx_Agent
server_2016 Windows Server 2016 eval-win2016-standard-ssh Windows_Server_2016_SignalFx_Agent

Building and Starting the Vagrant

The following snippet will create the vagrant base box, start the vagrant, provision, suspend, and destroy it.

$ cd $GOPATH/src/github.com/signalfx/signalfx-agent/scripts/windows/vagrant
$ WIN_VER=server_2008 make win-vagrant-base-box
  ...
$ WIN_VER=server_2008 make win-vagrant-up
  ...
$ WIN_VER=server_2008 make win-vagrant-suspend
  ...
$ WIN_VER=server_2008 make win-vagrant-destroy

win-vagrant-base-box, win-vagrant-up, and win-vagrant-provision can take a significant amount of time to complete and depend on the characteristics of your host environment.

Software Provisioned in the Vagrant

The vagrant will be provisioned with:

Host Files Included in the Vagrant

This github project github.com/signalfx/signalfx-agent will be mapped as a synced directory to C:\Users\vagrant\signalfx-agent.

Building For Windows

The vagrant box should have enough dependencies installed that you can build the agent bundle. To do this navigate to the project in the GOPATH.

$ cd C:\Users\vagrant\signalfx-agent

$ scripts/windows/make.ps1 bundle

External Compilation of the Agent

To compile the agent externally, create a separate Go module in your own repository. The most basic wrapper around the agent is the following file (put in your main package):

package main

import "github.com/signalfx/signalfx-agent/cmd/agent/agentmain"

func main() {
	agentmain.Run()
}

You should use Go modules and pin the agent version to a specific commit hash in the agent since the agent does not use proper major versioning with Go modules.