- Overview - What is puppet-rjil module?
- Details - how does it work?
- Orchestration - how cross host dependencies work
- Development Workflow
- Running Behind Proxy Server
- Build script - command line tool for generating test deployments
- Development - Resources for Developers
This project is the top level project where most of the deployment coding for jiocloud happens.
At a high level, it contains the following files/directories
- ./manifests/ composition Puppet manifests for building jiocloud
- hiera/hiera.yaml hiera configuration file that should be used for jiocloud
- hiera/data/ directory where default hiera data should go. Pretty much anything that doesn't contain lookups or secrets should be stored here.
- Vagrantfile
A Vagrantfile that can be used for faster deployments and iterative
development.
NOTE: Vagrant has its own hiera env data in
./hiera/data/env/vagrant.yaml: - Puppetfile List of all puppet modules that need to be installed as dependencies.
- ./build_scripts/deploy.sh - Script that performs deployment of jiocloud.
- ./files/maybe-upgrade.sh - script that actually runs Puppet to perform deployments on provisioned machines (when not using vagrant)
- ./environment/full.yaml - contains the definition of nodes that can be used for testing
- site.pp - puppet content that contains role assignments to node.
In general, most people be using the predefined full layout for deployments (which deploys the jiocloud reference architecture for openstack)
This deployment module was built using the following steps:
-
Describe the environment you need to deploy (what roles and how many)
-
Ensure that Puppet configuration exists to support this deployment
-
Ensure the relevant hiera data exists in the hierarchy
-
Use either the build script (deploy.sh) or vagrant to perform an environment build
-
Debug the build process to follow a build's progress
Puppet is responsible for ensuring that each of the defined hosts in our layout get assigned their correct role.
For example, for the case of an apache server. Puppet contains the description of how a machine is transitioned to its desired role.
+----------+ +-----------+
| base | | configured|
| image +------> apache |
| | | server |
| | | |
+----------+ +-----------+
The provisioned hosts have their hostname set as:
<role><number>_<project_id>
A set of matching nodes from site.pp might look like:
node /^bootstrap\d+/ {
}
node /^apache\d+/ {
}
Those nodes definitions should include whatever classes are required to configure those roles.
All services that Puppet configures are configured through modules that are installed as external dependencies to the puppet-rjil module.
These dependencies are defined in the Puppetfile (which results to modules that are installed using librarian-puppet-simple)
The Puppet-rjil module contains all of the composition roles that combine the external modules into the roles defined in site.pp.
In Puppet's lexicon, the type of content found in this module are referred to as profiles.
TODO: document more about the kinds of things that belong here and our expectations for how those things are developed.
All data used to override defaults of Puppet Profiles located in puppet-rjil/modules are passed in via hiera.
Understanding hiera is a requirement for using this system. Please get started here.
Hiera uses Facter to determine how data is set for a given node:
hiera/hiera.yaml supplies the override configuration that hiera uses to determine how to set hosts
for a given system. The override levels are as follows:
- clientcert/%{::clientcert} - client specific data
- secrets/%{env} - environment specific secrets. This data is provided from env specific packages
- role/%{jiocloud_role} - role specific overrides
- cloud_provider/%{cloud_provider} - overrides specific to a certain cloud provider.
- env/%{env} - environment specific overrides
- secrets/common - default test secrets that are overridden in prod via packages
- common - default overrides (the majority of the hiera data is stored here)
The build script is driven by a set of data that describes the machines that need to be configured.
These build scripts can be found in the environment directory. The filename should be of the form:
./environment/<layout>.yaml
NOTE: layout defaults to full (full openstack install) if not specified
where layout/env is an argument passed into the build script/Vagrantfile.
The following file contains our reference architecture for openstack:
./environment/full.yaml
This file should contain the top level key resources.
Here resources should be listed. Where the names of these resource should match a node expression (specified from Puppet in site.pp).
For example, if the following resources are specified:
resources:
haproxy:
number: 1
...
httpproxy:
number: 2
Layout files can be applied using the command:
python -m jiocloud.apply_resources apply --mappings=environment/${cloud_provider} environment/${layout}.yaml
This command will create only the desired nodes specified in the layout file that do not already exist.
Apply this file will result in host whose names are prefixed with the following strings:
- hapoxy1
- httpproxy1
- httpproxy2
This document is intended to capture the orchestration requirements for our current Openstack installation.
There are two main motivations for writing this document:
-
Our overall orchestration workflow is getting very complicated. It makes sense to document it to ensure that anyone from the team can understand how it works.
-
The workflow is complicated and not as well performing as it could be. This document is intended to capture those issues along with recommendations for how it can be improved.
Configuration orchestration is managed via a combination of registering services in consul and the following module.
Orchestration is currently managed by consul, a tool that provides DNS service registration and discovery.
Consul works as follows:
- The following Puppet Defined resource
rjil::jiocloud::consul::serviceis used to define a service in consul. - Each service registers its ip address as an A record for the address:
<service_name>.service.consul - Each service registers its hostname: .node.consul as an SRV record for
<service_name>.service.consul
Each service registers itself with consul, along with health checks that ensure that services are not actually registered until they are functional. These services are available both through the consul http api as well as via regular DNS tools (like dig)
Puppet uses both DNS as well as the consul API to understand what remote services are available and uses this information to make decisions about if local configuration should be applied.
- block until an address is resolvable
- block until we can retrieve registered A records or SRV records from an address.
- fail if a DNS address is not available
In order to understand the motivation for the design of Puppet + Consul for orchestration, you need to first understand the difference between compile vs. runtime in Puppet.
- compile time - Puppet goes through a separate process of compiling the catalog that will be used to apply the desired configuration state of each agent. During compile time, the following things occur:
- classes are evaluated
- hiera variables are bound
- conditional logic is evaluated
- functions are run
- variables are assigned to resource attributes
This phase processes Puppet manifests, functions, and hiera. In general, data can only be supplied during the compile time phase in Puppet. This means that all data must be available during compile to be used during runtime.
- run time - during Run time, a graph of resources is applied to the system. Most of the logic that is performed during these steps is contained within the actual ruby providers.
The orchestration_utils repo contains all code used to orchestrate configuration based on the current state of registered services in consul.
Functions can be used at runtime to collect data.
- dns_resolve - gets A records for an address
- service_discovery_consul - pull a host => ip hash for a specified hostname.
- runtime_fail - used to trigger a catalog failure which causes the entire subgraph to fail. This is used as a more performant way to fail and retry when certain data is not ready at compile time.
- dns_blocker - blocks until a specified address is registered. This blocks not only dependent resources, but also resources that are not dependencies that just happen to not have run.
- consul_kv_fail - fail a catalog subgraph if a certain key has not been set in consul. This is used to orchestrate arbitrary events besides registered services.
- consul_kv - used to register arbitrary keys from Puppet as a part of run time (meaning that keys can be sure to be inserted only after certain configuration has been applied.
3 kinds of orchestration actions are performed in our Puppet vs. Consul integration. This section will discussed along with its performance and design implications.
Orchestration decision is to simply fail to compile and retry until all external services are ready. We initially tried this approach, but discontinued for the following reason:
- Performance was terrible. Failing at compile time blocked all resource from being able to run.
- Unable to represent cross host circular dependencies.
- Impossible to decouple package installations. Currently, to ensure the best performance, we install all packages as a separate call to puppet apply with --tags package to ensure that package installs never have to be blocked on service level dependencies.
PROS:
- Easy to monitor cross host orchestration flow
- Less spurious failures.
CONS:
- Cannot work unless hard coded DNS addresses can be used
- Leads to some resource executions getting delayed.
We are tending towards a combination of function/type/providers. At compile time, functions are used to query for data which is then forwarded to a type that fails catalog compile if the expected values are not present.
This failure just results in a sub-graph failure, which means that failures are not preventing other resources from being executed.
This section is intended to document all of the cross host dependencies of our current Openstack architecture, and emphasize the performance implications of each step.
-
All machines block for the consul (bootstrap) server to come up.
-
Currently, the stmonleader, contrail controller, and haproxy machine can all start applying configuration immediately.
** The contrail node can currently install itself successfully as soon as consul it up. This is only because it doesn't actually install the service correctly.
** stmonleader can install itself as soon as consul is ready. It may have to run twice to configure OSD's (which we may do in testing, but not in prod)
** haproxy - can install itself, but it does not configure the balance members until the controller nodes are running.
It is worth noting that two of these roles will need to reapply configuration when the rest of the services come online.
- haproxy needs all addresses for all controllers it adds them as pool members
- stmonleader needs to have all it's mons registered as well as an OSD number that matches num replicas.
-
Once stmon.service.consul is registered in consul, stmon, ocdb, and oc can start compiling. These machines have not performed any configuration at all at this point. At the same time, stmonleader might be adding it's ods (which takes two runs)
-
oc will start compiling, but it blocks until the database is resolvable, once that is resolvable, it continues. At the same time stmon's are rerunning Puppet to set up their osd drives.
-
once oc and ocdb are up, haproxy registers pool members.
The below diagram is intended to keep track of what services are dependent on other services for configuration.
+--------+
| consul |
+------+-+---------------+--------------+
| | |
| | |
| | |
| | |
+-------v-----+ +-----v----+ +-----v----+
| stmonleader | | contrail | | haproxy |
+----+-----------+-+ +----------+ +----------+
| |
| |
| |
+---v---+ +---v--+
| stmon | | ocdb |
+-------+ +----+-+
| |
| |
| |
+---v---+ +-v--+
| stmon | | oc |
+-------+ +----+-----+
|
|
|
|
+----v----+
| haproxy |
+---------+
- the system still does not properly distinguish between addresses of services that will be running vs. addresses of things that will be running.
For example: glance, cannot actually be a functional service until keystone has been been registered as a service and it's address propagated to the load balancer. However, the load balancer cannot be properly verified until glance has registered (maybe this is actually not a problem...)
-
We have not yet implemented service watching. Currently, it is possible that all occurrences of a certain service are not property configured. This is because Puppet just continues to run until it can validate a service as functional. The correct way to resolve this is to ensure that you can watch a service, and ensure that Puppet runs to reconfigure things when service addresses change (this would currently apply to zeromq, ceph, and haproxy.)
-
Failing on compile time adds significant wall-clock time to tests, especially because multiple nodes have to be installed completely in serial.
-
The system doesn't really know when it is done running Puppet. We need to somehow understand what the desired cardinality is for a service. Perhaps this should be configured as a validation check (ie: haproxy is only validated when it has the same number of members as there should be configured services.)
You can setup a Dev environment on a bare-metal node with enough resources using Vagrant. Once you have the bare-metal server available, follow these steps.
It is possible that you can use LXC for this, but it is not fully validating.
This setup has been tested with the 1.6.5 installer from Vagrant Home Page The default version of Vagrant in the Ubuntu 14.04 Repo is 1.4.3 which causes an issue.
Vagrant makes it easier to perform iterative development on modules. It allows you to develop on your local laptop and see the effects of those changes on localized VMs in real time. We have designed a vagrant based solution that can be used to create VMs that very closely resemble the openstack environments that we are deploying.
When spinning up VMs for local development, Vagrant/VBox would need to be run on the physical host that has VT-x enabled in its BIOS. Currently, this setup cannot be provisioned and tested inside a VM itself (KVM/VBox).
The following initial setup steps are required to use the vagrant environment:
Option 1: You can create an rc file in the puppet-rjil directory. Name it something that doesn't actually mean something on the system (like vagrant_http_proxy, etc) (e.g., .mayankrc). Disadvantage: You'll have to source it everytime you login. Option 2: You can add the lines to your default .bashrc file in your home directory. That is automatically sourced each time you login.
Write following lines at the bottom of the rc file:
export http_proxy="http://10.135.121.138:3128"
export https_proxy="https://10.135.121.138:3128"
export no_proxy = "localhost, 127.0.0.1"
export env=vagrant-lxc # used to customize the environment to use
Then do $ source .bashrc
In order to use proxy with sudo command, use sudo with -E option, e.g.,
sudo -E apt-get update
In order to use proxy for apt, e.g., create a file in /etc/apt/apt.conf.d/90_proxy. Write following lines:
Acquire::http::proxy "http://10.135.121.138:3128";
Acquire::https::proxy "https://10.135.121.138:3128";
git clone git://github.com/jiocloud/puppet-rjil
cd puppet-rjil
If the git:// protocol doesn't work, use https://
source newtokens.sh
You'll see a consul ID. Copy that consul ID and paste onto last line of the rc file (.bashrc). Note: There may be problems with reusing the same id. You need to be careful that you recreate the env from scratch every-time, or old machines will join the new cluster. So whenever you create a dev env, always run newtokens.sh.
export consul_discovery_token=ca004..7f42f3
NOTE: you will need to run this operation as sudo if you intend to install the gems as system gems. Otherwise, consider installing ruby via rvm to ensure that you can install gems in the local users environment.
First, you need to make sure that your Puppet module dependencies are installed. NOTE: make sure that you install librarian-puppet-simple and not librarian-puppet!!!!
gem install librarian-puppet-simple
librarian-puppet install # from root level of this repo
This step is required b/c the top level directory (i.e. puppet-rjil is a module itself and it needs to exist in the modulepath for the vagrant environment.)
mkdir modules/rjil
The Vagrantfile accepts a few additional environment variables that can be used to further customize the environment.
Once you have initialized your vagrant environment using the above steps, you are ready to start using vagrant.
It is highly recommended that if you intend to use this utility that you be familiar with the basics of vagrant.
To get a list of support roles that can be booted by vagrant, run:
vagrant status
This list is populated using the ./environment/full.yaml file by default. It is possible to customize the roles to be populated by adjusting
To boot your desired role in vagrant
vagrant up <role>
To re-run Puppet provisioning (for interactive development of modules):
vagrant provision <role>
To login to VMs:
vagrant ssh <role_name>
- Ensure that the role exists in site.pp as a node declaration
node /^somerole/ {
include somerole
}
- Ensure the node is defined in Vagrant
Roles in vagrant are populated from layouts. In order to be able to see a machine in vagrant, you need to add it to environment/layout.yaml for your custom layout or for the layout you are working on.
Then you need to set the layout env variable.
- Add relevant hiera data. Vagrant environment specific data should be added to the following location.
./hiera/data/env/<env_name>.yaml
In order to run the system behind proxy server, the following extra configuration is required.
- While running Vagrant or deploy.sh, export http_proxy and https_proxy variables on the system where running them. This will make sure those scripts will use system wide proxy settings.
example
export http_proxy=http://10.135.121.138:3128
export https_proxy=http://10.135.121.138:3128
where the proxy url is http://10.135.121.138:3128.
- Set rjil::system::proxies hiera data, preferably in hiera/data/env/<env_name>.yaml as below. If this is configured, puppet will configure system wide proxy settings on all systems.
rjil::system::proxies:
"no":
url: "127.0.0.1,localhost,consul,jiocloud.com"
"http":
url: "http://10.135.121.138:3128"
"https":
url: "http://10.135.121.138:3128"
where the proxy url is http://10.135.121.138:3128.
- Sometimes, the systems will not have access to upstream ntp servers, in which case that setting have to be changed to internal ntp server (By default, it take pool.ntp.org).
Configure appropriate ntp server in hiera/data/env/<env_name>.yaml.
example:
upstream_ntp_servers:
- 10.135.121.138
This project contains the following script that is used to deploy out openstack environments.
build\_scripts/deploy.sh
The script can be configured using the following environment variables:
Used to attach a unique identifier to all vms created as a part of a test deployment. This is useful for distinguishing between separate sets of machines that may be on the same network.
If you are going to be performing multiple deployments for testing purposes, it is recommended to use a timestamp to make the build number unique per invocation:
export BUILD\_NUMBER=`date +"%d%m%y%H%M%S"`;
This build number is appended to all of the hostname prefixes of the provisioned machines.
The environment indicator.
export env=dan
This is used to determine the file that specifies what machines should exist for a specific deployment. It refers to the file environment/cloud.${env}.yaml.
It is also used to pull env specific hiera data from ./hiera/data/env/${env}.yaml
The name of the keypair that will used assigned to the VM.
export KEY\_NAME=openstack\_ssh\_key
File that contains the following environment variables used to configuration access and authentication against your openstack environment.
This argument is optional provided that you manually provide those environment variables.
export env\_file=~/credentials.sh
Name of the command that provides the timeout capability. You should only even have to set this if you are using the script from a Mac (or other Unix system).
export timeout\_command=gtimeout
Location of source repository for puppet-rjil. This is required if you want to test against a branch instead of against the puppet-rjil contents that are packaged. This is very handy for pre-validating some configuration before running tests.
export puppet\_modules\_source\_repo=https://github.com/bodepd/puppet-rjil
Branch that should be used from the provided puppet-rjil repo.
export puppet\_modules\_source\_branch=origin/deploy\_script
Allows a user to specify an alternative location for python jiocloud so that users can test builds with patches to this repository.
This is intended to make testing changes to this external python library easier.
Branch that should be used to install python-jiocloud.
User to that build process should use to ssh into bootstrap server.
ssh\_user=ubuntu
Protocol to use when fetching puppet modules. Defaults to git.
export git_protocol=https
Provide a DNS server to use during bootstrapping. This will replace the contents of /etc/resolv.conf during the bootstrapping process and until it is changed back to localhost for consul.
A file location for a tgz file that stores an apt repo that needs to be added for development or testing of new repo features. This repo will be downloaded onto all compute instances of your build, and pinned at a higher precedence than your other package repos (specifically at 999).
The current workflow assumes that this file location is the URL of the
location of an artifact created by jenkins. It is assumed that apt
repos will be created by using the build_scripts/override_packages.sh
which builds a set of packages based on the repoconf repo
By default, builds assume that no override repo is being used.
Slack url to post consul alerts.
This is what I used for testing:
export BUILD\_NUMBER=test\_`date +"%d%m%y%H%M%S"`;export env=dan;export KEY\_NAME=openstack\_ssh\_key; export env\_file=~/credentials.sh;export timeout\_command=gtimeout; export puppet\_modules\_source\_repo=https://github.com/bodepd/puppet-rjil;export puppet\_modules\_source\_branch=origin/deploy\_script;export ssh\_user=ubuntu;bash -x build\_scripts/deploy.sh
The first time this script runs (or when it needs to update the version of jiocloud install), there are a few deps that need to be manually installed on Mac.
brew install python
pip install virtualenv
brew install libffi
brew install libyaml
brew install coreutils
alias timeout=gtimeout
It is possible that you will also have to manually update the include path to ensure that libffi is available:
export CFLAGS="-I/usr/local/opt/libffi/lib/libffi-3.0.13/include/"
In general, the process has been designed to be as unobtrusive and lenient as possible, and is split into the following:
https://github.com/jiocloud/puppet-rjil/pulls
The easiest way to make a contribution is to submit a pull request via github. It is not a requirement that an issue of task exist for a pull request to be merged (although it might be helpful for some situations). Contributions are merged as long as they:
- Are approved by at least one core team member
- Pass unit tests
- Pass integration tests (if they risk regressions to any of the systems under integration tests)
- Contain commit messages that provide enough context for reviewers to understand
-
- the motivation for the patch
-
- the desired outcome of the patch
https://github.com/jiocloud/puppet-rjil/issues
Github issues are used to track issues where there is not a patch immediately available. This might be because:
- The issue is not well understood/diagnosed
- The issue is non-trivial to fix
- The person opening the issue does not immediately know how to resolve it.
In general, at least a menial attempt to debug an issue should be attempted before opening an issue. There are standard procesures that anyone can use to attempt to categorize and provide context around a given failure so that they can open a useful (ie: actionable) as opposed to a useless issue.
Example of a useless issue:
Umm... dude, it doesn't work.
<enter random log spew here, or even worse, logs from jenkins, showing that you put
zero effort into it>
Issues like the above will result in a mild scolding followed by a request that you perform the debugging steps mentioned in this section.
https://rjil.slack.com/messages/deployment-team/
We use slack as the communication center for this project. Slack is a great place to ask a questions, or to collaborate on issues where pairing is desired. Events related to the development of the cloud platform are also streamed to slack in real time.
https://trello.com/b/PUXa12Cl/jiocloud-deployments
We use Trello to track tasks and task backlog related to this project. It is used to understand our tasks, how they fit into overall categories as well as who is currently working on what things. The process for how tasks are added into Trello is currently a bit adhoc, but needs to be enhanced as additional resources come online.
https://github.com/jiocloud/jiocloud-specs
A specification should be created in advance for large features.
Specifications are intended to allow more people to be involved in the design and consensus for larger feature sets. This becomes more important as more people get on boarded, especially as those resources intend to assist in the implementation of larger features sets.
-
login to one of the roles with a floating ip assigned (review your specified layout to be sure)
-
from there, run the following
*jorc get_failures --hosts*
This command will list three kinds of failures for each host. In order to debug each host, you generally need to log into the host that is in an error state.
NOTE: it is possible to login to each host by the same hostname returned by get_failures. It even autocompletes for you.
Indicate that the last puppet run did not complete successfully.
Node: XXXX, Check: puppet
Indicates the last puppet run has failing resources. Review /var/log/syslog to track down failures.
NOTE: all puppet lines from the logs contain puppet-user
NOTE: some errors from the logs are just cascading failures, you need to track these down to the root cause
Node: XXX, Check: validation
Validation checks are run to ensure that each service installed on a machine is in an active state before calling the failure a success. Validation failures indicate that while the configuration has been successfully applied, the service is not actually functional. This may indicate that it is still waiting for external dependencies or in the case of ceph, that it is still performing bootstrapping operations.
The output from each validation command can be viewed in /var/log/syslog of the machine whose validation is failing. It is also easy to run the validation commands yourself:
sudo run-parts --regex=. --verbose --exit-on-error --report /usr/lib/jiocloud/tests/
After seeing that a service is not in a functional state, you should check the logs for that individual service to see if there are any clues there.
Lists additional consul checks that are currently in the critical state
To see all consul services in the critical state:
curl http://localhost:8500/v1/health/state/critical?pretty=1
Each service mentioned here also contains the output from it's failed command.
- logging into the machines
- running jorc get_failures --hosts
-
- tracking down failures to root causes (in either /var/log/syslog or /var/log/cloud-init.log)
/var/log/cloud-init-output.log
we want to better support project devs.
here is what I envision:
- developers can check out their local code into the puppet-rjil environment
- set it up as a mount in vagrant (or rely on the fact it will be automounted into /vagrant/)
- customize the load path for whatever is using it