Background
Fedora Infrastructure is looking to setup a private cloud instance in 2012. This cloud instance will be used in a number of ways to benefit Fedora. We are continuing to evaluate a number of cloud technologies for the software side of this cloud.
Two Cloudlets
Our original setup was going to be a single eucalyptus cloud instance. However, when testing deployment, we determined it would be better to split our resources into 2 clouds. This will allow us to do things like upgrade or re-install one cloud while the other is running. Resources could be redirected/rebooted into the other cloud to allow one to be in downtime.
Software
When we evaluated software early in 2012, eucalyptus was the clear leader. However, later in 2012 things are not as clear, so we are investigating other cloud software to determine which we wish to go with. We may well decide on one for one cloudlet and another for the other one, depending on ongoing setup and maint costs.
Use cases
Doesn't need persistent storage
- Fedora QA may use instances with it's AutoQA setup. Instances would be created, tests run and destroyed. It's unknown how many instances we would need here.
- Chainbuilding / Kopers may use this cloud to build chains of packages that are not yet in Fedora and thus cannot be build via scratch builds in the existing buildsystem. These may also be used for spinning test live or install images by QA. This may be open to Fedora contributors or restricted to a subset such as packagers.
- Mass rebuilds of Fedora packages. This could be done for testing a new global rpm/package change, or to discover FTBFS (Fails to build from source) packages. This would use as many builders as we could easily spin up to reduce time for building all 10,000+ Fedora packages. Could use the chainbuilding setup as above as a scaffolding. Additionally, extra builder instances could be potentially used by the official build system during mass rebuilds to reduce rebuild time.
- Docs folks need to generate i18n versions of docs. This would require an instance, tools and a script running. Then data is synced off and the instance could be destroyed.
Needs persistent storage, but possibly can use a /mnt ed volume
- Test instances may be used for testing new tech or applications as a proof of concept before persuing a RFR. We currently have several publictest instances.
Needs persistent storage and snapshots
- Infrastructure Development hosts may be moved to this cloud. These instances could possibly be 'on demand' when development needs to take place. Currently we have about 8 development instances.
- Infrastructure Staging hosted may be moved to this cloud. Some of these may be 'always on' and some may be on demand. Currently we have about 13 of these instances.
- We may want to move some of our one-off instances that are outside phx2 into the cloud for easier management. Things like keyservers, unbound instances, listservers or hosted resources.
Further down the road:
- Instances for qa/packagers to test new packages or track down bugs.
- Instances for demos or events to show off Fedora.
For initial deployment, we would need to be able to run ~30 or so instances at a time with ability to grow rapidly above that for qa and building needs.
Dependencies
- Need a way to easily provision new instances with limited admin intervention. Looking at ansible for this task.
- Would like to be able to create images via kickstart and normal install/deployment methods if needed.
- Hardware needs to be ordered and installed.
- Public IP addresses need to be made available.
- Would be nice to get full EPEL packages to deploy with.
Setup / deployment
This hardware will be on the 'edge' of the network and not connected to the rest of Fedora Infrastructure except via external networks. This will allow us to us external ip's and make sure the cloud instance doesn't have access to anything in the regular Fedora Infrastructure. Storage will be on the local servers for caching with additional netapp space for images and data.
We have 8 physical servers for this deployment. 4 will be in each 'cloudlet'. One node will be a controller node with access to external IP's and the other 3 will be compute nodes.
Implementation overview / timelines
2012-04 - Hardware is being determined and finalized.
2012-07 - Initial hardware setup and install
2012-08 - Initial use cases gathered
2012-09 - Finish software evaluations, setup 'production' instances.
2012-10 - Announce availability and collect more use cases.
2012-11 - Evaluate load and expansion needs.
Policies
This section is currently under discussion. We need to setup clear policies on usage and access to the private cloud. In general we plan to open things to a small group of trusted contributors, take their feedback and usage and expand access out to larger groups as capacity and desire allows.
(This section is a DRAFT)
Users or groups that need rare one off images can simply request one via a ticket. Users or groups that often need instances will be granted accounts to spin up and down their own images.
Instances may be rebooted at any time. Save your data off often.
Persistent storage may be available as seperate volumes. Data retention and Quotas may be imposed on this data.
Default instance time to live would be a week.
Default network policy will allow only ports 80, 443, and 22 tcp.
Instances are assist in furthering the work related to the Fedora Project. Please don't use them for unrelated activities.
We reserve the right to shutdown, delete or revoke access to any instances at any time for any reason.
Eucalyptus Cloud Information
The Eucalyptus cloudlet is up and under limited testing now. If you are given an account this section will show some basic workflow and setup and example commands to help you getting started.
Obtaining an account
Currently we are under limited testing. Please ask in #fedora-admin on IRC if you are willing to help us test. Note that we are going to be pretty limited on testers for a while until we have more bugs worked out and policies in place. See above section.
First steps
Once you have an account issued to you, you will go to the web interface at: https://ec2.cloud.fedoraproject.org:8443/ and enter the account, user and password you were given. The interface will ask you to change your password, please do so and pick a nice long passphrase. After logging in, select your username at the top and from the pull down list, choose: "Download New credentials". This will download a .zip file to your local computer. Unpack this zip file and 'source eucarc' to setup your env.
Simple euca commands
Install the command line euca tools: 'yum install euca2ools'
Check what kinds of images are available: 'euca-describe-images'
Check what instances are running: 'euca-describe-instances'
Create a ssh keypair for use with instances: 'euca-add-keypair keyname > keyname.pub'
Create an instance with that key setup: 'euca-run-instances -k kevins ami-00000006' (the ami is the image ami from the describe images above)
Login to your instance: "ssh ec2-user@externalIP"
Terminate your instance: 'euca-terminate-instances i-NNNNN' (where this is from 'euca-describe-instances')
Many more commands at: http://cheat.errtheblog.com/s/euca2ools/1
Images
Currently we have available: Fedora 16, Fedora 17 and RHEL6.3
We should customize available images for the above use cases:
All images
Standardize on root or ec2user access.
Set hostname to ami name or something uniq and descriptive. Perhaps "fedoracloud-AMI" or something?
Sudo configured for the ec2user
git-core installed.
Infrastructure Dev Instances
Based on rhel6 image.
Should contain:
mod_wsgi, httpd, git-core, puppet, persistent volume mounted on /srv
QA images
TBD
Builder Images
For mockchain/kopers use. Should be limited to 24 hours.
Moving to "production"
This section is a checklist of things we need to do before we can consider either of the cloudlets "production". Once we move to production mode on them we will move to scheduling outages, try and keep instances running smoothly and just perform upgrades and maint on the cloudlets. We want to make sure before we do this that things are stable and processes are ready for users.
- SOP needs to be written for creating images. (whats in them, update policy, ssh keys policy, etc)
- Decide who gets a login to manage instances, and who can just request instances be made for them.
- Should we get this FAS managed and bind our 3 level permission to it.
- SOP on making an instance for a requestor (via infra ticket?)
- Decide on time limits or other resource limits per account/tennat. Setup initial accounts/tenants.
- OpenStack cloudlet needs ansible playbooks written to install/configure it.
OpenStack needs folsom testing performed. </stike> folsom installed now.
OpenStack needs vlan testing performed.vlans in use.
OpenStack needs non glusterfs testing done.current install has base fs.
Do we need to decide between euca and openstack? When?For now we are going to do both.
- We need monitoring added. nagios? Controllers down, nodes down, capacity issues, etc.
- We need reporting added. Note when instances are made, etc. Either logging to log02, or some seperate report for cloud-sysadmins. Possibly some export from the software about cpu/mem/disk, etc.
- Need to determine who has access to physical cloud machines. Repurpose sysadmin-cloud, setup fas and sudo?
- if you (and we should) add fas, why not configure it to only create shell account for people who is admin or sponsor from sysadmin-cloud. Which bring us back to 2nd point where approved people from sysadmin-cloud could have access to request instances to be made.
- Setup group that can run ansible against physical cloud machines for updates, etc. (see above question too)
- Consider a re-occuring maint window for reboots/updates... ie, tell everyone that every month we have a window to do so, save work before then?
- Figure out how to handle dns. Should we setup some kind of dyndns? Should we just leave it with generic dns? Should we ask for control of reverse dns?
- How do we back these systems up and what should we actually be backing up
- Figure out how to make some system be or seem to be persistent