opendev/system-config
Code Issues Proposed changes
c51521fffe
system-config/doc/source/sysadmin.rst
Ian Wienand 3d63b3b8a4 borg-backup-server: log prune output to file 
This saves prune output to a log file automatically.  Add a bit more
info on the process too.

Change-Id: I2607ddbc313dfebc122609af78bb5eed63906f6b
2021-08-04 14:47:50 +10:00

598 lines
23 KiB
ReStructuredText
Raw Blame History

:title: System Administration
.. _sysadmin:
System Administration
#####################
Our infrastructure is code and contributions to it are handled just
like the rest of OpenDev. This means that anyone can contribute to
the installation and long-running maintenance of systems without shell
access, and anyone who is interested can provide feedback and
collaborate on code reviews.
The configuration of every system operated by the infrastructure team
is managed by Ansible and driven by continuous integration and
deployment by Zuul.
https://opendev.org/opendev/system-config
All system configuration should be encoded in that repository so that
anyone may propose a change in the running configuration to Gerrit.
Guide to CI and CD
==================
All development work is based around Zuul jobs and a continuous
integration and development workflow.
The starting point for all services is generally the playbooks and
roles kept in :git_file:`playbooks`.
Most playbooks are named ``service-<name>.yaml`` and will indicate
which production areas they drive.
These playbooks run on groups of hosts which are defined in
:git_file:`inventory/service/groups`. The production hosts are kept
in an inventory at :git_file:`inventory/base/hosts.yaml`. During
testing, these same playbooks are run against the test nodes. You can
note that the testing hosts are given names that match the group
configuration in the jobs defined in
:git_file:`zuul.d/system-config-run.yaml`.
Deployment is run through a bastion host ``bridge.openstack.org``.
After changes are approved, Zuul will run Ansible on this host; which
will then connect to the production hosts and run the orchestration
using the latest committed code. The bridge is a special host because
it holds production secrets, such as passwords or API keys, and
unredacted logs. As many logs as possible are provided in the public
Zuul job results, but they need to be audited to ensure they do not
leak secrets and thus in some cases may not be published.
For CI testing, each job creates a "fake" bridge, along with the
servers required for orchestration. Thus CI testing is performed by a
"nested" Ansible -- Zuul initially connects to the testing bridge node
and deploys it, and then this node runs its own Ansible that tests the
orchestration to the other testing nodes, simulating the production
environment. This is driven by playbooks kept in
:git_file:`playbooks/zuul`. Here you will also find testing
definitions of host variables that are kept secret for production
hosts.
After the test environment is orchestrated, the
`testinfra <https://testinfra.readthedocs.io/en/latest/>`__ tests from
:git_file:`testinfra` are run. This validates the complete
orchestration testing environment; things such as ensuring user
creation, container readiness and service wellness checks are all
performed.
.. _adding_new_server:
Adding a New Server
===================
Creating a new server for your service requires discussion with the
OpenDev administrators to ensure donor resources are being used
effectively.
* Hosts should only be configured by Ansible. Nonetheless, in some
cases SSH access can be granted. Add your public key to
:git_file:`inventory/base/group_vars/all.yaml` and include a stanza
like this in your server ``host_vars``::
extra_users:
- your_user_name
* Add an RST file with documentation about the server and services in
:git_file:`doc/source` and add it to the index in that directory.
SSH Access
==========
For any of the systems managed by the OpenDev Infrastructure team, the
following practices must be observed for SSH access:
* SSH access is only permitted with SSH public/private key
authentication.
* Users must use a strong passphrase to protect their private key. A
passphrase of several words, at least one of which is not in a
dictionary is advised, or a random string of at least 16
characters.
* To mitigate the inconvenience of using a long passphrase, users may
want to use an SSH agent so that the passphrase is only requested
once per desktop session.
* Users private keys must never be stored anywhere except their own
workstation(s). In particular, they must never be stored on any
remote server.
* If users need to 'hop' from a server or bastion host to another
machine, they must not copy a private key to the intermediate
machine (see above). Instead SSH agent forwarding may be used.
However due to the potential for a compromised intermediate machine
to ask the agent to sign requests without the users knowledge, in
this case only an SSH agent that interactively prompts the user
each time a signing request (ie, ssh-agent, but not gnome-keyring)
is received should be used, and the SSH keys should be added with
the confirmation constraint ('ssh-add -c').
* The number of SSH keys that are configured to permit access to
OpenDev machines should be kept to a minimum.
* OpenDev Infrastructure machines must use Ansible to centrally manage
and configure user accounts, and the SSH authorized_keys files from
the opendev/system-config repository.
* SSH keys should be periodically rotated (at least once per year).
During rotation, a new key can be added to puppet for a time, and
then the old one removed.
Gerrit Admins
=============
To provide a reasonable firewall from outside authentication systems,
Gerrit administrators keep two accounts: one for normal code review
activity and one for performing Gerrit administration. Following the same
pattern as our Kerberos administrator account logins, the admin account
corresponding to ``$USER`` would be ``$USER.admin`` (Gerrit doesn't allow
``/`` in usernames) so they can be easily identified when auditing
activity. Unlike the normal code review account, the admin account should
have no OpenID so that it is only accessable by API/CLI methods so they
cannot be compromised at the third-party ID provider.
To create a personal Gerrit admin account from a shell on the server, run
the following command::
sudo -u gerrit2 ssh -i ~gerrit2/review_site/etc/ssh_host_rsa_key \
-p 29418 -l 'Gerrit Code Review' localhost \
"suexec --as openstack-project-creator -- \
gerrit create-account --group Administrators --full-name myname.admin \
--ssh-key 'ssh-rsa AAAA...BCDE myname@computer' myname.admin"
We ``suexec`` as the ``openstack-project-creator`` account because the
magic ``Gerrit Code Review`` pseudoaccount can't set group memberships so
we need to run that command as a user which is already in the
``Administrators`` group. With an account like this, routine actions like
populating new groups with initial members is still quite simple::
ssh -p 29418 myname.admin@review.opendev.org \
"gerrit set-members some-new-group --add somebody@example.org"
Another common example is bypassing Zuul to submit a change for merging
directly to a project. In this case we must first add our account to
another group which has permission to set the relevant labels (it doesn't
get that simply by being an administrator), and then do the
commenting/voting/submitting, followed by cleaning up the extra group
membership again at the end::
ssh -p 29418 myname.admin@review.opendev.org \
"gerrit set-members 'Project Bootstrappers' --add myname.admin"
ssh -p 29418 myname.admin@review.opendev.org \
"gerrit review 12345,6 --message 'Bypassing Zuul to merge this.' \
--code-review=2 --verified=2 --label workflow=1 --submit"
ssh -p 29418 myname.admin@review.opendev.org \
"gerrit set-members 'Project Bootstrappers' --remove myname.admin"
Note that it's possible to temporarily add your normal OpenID-associated
WebUI account to the ``Administrators`` group or other groups with similar
superuser permissions like ``Project Bootstrappers``, but keep in mind that
an attacker who has quietly gained control of your account at the OpenID
provider could be waiting for that opportunity to take advantage of the
added permissions, or you may simply forget to remove the account afterward
negating the added safety of this account separation.
For more examples, see the detailed documentation for Gerrit's SSH CLI,
available on our server:
https://review.opendev.org/Documentation/cmd-index.html
GitHub Access
=============
To ensure that code review and testing are not bypassed in the public
Git repositories, only Gerrit will be permitted to commit code to
OpenDev repositories. Because GitHub always allows project
administrators to commit code, accounts that have access to manage the
GitHub projects necessarily will have commit access to the
repositories.
A shared Github administrative account is available (credentials
stored in the global authentication location). If administrators
would prefer to keep a separate account, it can be added to the
organisation after discussion and noting the caveats around elevated
access. The account must have 2FA enabled.
In either case, the administrator accounts should not be used to check
out or commit code for any project.
Note that it is unlikely to be useful to use an account also used for
active development, as you will be subscribed to many notifications
for all projects.
Root only information
#####################
Below is information relevant to members of the core team with root
access.
Accessing Clouds
================
As an unprivileged user who is a member of the `sudo` group on bridge,
you can inspect any of the clouds with::
sudo openstack --os-cloud <cloud name> --os-cloud-region <region name>
Backups
=======
Infra uses the `borg <https://borgbackup.readthedocs.io>`__ backup
tool.
Hosts in the ``borg-backup`` Ansible inventory group will be backed up
to servers in the ``borg-backup-server`` group with ``borg``. The
``playbooks/roles/borg-backup`` and
``playbooks/roles/borg-backup-server`` roles implement the required
setup.
The backup server has a unique Unix user for each host to be backed
up. The roles will setup required users, their home directories in
the backup volume and relevant ``authorized_keys``.
Host backup happens via a daily cron job (managed by Ansible) on each
individual host to be backed up. The host to be backed up initiates
the backup process to the remote backup server(s) using a separate ssh
key setup just for backup communication (see ``/root/.ssh/config``).
Setting up hosts for backup
---------------------------
To setup a host for backup, put it in the ``borg-backup`` group.
Hosts can specify ``borg_backup_excludes_extra`` and
``borg_backup_dirs_extra`` to exclude or include specific directories
as required (see role documentation for more details).
``borg`` splits backup data into chunks and de-duplicates as much as
possible. For backing up large items, particularly things like
database dumps, we want to give ``borg`` as much chance to
de-duplicate as possible. Approaches such as dumping to compressed
files on disk defeat de-duplication because all the data changes for
each dump.
For dumping large data, hosts should put a file into
``/etc/borg-streams`` that performs the dump in an uncompressed manner
to stdout. The backup scripts will create a separate archive for each
stream defined here. For more details, see the ``backup`` role
documentation. These streams should attempt to be as friendly to
de-duplication as possible; see some of the examples of ``mysqldump``
to find arguments that help keep the output data more stable (and
hence more easily de-duplicated).
Restore from Backup
-------------------
Hosts have ``/usr/local/bin/borg-mount`` (specify one of the backup
servers as an argument) that will mount the backups to
``/opt/backups`` via FUSE.
``borg`` has other options for restoring. If you need to extract on
the backup server itself, a basic way to dump a host at a particular
time is to
* log into the backup server
* sudo ``su -`` to switch to the backup user for the host to be restored
* you will now be in the home directory of that user
* run ``/opt/borg/bin/borg list ./backup`` to list the archives available
* these should look like ``<hostname>-<stream>-YYYY-MM-DDTHH:MM:SS``
* move to working directory
* extract one of the appropriate archives with ``/opt/borg/bin/borg extract ~/backup <archive-tag>``
Managing backup storage
-----------------------
We run ``borg`` in append-only mode. This means clients can not
remove old backups on the server.
However, due to the way borg works, append-only mode plays all client
transactions into a transaction log until a read-write operation
occurs. Examining the repository will appear to have all these
transactions applied (e.g. pruned archives will not appear; even if
they have not actually been pruned from disk). If you have reason to
not trust the state of the backup, you should *not* run any read-write
operations. You will need to manually examine the transaction log and
roll-back to a known good state; see
`<https://borgbackup.readthedocs.io/en/stable/usage/notes.html#append-only-mode>`__.
However, we have limited backup space. Each backup server has a
script ``/usr/local/bin/prune-borg-backups`` which can be run to
reclaim space. This should be run in a ``screen`` instance as it can
take a considerable time. It will prompt when run; you can confirm
the process with a ``noop`` run; confirming the prune will log the
output to ``/opt/backups``. This will keep the last 7 days of backups,
then monthly backups for 1 year and yearly backups for each archive.
The backup servers will send a warning when backup volume usage is
high, at which point this can be run manually.
.. _force-merging-a-change:
Force-Merging a Change
======================
Occasionally it is necessary to bypass the CI system and merge a
change directly. Usually, this is only required if we have a hole in
our testing of the CI or related systems themselves and have merged a
change which causes them to be unable to operate normally and
therefore unable to merge a reversion of the problematic change. In
these cases, use the following procedure to force-merge a change.
* Add yourself to the *Project Bootstrappers* group in Gerrit.
* Navigate to the change which needs to be merged and reload the page.
* Remove any -2 votes on the change.
* Add +2 Code-Review, and +1 Workflow votes if necessary, then add +2
Verified. Also leave a review comment briefly explaining why this
was necessary, and make sure to mention it in the #opendev
IRC channel (ideally as a #status log entry for the benefit of
those not paying close attention to scrollback).
* At this point, a *Submit* Button should appear, click it. The
change should now be merged.
* Remove yourself from *Project Bootstrappers*
This procedure is the safest way to force-merge a change, ensuring
that all of the normal steps that Gerrit performs on repos still
happen.
Launching New Servers
=====================
New servers are launched using the ``launch/launch-node.py`` tool from the git
repository ``https://opendev.org/opendev/system-config``. This
tool is run from a checkout on the bridge - please see :git_file:`launch/README.rst`
for detailed instructions.
.. _disable-enable-ansible:
Disable/Enable Ansible
======================
You should normally not make manual changes to servers, but instead,
make changes through ansible or puppet. However, under some circumstances,
you may need to temporarily make a manual change to a managed
resource on a server.
OpenDev uses a Static Inventory in Ansible to control execution of Ansible
on hosts. A full understanding
of the concepts in
`Ansible Inventory Introduction
<http://docs.ansible.com/ansible/intro_inventory.html>`_
is essential for being able to make informed decisions about actions
to take.
In the case of needing to disable the running of ansible or puppet on a node,
it's a simple matter of adding an entry to the ansible inventory "disabled" group
in :git_file:`inventory/groups.yaml`. The
disabled entry is an input to `ansible --list-hosts` so you can check your
entry simply by running it with `ansible $hostlist --list-hosts` as root
on the bridge host and ensuring that the list of hosts returned is as
expected. Globs, group names and server UUIDs should all be acceptable input.
If you need to disable a host immediately without waiting for a patch to land
to `system-config`, there is a file on the bridge host,
`/etc/ansible/hosts/emergency.yaml` that can be edited directly.
`/etc/ansible/hosts/emergency.yaml` is a file that should normally be empty,
but the contents are not managed by ansible. It's purpose is to allow for
disabling ansible at times when landing a change to the ansible repo would be
either unreasonable or impossible.
Disabling puppet via ansible inventory does not disable puppet from being
able to be run directly on the host, it merely prevents ansible from
attempting to run it during the regular zuul jobs. If you choose to run
puppet manually on a host, take care to ensure that it has not been disabled
at the bridge level first.
If you need to pause all execution of ansible playbooks by Zuul you can
run the utility script ``disable-ansible``. The script touches the file
``/home/zuul/DISABLE-ANSIBLE`` on bridge.openstack.org. Doing
this forces the Zuul jobs that run ansible for us to wait until that file is
removed. This acts like a global pause. The script exists to prevent admins
from misspelling the name of the file and is recommended.
Examples
--------
To disable an OpenDev instance called `foo.opendev.org` temporarily,
ensure the following is in `/etc/ansible/hosts/emergency.yaml`
::
# Please add an inline comment so we know who added the host and why
plugin: yamlgroup
groups:
disabled:
- foo.opendev.org # 2020-05-23 bob is testing change 654321
Ad-hoc Ansible runs
===================
If you need to run Ansible manually against a host, you should
* disable automated Ansible runs following the section above
* ``su`` to the ``zuul`` user and run the playbook with something like
``ansible-playbook -vv
src/opendev.org/opendev/system-config/playbooks/service-<name>.yaml``
* Restore automated ansible runs.
* You can also use the ``--limit`` flag to restrict which hosts run
when there are many in a group. However, be aware that some
roles/playbooks like ``letsencrypt`` and ``backup`` run across
multiple hosts (deploying DNS records or authorization keys), so
incorrect ``--limit`` flags could cause further failures.
.. _cinder:
Cinder Volume Management
========================
Adding a New Device
-------------------
If the main volume group doesn't have enough space for what you want
to do, this is how you can add a new volume.
Log into bridge.openstack.org and run::
export OS_CLOUD=openstackci-rax
export OS_REGION_NAME=DFW
openstack server list
openstack volume list
Change the variables to use a different environment. ORD for example::
export OS_CLOUD=openstackci-rax
export OS_REGION_NAME=ORD
* Add a new 1024G cinder volume (substitute the hostname and the next number
in series for NN)::
openstack volume create --size 1024 "$HOSTNAME.ord.openstack.org/mainNN"
openstack server add volume "HOSTNAME.openstack.org" "HOSTNAME.openstack.org/mainNN"
* or to add a 100G SSD volume::
openstack volume create --type SSD --size 100 "HOSTNAME.openstack.org/mainNN"
openstack server add volume "HOSTNAME.openstack.org" "HOSTNAME.openstack.org/mainNN"
* Then, on the host, create the partition table::
DEVICE=/dev/xvdX
sudo parted $DEVICE mklabel msdos mkpart primary 0% 100% set 1 lvm on
sudo pvcreate ${DEVICE}1
* It should show up in pvs::
$ sudo pvs
PV VG Fmt Attr PSize PFree
/dev/xvdX1 lvm2 a- 1024.00g 1024.00g
* Add it to the main volume group::
sudo vgextend main ${DEVICE}1
* However, if the volume group does not exist yet, you can create it::
sudo vgcreate main ${DEVICE}1
Creating a New Logical Volume
-----------------------------
Make sure there is enough space in the volume group::
$ sudo vgs
VG #PV #LV #SN Attr VSize VFree
main 4 2 0 wz--n- 2.00t 347.98g
If not, see `Adding a New Device`_.
Create the new logical volume and initialize the filesystem::
NAME=newvolumename
sudo lvcreate -L1500GB -n $NAME main
sudo mkfs.ext4 -m 0 -j -L $NAME /dev/main/$NAME
sudo tune2fs -i 0 -c 0 /dev/main/$NAME
Be sure to add it to ``/etc/fstab``.
Expanding an Existing Logical Volume
------------------------------------
Make sure there is enough space in the volume group::
$ sudo vgs
VG #PV #LV #SN Attr VSize VFree
main 4 2 0 wz--n- 2.00t 347.98g
If not, see `Adding a New Device`_.
The following example increases the size of a volume by 100G::
NAME=volumename
sudo lvextend -L+100G /dev/main/$NAME
sudo resize2fs /dev/main/$NAME
The following example increases the size of a volume to the maximum allowable::
NAME=volumename
sudo lvextend -l +100%FREE /dev/main/$NAME
sudo resize2fs /dev/main/$NAME
Replace an Existing Device
--------------------------
We generally need to do this if our cloud provider is planning maintenance to a
volume. We usually get a few days heads up on maintenance window, so depending
on the size of the volume, it may take some time to replace.
First thing to do is add the replacement device to the server, see
`Adding a New Device`_. Be sure the replacement volume is the same type / size
as the existing.
If the step above were followed, you should see something like::
$ sudo pvs
PV VG Fmt Attr PSize PFree
/dev/xvdb1 main lvm2 a-- 50.00g 0
/dev/xvdc1 main lvm2 a-- 50.00g 50.00g
Be sure both devices are in the same VG (volume group), if not you did not
properly extend the device.
.. note::
Be sure to use a screen session for the following step!
Next is to move the data from once device to another::
$ sudo pvmove /dev/xvdb1 /dev/xvdc1
/dev/xvdb1: Moved: 0.0%
/dev/xvdb1: Moved: 1.8%
...
...
/dev/xvdb1: Moved: 99.4%
/dev/xvdb1: Moved: 100.0%
Confirm all the data was moved, and the original device is empty (PFree)::
$ sudo pvs
PV VG Fmt Attr PSize PFree
/dev/xvdb1 main lvm2 a-- 50.00g 50.00g
/dev/xvdc1 main lvm2 a-- 50.00g 0
And remove the device from the main volume group::
$ sudo vgreduce main /dev/xvdb1
Removed "/dev/xvdb1" from volume group "main"
To be safe, we can also wipe the label from LVM::
$ sudo pvremove /dev/xvdb1
Labels on physical volume "/dev/xvdb1" successfully wiped
Leaving us with just a single device::
$ sudo pvs
PV VG Fmt Attr PSize PFree
/dev/xvdc1 main lvm2 a-- 50.00g 0
At this time, you are able to remove the original volume from openstack if
no longer needed.
Email
=====
There is a shared email account used for Infrastructure related mail
(account sign-ups, support tickets, etc). Root admins should ensure
they have access to this account; access credentials are available
from any existing member.
View Git Blame Copy Permalink