Jiri Stransky 653683981c Major Upgrades Including Operating System Upgrade

Co-Authored-By: Sofer Athlan Guyot <sathlang@redhat.com>
Change-Id: I7f2dc6b4ad301e7c5a9a041c870cf20f054f3614
Implements: blueprint upgrades-with-os

2019-02-07 16:45:29 +01:00

26 KiB

Raw Blame History

Major Upgrades Including Operating System Upgrade

https://blueprints.launchpad.net/tripleo/+spec/upgrades-with-os

Note

Abbreviation "OS" in this spec stands for "operating system", not "OpenStack".

So far all our update and upgrade workflows included doing minor operating system updates (essentially a yum update) on the machines managed by TripleO. This will need to change as we can't stay on a single OS release indefinitely -- we'll need to perform a major OS upgrade. The intention is for the TripleO tooling to help with the OS upgrade significantly, rather than leaving this task entirely to the operator.

Problem Description

We need to upgrade undercloud and overcloud machines to a new release of the operating system.

We would like to provide an upgrade procedure both for environments where Nova and Ironic are managing the overcloud servers, and "Deployed Server" environments where we don't have control over provisioning.

Further constraints are imposed by Pacemaker clusters: Pacemaker is non-containerized, so it is upgraded via packages together with the OS. While Pacemaker would be capable of a rolling upgrade, Corosync also changes major version, and starts to rely on knet for the link protocol layer, which is incompatible with previous version of Corosync. This introduces additional complexity: we can't do OS upgrade in a rolling fashion naively on machines which belong to the Pacemaker cluster (controllers).

Proposed Change - High Level View

The Pacemaker constraints will be addressed by performing a one-by-one (though not rolling) controller upgrade -- temporarily switching to a single-controller cluster on the new OS, and gradually upgrading the rest. This will also require implementation of persistent OpenStack data transfer from older to newer OS releases (to preserve uptime and for easier recoverability in case of failure).

We will also need to ensure that at least 2 ceph-mon services run at all times, so ceph-mon services will keep running even after we switch off Pacemaker and OpenStack on the 2 older controllers.

We should scope two upgrade approaches: full reprovisioning, and in-place upgrade via an upgrade tool. Each come with different benefits and drawbacks. The proposed CLI workflows should ideally be generic enough to allow picking the final preferred approach of overcloud upgrade late in the release cycle.

While the overcloud approach is still wide open, undercloud seems to favor an in-place upgrade due to not having a natural place to persist the data during reprovisioning (e.g. we can't assume overcloud contains Swift services), but that could be overcome by making the procedure somewhat more manual and shifting some tasks onto the operator.

The most viable way of achieving an in-place (no reprovisioning) operating system upgrade currently seems to be Leapp, "an app modernization framework", which should include in-place upgrade capabilites.

Points in favor of in-place upgrade:

While some data will need to be persisted and restored regardless of approach taken (to allow safe one-by-one upgrade), reprovisioning may also require managing data which would otherwise persist on its own during an in-place upgrade.
In-place upgrade allows using the same approach for Nova+Ironic and Deployed Server environments. If we go with reprovisioning, on Deployed Server environments the operator will have to reprovision using their own tooling.
Environments with a single controller will need different DB mangling procedure. Instead of system_upgrade_transfer_data step below, their DB data will be included into the persist/restore operations when reprovisioning the controller.

Points in favor of reprovisioning:

Not having to integrate with external in-place upgrade tool. E.g. in case of CentOS, there's currently not much info available about in-place upgrade capabilities.
Allows to make changes which wouldn't otherwise be possible, e.g. changing a filesystem.
Reprovisioning brings nodes to a clean state. Machines which are continuously upgraded without reprovisioining can potentially accumulate unwanted artifacts, resulting in increased number of problems/bugs which only appear after an upgrade, but not on fresh deployments.

Proposed Change - Operator Workflow View

The following is an example of expected upgrade workflow in a deployment with roles: ControllerOpenstack, Database, Messaging, Networker, Compute, CephStorage. It's formulated in a documentation-like manner so that we can best imagine how this is going to work from operator's point of view.

Upgrading the Undercloud

The in-place undercloud upgrade using Leapp will likely consist of the following steps. First, prepare for OS upgrade via Leapp, downloading the necessary packages:

leapp upgrade

Then reboot, which will upgrade the OS:

reboot

Then run the undercloud upgrade, which will bring back the undercloud services (using the newer OpenStack release):

openstack tripleo container image prepare default \
    --output-env-file containers-prepare-parameter.yaml
openstack undercloud upgrade

If we wanted or needed to upgrade the undercloud via reprovisioning, we would use a backup and restore procedure as currently documented, with restore perhaps being utilized just partially.

Upgrading the Overcloud

Update the Heat stack, generate Heat outputs for building upgrade playbooks:
```
openstack overcloud upgrade prepare <DEPLOY ARGS>
```
Notes:
- Among the <DEPLOY ARGS> should be containers-prepare-parameter.yaml bringing in the containers of newer OpenStack release.
Prepare an OS upgrade on one machine from each of the "schema-/cluster-sensitive" roles:
```
openstack overcloud upgrade run \
    --tags system_upgrade_prepare \
    --limit controller-openstack-0,database-0,messaging-0
```
Notes:
- This stops all services on the nodes selected.
- For external installers like Ceph, we'll have a similar external-upgrade command, which can e.g. remove the nodes from the Ceph cluster:
```
openstack overcloud external-upgrade run \
    --tags system_upgrade_prepare \
    -e system_upgrade_nodes=controller-openstack-0,database-0,messaging-0
```
- If we use in-place upgrade:
  - This will run the leapp upgrade command. It should use newer OS and newer OpenStack repos to download packages, and leave the node ready to reboot into the upgrade process.
  - Caution: Any reboot after this is done on a particular node will cause that node to automatically upgrade to newer OS.
- If we reprovision:
  - This should persist node's important data to the undercloud. (Only node-specific data. It would not include e.g. MariaDB database content, which would later be transferred from one of the other controllers instead.)
  - Services can export their upgrade_tasks to do the persistence, we should provide an Ansible module or role to make it DRY.
Upload new overcloud base image:
```
openstack overcloud image upload --update-existing \
    --image-path /home/stack/new-images
```
Notes:
- For Nova+Ironic environments only. After this step any new or reprovisioned nodes will receive the new OS.
Run an OS upgrade on one node from each of the "schema-/cluster-sensitive" roles or reprovision those nodes.

Only if we do reprovisioning:
```
openstack server rebuild controller-openstack-0
openstack server rebuild database-0
openstack server rebuild messaging-0

openstack overcloud admin authorize \
    --overcloud-ssh-user <user> \
    --overcloud-ssh-key <path-to-key> \
    --overcloud-ssh-network <ssh-network> \
    --limit controller-openstack-0,database-0,messaging-0
```
Both reprovisioning and in-place:
```
openstack overcloud upgrade run \
    --tags system_upgrade_run \
    --limit controller-openstack-0,database-0,messaging-0
```
Notes:
- This step either performs a reboot of the nodes and lets Leapp upgrade them to newer OS, or reimages the nodes with a fresh new OS image. After they come up, they'll have newer OS but no services running. The nodes can be checked before continuing.
- In case of reprovisioning:
  - The overcloud admin authorize will ensure existence of tripleo-admin user and authorize Mistral's ssh keys for connection to the newly provisioned nodes. The --overcloud-ssh-* work the same as for overcloud deploy.
  - The --tags system_upgrade_run is still necessary because it will restore the node-specific data from the undercloud.
  - Services can export their upgrade_tasks to do the restoration, we should provide an Ansible module or role to make it DRY.
- Ceph-mon count is reduced by 1 (from 3 to 2 in most environments).
- Caution: This will have bad consequences if run by accident on unintended nodes, e.g. on all nodes in a single role. If possible, it should refuse to run if --limit is not specified. If possible further, it should refuse to run if a full role is included, rather than individual nodes.
Stop services on older OS and transfer data to newer OS:
```
openstack overcloud external-upgrade run \
    --tags system_upgrade_transfer_data \
    --limit ControllerOpenstack,Database,Messaging
```
Notes:
- This is where control plane downtime starts.
- Here we should:
  - Detect which nodes are on older OS and which are on newer OS.
  - Fail if we don't find at least one older OS and exactly one newer OS node in each role.
  - On older OS nodes, stop all services except ceph-mon. (On newer node, no services are running yet.)
  - Transfer data from an older OS node (simply the first one in the list we detect, or do we need to be more specific?) to the newer OS node in a role. This is probably only going to do anything on the Database role which includes DBs, and will be a no-op for others.
  - Services can export their external_upgrade_tasks for the persist/restore operations, we'll provide an Ansible module or role to make it DRY. The transfer will likely go via undercloud initially, but it would be nice to make it direct in order to speed it up.
Run the usual upgrade tasks on the newer OS nodes:
```
openstack overcloud upgrade run \
    --limit controller-openstack-0,database-0,messaging-0
```
Notes:
- Control plane downtime stops at the end of this step. This means the control plane downtime spans two commands. We should not make it one command because the commands use different parts of upgrade framework underneath, and the separation will mean easier re-running of individual parts, should they fail.
- Here we start pcmk cluster and all services on the newer OS nodes, using the data previously transferred from the older OS nodes.
- Likely we won't need any special per-service upgrade tasks, unless we discover we need some data conversions or adjustments. The node will be with all services stopped after upgrade to newer OS, so likely we'll be effectively "setting up a fresh cloud on pre-existing data".
- Caution: At this point the newer OS nodes became the authority on data state. Do not re-run the previous data transfer step after services have started on newer OS nodes.
- (Currently upgrade run has --nodes and --roles which both function the same, as Ansible --limit. Notably, nothing stops you from passing role names to --nodes and vice versa. Maybe it's time to retire those two and implement --limit to match the concept from Ansible closely.)
Perform any service-specific && node-specific external upgrades, most importantly Ceph:
```
openstack overcloud external-upgrade run \
    --tags system_upgrade_run \
    -e system_upgrade_nodes=controller-openstack-0,database-0,messaging-0
```
Notes:
- Ceph-ansible here runs on a single node and spawns a new version of ceph-mon. Per-node run capability will need to be added to ceph-ansible.
- Ceph-mon count is restored here (in most environments, it means going from 2 to 3).
Upgrade the remaining control plane nodes. Perform all the previous control plane upgrade steps for the remaining controllers too. Two important notes here:
- Do not run the ``system_upgrade_transfer_data`` step anymore. The remaining controllers are expected to join the cluster and sync the database data from the primary controller via DB replication mechanism, no explicit data transfer should be necessary.
- To have the necessary number of ceph-mons running at any given time (often that means 2 out of 3), the controllers (ceph-mon nodes) should be upgraded one-by-one.
After this step is finished, all of the nodes which are sensitive to Pacemaker version or DB schema version should be upgraded to newer OS, newer OpenStack, and newer ceph-mons.

Upgrade the rest of the overcloud nodes (Compute, Networker, CephStorage), either one-by-one or in batches, depending on uptime requirements of particular nodes. E.g. for computes this would mean evacuating and then also running:

openstack overcloud upgrade run \
    --tags system_upgrade_prepare \
    --limit novacompute-0

openstack overcloud upgrade run \
    --tags system_upgrade_run \
    --limit novacompute-0

openstack overcloud upgrade run \
    --limit novacompute-0

Notes:

Ceph OSDs can be removed by the external-upgrade run --tags system_upgrade_prepare step before reprovisioning, and after upgrade run command, ceph-ansible can recreate the OSD via the external-upgrade run --tags system_upgrade_run step, always limited to the OSD being upgraded:

# Remove OSD
openstack overcloud external-upgrade run \
    --tags system_upgrade_prepare \
    -e system_upgrade_nodes=novacompute-0

# <<Here the node is reprovisioned and upgraded>>

# Re-deploy OSD
openstack overcloud external-upgrade run \
    --tags system_upgrade_run \
    -e system_upgrade_nodes=novacompute-0

Perform online upgrade (online data migrations) after all nodes have been upgraded:
```
openstack overcloud external-upgrade run \
    --tags online_upgrade
```
Perfrom upgrade converge to re-assert the overcloud state:
```
openstack overcloud upgrade converge <DEPLOY ARGS>
```

Clean up upgrade data persisted on undercloud:

openstack overcloud external-upgrade run \
    --tags system_upgrade_cleanup

Additional notes on data persist/restore

There are two different use cases:
- Persistence for things that need to survive reprovisioning (for each node)
- Transfer of DB data from node to node (just once to bootstrap the first new OS node in a role)
The synchronize Ansible module shipped with Ansible seems fitting, we could wrap it in a role to handle common logic, and execute the role via include_role from upgrade_tasks.
We would persist the temporary data on the undercloud under a directory accessible only by the user which runs the upgrade playbooks (mistral user). The root dir could be /var/lib/tripleo-upgrade and underneath would be subdirs for individual nodes, and one more subdir level for services.
- (Undercloud's Swift also comes to mind as a potential place for storage. However, it would probably add more complexity than benefit.)
The data persist/restore operations within the upgrade do not supplement or replace backup/restore procedures which should be performed by the operator, especially before upgrading. The automated data persistence is solely for upgrade purposes, not for disaster recovery.

Alternatives

Parallel cloud migration. We could declare the in-place upgrade of operating system + OpenStack as too risky and complex and time consuming, and recommend standing up a new cloud and transferring content to it. However, this brings its own set of challenges.

This option is already available for anyone whose environment is constrained such that normal upgrade procedure is not realistic, e.g. in case of extreme uptime requirements or extreme risk-aversion environments.

Implementing parallel cloud migration is probably best handled on a per-environment basis, and TripleO doesn't provide any automation in this area.
Upgrading the operating system separately from OpenStack. This would simplify things on several fronts, but separating the operating system upgrade while preserving uptime (i.e. upgrading the OS in a rolling fashion node-by-node) currently seems not realistic due to:
- The pacemaker cluster (corosync) limitations mentioned earlier. We would have to containerize Pacemaker (even if just ad-hoc non-productized image).
- Either we'd have to make OpenStack (and dependencies) compatible with OS releases in a way we currently do not intend, or at least ensure such compatibility when running containerized. E.g. for data transfer, we could then probably use Galera native replication.
- OS release differences might be too large. E.g. in case of differing container runtimes, we might have to make t-h-t be able to deploy on two runtimes within one deployment.
Upgrading all control plane nodes at the same time as we've been doing so far. This is not entirely impossible, but rebooting all controllers at the same time to do the upgrade could mean total ceph-mon unavailability. Also given that the upgraded nodes are unreachable via ssh for some time, should something go wrong and the nodes got stuck in that state, it could be difficult to recover back into a working cloud.

This is probably not realistic, mainly due to concerns around Ceph mon availability and risk of bricking the cloud.

Security Impact

How we transfer data from older OS machines to newer OS machines is a potential security concern.
The same security concern applies for per-node data persist/restore procedure in case we go with reprovisioning.
The stored data may include overcloud node's secrets and should be cleaned up from the undercloud when no longer needed.
In case of using the synchronize Ansible module: it uses rsync over ssh, and we would store any data on undercloud in a directory only accessible by the same user which runs the upgrade playbooks (mistral). This undercloud user has full control over overcloud already, via ssh keys authorized for all management operations, so this should not constitute a significant expansion of mistral user's knowledge/capabilities.

Upgrade Impact

The upgrade procedure is riskier and more complex.
- More things can potentially go wrong.
- It will take more time to complete, both manually and automatically.
Given that we upgrade one of the controllers while the other two are still running, the control plane services downtime could be slightly shorter than before.
When control plane services are stopped on older OS machines and running on newer OS machine, we create a window without high availability.
Upgrade framework might need some tweaks but on high level it seems we'll be able to fit the workflow into it.
All the upgrade steps should be idempotent, rerunnable and recoverable as much as we can make them so.

Other End User Impact

Floating IP availability could be affected. Neutron upgrade procedure typically doesn't immediately restart sidecar containers of L3 agent. Restarting will be a must if we upgrade the OS.

Performance Impact

When control plane services are stopped on older OS machines and running on newer OS machine, only one controller is available to serve all control plane requests.
Depending on role/service composition of the overcloud, the reduced throughput could also affect tenant traffic, not just control plane APIs.

Other Deployer Impact

Automating such procedure introduces some code which had better not be executed by accident. The external upgrade tasks which are tagged system_upgrade_* should also be tagged never, so that they only run when explicitly requested.
For the data transfer step specifically, we may also introduce a safety "flag file" on the target overcloud node, which would prevent re-running of the data transfer until the file is manually removed.

Developer Impact

Developers who work on specific composable services in TripleO will need to get familiar with the new upgrade workflow.

Main Risks

Leapp has been somewhat explored but its viability/readiness for our purpose is still not 100% certain.
CI testing will be difficult, if we go with Leapp it might be impossible (more below).
Time required to implement everything may not fit within the release cycle.
We have some idea how to do the data persist/restore/transfer parts, but some prototyping needs to be done there to gain confidence.
We don't know exactly what data needs to be persisted during reprovisioning.

Implementation

Assignee(s)

Primary assignees::: jistr, chem, jfrancoa
Other contributors::: fultonj for Ceph

Work Items

With aditional info in format: (how much do we know about this task, estimate of implementation difficulty).

(semi-known, est. as medium) Change tripleo-heat-templates + puppet-tripleo to be able to set up a cluster on just one controller (with newer OS) while the Heat stack knows about all controllers. This is currently not possible.
(semi-known, est. as medium) Amend upgrade_tasks to work for Rocky->Stein with OS upgrade.
system_upgrade_transfer_data:
- (unknown, est. as easy) Detect upgraded vs. unupgraded machines to transfer data to/from.
- (known, est. as easy) Stop all services on the unupgraded machines transfer data to/from. (Needs to be done via external upgrade tasks which is new, but likely not much different from what we've been doing.)
- (semi-known, est. as medium/hard) Implement an Ansible role for transferring data from one node to another via undercloud.
- (unknown, est. as medium) Figure out which data needs transferring from old controller to new, implement it using the above Ansible role -- we expect only MariaDB to require this, any special services should probably be tackled by service squads.
(semi-known, est. as medium/hard) Implement Ceph specifics, mainly how to upgrade one node (mon, OSD, ...) at a time.
(unknown, either easy or hacky or impossible :) ) Implement --limit for external-upgrade run. (As external upgrade runs on undercloud by default, we'll need to use delegate_to or nested Ansible for overcloud nodes. I'm not sure how well --limit will play with this.)
(known, est. as easy) Change update/upgrade CLI from --nodes and --roles to --limit.
(semi-known, est. as easy/medium) Add -e variable pass-through support to external-upgrade run.
(unknown, unknown) Test as much as we can in CI -- integrate with tripleo-upgrade and OOOQ.
For reprovisioning:
- (semi-known, est. as medium) Implement openstack overcloud admin authorize. Should take --stack, --limit, --overcloud-ssh-* params.
- (semi-known, est. as medium/hard) Implement an Ansible role for temporarily persisting overcloud nodes' data on the undercloud and restoring it.
- (known, est. as easy) Implement external-upgrade run --tags system_upgrade_cleanup.
- (unknown, est. as hard in total, but should probably be tackled by service squads) Figure out which data needs persisting for particular services and implement the persistence using the above Ansible role.
For in-place:
- (semi-known, est. as easy) Calls to Leapp in system_upgrade_prepare, system_upgrade_run.
- (semi-known, est. as medium) Implement a Leapp actor to set up or use the repositories we need.

Dependencies

For in-place: Leapp tool being ready to upgrade the OS.
Changes to ceph-ansible might be necessary to make it possible to run it on a single node (for upgrading mons and OSDs node-by-node).

Testing

Testing is one of the main estimated pain areas. This is a traditional problem with upgrades, but it's even more pronounced for OS upgrades.

Since we do all the OpenStack infra cloud testing of TripleO on CentOS 7 currently, it would make sense to test an upgrade to CentOS 8. However, CentOS 8 is nonexistent at the time of writing.
It is unclear when Leapp will be ready for testing an upgrade from CentOS 7, and it's probably the only thing we'd be able to execute in CI. The openstack server rebuild alternative is probably not easily executable in CI, at least not in OpenStack infra clouds. We might be able to emulate reprovisioning by wiping data.
Even if we find a way to execute the upgrade in CI, it might still take too long to make the testing plausible for validating patches.

Documentation Impact

Upgrade docs will need to be amended, the above spec is written mainly from the perspective of expected operator workflow, so it should be a good starting point.

26 KiB Raw Blame History