There was recent discussion in some reviews about fixing latent bugs in the legacy v2.0 API code. Since the v2.0 API is deprecated and v2.1 is the default since Liberty, we shouldn't need to fix latent low-priority bugs in v2.0 anymore. However, we'll still fix critical bugs, and we shouldn't knowingly introduce new regressions that would result in a 500 response. Change-Id: I9937d9226a99754dadcc48d599090296f5ae01f7
12 KiB
Code Review Guide for Nova
This is a very terse set of points for reviewers to consider when looking at nova code. These are things that are important for the continued smooth operation of Nova, but that tend to be carried as "tribal knowledge" instead of being written down. It is an attempt to boil down some of those things into nearly checklist format. Further explanation about why some of these things are important belongs elsewhere and should be linked from here.
Upgrade-Related Concerns
RPC API Versions
- If an RPC method is modified, the following needs to happen:
- The manager-side (example: compute/manager) needs a version bump
- The manager-side method needs to tolerate older calls as well as newer calls
- Arguments can be added as long as they are optional. Arguments cannot be removed or changed in an incompatible way.
- The RPC client code (example: compute/rpcapi.py) needs to be able to honor a pin for the older version (see self.client.can_send_version() calls). If we are pinned at 1.5, but the version requirement for a method is 1.7, we need to be able to formulate the call at version 1.5.
- Methods can drop compatibility with older versions when we bump a major version.
- RPC methods can be deprecated by removing the client (example: compute/rpcapi.py) implementation. However, the manager method must continue to exist until the major version of the API is bumped.
Object Versions
- If a tracked attribute (i.e. listed in fields) or remotable method is added, or a method is changed, the object version must be bumped. Changes for methods follow the same rules as above for regular RPC methods. We have tests to try to catch these changes, which remind you to bump the version and then correct the version-hash in the tests.
- Field types cannot be changed. If absolutely required, create a new attribute and deprecate the old one. Ideally, support converting the old attribute to the new one with an obj_load_attr() handler. There are some exceptional cases where changing the type can be allowed, but care must be taken to ensure it does not affect the wireline API.
- New attributes should be removed from the primitive in obj_make_compatible() if the attribute was added after the target version.
- Remotable methods should not return unversioned structures wherever possible. They should return objects or simple values as the return types are not (and cannot) be checked by the hash tests.
- Remotable methods should not take complex structures as arguments. These cannot be verified by the hash tests, and thus are subject to drift. Either construct an object and pass that, or pass all the simple values required to make the call.
- Changes to an object as described above will cause a hash to change in TestObjectVersions. This is a reminder to the developer and the reviewer that the version needs to be bumped. There are times when we need to make a change to an object without bumping its version, but those cases are only where the hash logic detects a change that is not actually a compatibility issue and must be handled carefully.
Database Schema
- Changes to the database schema must generally be additive-only. This means you can add columns, but you can't drop or alter a column. We have some hacky tests to try to catch these things, but they are fragile. Extreme reviewer attention to non-online alterations to the DB schema will help us avoid disaster.
- Dropping things from the schema is a thing we need to be extremely careful about, making sure that the column has not been used (even present in one of our models) for at least a release.
- Data migrations must not be present in schema migrations. If data needs to be converted to another format, or moved from one place to another, then that must be done while the database server remains online. Generally, this can and should be hidden within the object layer so that an object can load from either the old or new location, and save to the new one.
REST API
When making a change to the nova API, we should always follow the API WG guidelines rather than going for "local" consistency. Developers and reviewers should read all of the guidelines, but they are very long. So here are some key points:
-
project
should be used in the REST API instead oftenant
.server
should be used in the REST API instead ofinstance
.compute
should be used in the REST API instead ofnova
.
-
- URL should not include underscores; use hyphens ('-') instead.
- The field names contained in a request/response body should use snake_case style, not CamelCase or Mixed_Case style.
-
- Synchronous resource creation:
201 Created
- Asynchronous resource creation:
202 Accepted
- Synchronous resource deletion:
204 No Content
- For all other successful operations:
200 OK
- Synchronous resource creation:
Fixing bugs in v2.0
The v2.1 API was made the default in Liberty1 and the ability to disable the v2.1 API was removed in Newton2. The v2.0 legacy API code is also deprecated3. As such, latent bugs do not need to be fixed in the legacy v2.0 API code, with the exception of critical bugs like security vulnerabilities. Also, new regressions should be avoided, like if raising a new exception would result in a 500 response from the REST API, that should be handled in the v2.0 API code as well as the v2.1 API.
Config Options
Location
The central place where all config options should reside is the
/nova/conf/
package. Options that are in named sections of
nova.conf
, such as [serial_console]
, should be
in their own module. Options that are in the [DEFAULT]
section should be placed in modules that represent a natural grouping.
For example, all of the options that affect the scheduler would be in
the scheduler.py
file, and all the networking options would
be moved to network.py
.
Implementation
A config option should be checked for:
A short description which explains what it does. If it is a unit (e.g. timeouts or so) describe the unit which is used (seconds, megabyte, mebibyte, ...).
A long description which shows the impact and scope. The operators should know the expected change in the behavior of Nova if they tweak this.
Hints which services will consume this config option. Operators/Deployers should not be forced to read the code to know which one of the services will change its behavior nor should they set this in every
nova.conf
file to be sure.Descriptions/Validations for the possible values.
- If this is an option with numeric values (int, float), describe the edge cases (like the min value, max value, 0, -1).
- If this is a DictOpt, describe the allowed keys.
- If this is a StrOpt, list any possible regex validations, or provide a list of acceptable and/or prohibited values.
Interdependencies to other options. If other config options have to be considered when this config option gets changed, is this described?
Third Party Tests
Any change that is not tested well by the Jenkins check jobs must have a recent +1 vote from an appropriate third party test (or tests) on the latest patchset, before a core reviewer is allowed to make a +2 vote.
Virt drivers
At a minimum, we must ensure that any technology specific code has a +1 from the relevant third party test, on the latest patchset, before a +2 vote can be applied. Specifically, changes to nova/virt/driver/<NNNN> need a +1 vote from the respective third party CI. For example, if you change something in the XenAPI virt driver, you must wait for a +1 from the XenServer CI on the latest patchset, before you can give that patch set a +2 vote.
This is important to ensure:
- We keep those drivers stable
- We don't break that third party CI
Notes
Please note:
- Long term, we should ensure that any patch a third party CI is allowed to vote on, can be blocked from merging by that third party CI. But we need a lot more work to make something like that feasible, hence the proposed compromise.
- While its possible to break a virt driver CI system by changing code that is outside the virt drivers, this policy is not focusing on fixing that. A third party test failure should always be investigated, but the failure of a third party test to report in a timely manner should not block others.
- We are only talking about the testing of in-tree code. Please note
the only public API is our REST API, see:
policies
Microversion API
- If an new microversion API is added, the following needs to happen:
- A new patch for the microversion API change in python-novaclient side should be submitted.
Release Notes
What is reno ?
Nova uses reno for providing release notes in-tree. That means that a patch can include a reno file or a series can have a follow-on change containing that file explaining what the impact is.
A reno file is a YAML file written in the releasenotes/notes tree which is generated using the reno tool this way:
$ tox -e venv -- reno new <name-your-file>
where usually <name-your-file>
can be
bp-<blueprint_name>
for a blueprint or
bug-XXXXXX
for a bugfix.
Refer to the reno documentation for the full list of sections.
When a release note is needed
A release note is required anytime a reno section is needed. Below are some examples for each section. Any sections that would be blank should be left out of the note file entirely. If no section is needed, then you know you don't need to provide a release note :-)
upgrade
-
- The patch has an UpgradeImpact tag
- A DB change needs some deployer modification (like a migration)
- A configuration option change (deprecation, removal or modified default)
- some specific changes that have a DocImpact tag but require further action from an deployer perspective
- any patch that requires an action from the deployer in general
security
-
- If the patch fixes a known vulnerability
features
-
- If the patch has an APIImpact tag
- For nova-manage and python-novaclient changes, if it adds or changes a new command, including adding new options to existing commands
- not all blueprints in general, just the ones impacting a contractual API
- a new virt driver is provided or an existing driver impacts the HypervisorSupportMatrix
critical
-
- Bugfixes categorized as Critical in Launchpad impacting users
fixes
-
- No clear definition of such bugfixes. Hairy long-standing bugs with high importance that have been fixed are good candidates though.
Three sections are left intentionally unexplained
(prelude
, issues
and other
).
Those are targeted to be filled in close to the release time for
providing details about the soon-ish release. Don't use them unless you
know exactly what you are doing.
Notifications
- Every new notification type shall use the new versioned notification
infrastructure documented in
notifications