neutron/doc/source/devref/effective_neutron.rst

Effective Neutron: 100 specific ways to improve your Neutron contributions

There are a number of skills that make a great Neutron developer: writing good code, reviewing effectively, listening to peer feedback, etc. The objective of this document is to describe, by means of examples, the pitfalls, the good and bad practices that 'we' as project encounter on a daily basis and that make us either go slower or accelerate while contributing to Neutron.

By reading and collaboratively contributing to such a knowledge base, your development and review cycle becomes shorter, because you will learn (and teach to others after you) what to watch out for, and how to be proactive in order to prevent negative feedback, minimize programming errors, writing better tests, and so on and so forth...in a nutshell, how to become an effective Neutron developer.

The notes below are meant to be free-form and brief by design. They are not meant to replace or duplicate OpenStack documentation, or any project-wide documentation initiative like peer-review notes or the team guide. For this reason, references are acceptable and should be favored, if the shortcut is deemed useful to expand on the distilled information. We will try to keep these notes tidy by breaking them down into sections if it makes sense. Feel free to add, adjust, remove as you see fit. Please do so, taking into consideration yourself and other Neutron developers as readers. Capture your experience during development and review and add any comment that you believe will make your life and others' easier.

Happy hacking!

Developing better software

Plugin development

Document common pitfalls as well as good practices done during plugin development.

• Use mixin classes as last resort. They can be a powerful tool to add behavior but their strength is also a weakness, as they can introduce unpredictable behavior to the MRO, amongst other issues.
• In lieu of mixins, if you need to add behavior that is relevant for ML2, consider using the extension manager.
• If you make changes to the DB class methods, like calling methods that can be inherited, think about what effect that may have to plugins that have controller backends.
• If you make changes to the ML2 plugin or components used by the ML2 plugin, think about the effect that may have to other plugins.
• When adding behavior to the L2 and L3 db base classes, do not assume that there is an agent on the other side of the message broker that interacts with the server. Plugins may not rely on agents at all.
• Be mindful of required capabilities when you develop plugin extensions. The Extension description provides the ability to specify the list of required capabilities for the extension you are developing. By declaring this list, the server will not start up if the requirements are not met, thus avoiding leading the system to experience undetermined behavior at runtime.

Database interaction

Document common pitfalls as well as good practices done during database development.

• first() does not raise an exception.

• Do not use delete() to remove objects. A delete query does not load the object so no sqlalchemy events can be triggered that would do things like recalculate quotas or update revision numbers of parent objects. For more details on all of the things that can go wrong using bulk delete operations, see the "Warning" sections in the link above.

• For PostgreSQL if you're using GROUP BY everything in the SELECT list must be an aggregate SUM(...), COUNT(...), etc or used in the GROUP BY.

  The incorrect variant:

  q = query(Object.id, Object.name,
            func.count(Object.number)).group_by(Object.name)

  The correct variant:

  q = query(Object.id, Object.name,
            func.count(Object.number)).group_by(Object.id, Object.name)

• Beware of the InvalidRequestError exception. There is even a Neutron bug registered for it. Bear in mind that this error may also occur when nesting transaction blocks, and the innermost block raises an error without proper rollback. Consider if savepoints can fit your use case.

• When designing data models that are related to each other, be careful to how you model the relationships' loading strategy. For instance a joined relationship can be very efficient over others (some examples include router gateways or network availability zones).

• If you add a relationship to a Neutron object that will be referenced in the majority of cases where the object is retrieved, be sure to use the lazy='joined' parameter to the relationship so the related objects are loaded as part of the same query. Otherwise, the default method is 'select', which emits a new DB query to retrieve each related object adversely impacting performance. For example, see patch 88665 which resulted in a significant improvement since router retrieval functions always include the gateway interface.

• Conversely, do not use lazy='joined' if the relationship is only used in corner cases because the JOIN statement comes at a cost that may be significant if the relationship contains many objects. For example, see patch 168214 which reduced a subnet retrieval by ~90% by avoiding a join to the IP allocation table.

• When writing extensions to existing objects (e.g. Networks), ensure that they are written in a way that the data on the object can be calculated without additional DB lookup. If that's not possible, ensure the DB lookup is performed once in bulk during a list operation. Otherwise a list call for a 1000 objects will change from a constant small number of DB queries to 1000 DB queries. For example, see patch 257086 which changed the availability zone code from the incorrect style to a database friendly one.

• Sometimes in code we use the following structures:

  def create():
     with context.session.begin(subtransactions=True):
         create_something()
         try:
             _do_other_thing_with_created_object()
         except Exception:
             with excutils.save_and_reraise_exception():
                 delete_something()
  
  def _do_other_thing_with_created_object():
     with context.session.begin(subtransactions=True):
         ....

  The problem is that when exception is raised in _do_other_thing_with_created_object it is caught in except block, but the object cannot be deleted in except section because internal transaction from _do_other_thing_with_created_object has been rolled back. To avoid this nested transactions should be used. For such cases help function safe_creation has been created in neutron/db/_utils.py. So, the example above should be replaced with:

  _safe_creation(context, create_something, delete_something,
                 _do_other_thing_with_created_object)

  Where nested transaction is used in _do_other_thing_with_created_object function.

  The _safe_creation function can also be passed thetransaction=False`` argument to prevent any transaction from being created just to leverage the automatic deletion on exception logic.

• Beware of ResultProxy.inserted_primary_key which returns a list of last inserted primary keys not the last inserted primary key:

  result = session.execute(mymodel.insert().values(**values))
  # result.inserted_primary_key is a list even if we inserted a unique row!

• Beware of pymysql which can silently unwrap a list with an element (and hide a wrong use of ResultProxy.inserted_primary_key for example):

  e.execute("create table if not exists foo (bar integer)")
  e.execute(foo.insert().values(bar=1))
  e.execute(foo.insert().values(bar=[2]))

  The 2nd insert should crash (list provided, integer expected). It crashes at least with mysql and postgresql backends, but succeeds with pymysql because it transforms them into:

  INSERT INTO foo (bar) VALUES (1)
  INSERT INTO foo (bar) VALUES ((2))

System development

Document common pitfalls as well as good practices done when invoking system commands and interacting with linux utils.

• When a patch requires a new platform tool or a new feature in an existing tool, check if common platforms ship packages with the aforementioned feature. Also, tag such a patch with UpgradeImpact to raise its visibility (as these patches are brought up to the attention of the core team during team meetings). More details in review guidelines.
• When a patch or the code depends on a new feature in the kernel or in any platform tools (dnsmasq, ip, Open vSwitch etc.), consider introducing a new sanity check to validate deployments for the expected features. Note that sanity checks must not check for version numbers of underlying platform tools because distributions may decide to backport needed features into older versions. Instead, sanity checks should validate actual features by attempting to use them.

Eventlet concurrent model

Document common pitfalls as well as good practices done when using eventlet and monkey patching.

• Do not use with_lockmode('update') on SQL queries without protecting the operation with a lockutils semaphore. For some SQLAlchemy database drivers that operators may choose (e.g. MySQLdb) it may result in a temporary deadlock by yielding to another coroutine while holding the DB lock. The following wiki provides more details: https://wiki.openstack.org/wiki/OpenStack_and_SQLAlchemy#MySQLdb_.2B_eventlet_.3D_sad

Mocking and testing

Document common pitfalls as well as good practices done when writing tests, any test. For anything more elaborate, please visit the testing section.

• Preferring low level testing versus full path testing (e.g. not testing database via client calls). The former is to be favored in unit testing, whereas the latter is to be favored in functional testing.

• Prefer specific assertions (assert(Not)In, assert(Not)IsInstance, assert(Not)IsNone, etc) over generic ones (assertTrue/False, assertEqual) because they raise more meaningful errors:

  def test_specific(self):
      self.assertIn(3, [1, 2])
      # raise meaningful error: "MismatchError: 3 not in [1, 2]"
  
  def test_generic(self):
      self.assertTrue(3 in [1, 2])
      # raise meaningless error: "AssertionError: False is not true"

• Use the pattern "self.assertEqual(expected, observed)" not the opposite, it helps reviewers to understand which one is the expected/observed value in non-trivial assertions. The expected and observed values are also labeled in the output when the assertion fails.

• Prefer specific assertions (assertTrue, assertFalse) over assertEqual(True/False, observed).

• Don't write tests that don't test the intended code. This might seem silly but it's easy to do with a lot of mocks in place. Ensure that your tests break as expected before your code change.

• Avoid heavy use of the mock library to test your code. If your code requires more than one mock to ensure that it does the correct thing, it needs to be refactored into smaller, testable units. Otherwise we depend on fullstack/tempest/api tests to test all of the real behavior and we end up with code containing way too many hidden dependencies and side effects.

• All behavior changes to fix bugs should include a test that prevents a regression. If you made a change and it didn't break a test, it means the code was not adequately tested in the first place, it's not an excuse to leave it untested.

• Test the failure cases. Use a mock side effect to throw the necessary exceptions to test your 'except' clauses.

• Don't mimic existing tests that violate these guidelines. We are attempting to replace all of these so more tests like them create more work. If you need help writing a test, reach out to the testing lieutenants and the team on IRC.

• Mocking open() is a dangerous practice because it can lead to unexpected bugs like bug 1503847. In fact, when the built-in open method is mocked during tests, some utilities (like debtcollector) may still rely on the real thing, and may end up using the mock rather what they are really looking for. If you must, consider using OpenFixture, but it is better not to mock open() at all.

Backward compatibility

Document common pitfalls as well as good practices done when extending the RPC Interfaces.

• Make yourself familiar with Upgrade review guidelines <upgrade_review_guidelines>.

Deprecation

Sometimes we want to refactor things in a non-backward compatible way. In most cases you can use debtcollector to mark things for deprecation. Config items have deprecation options supported by oslo.config.

The deprecation process must follow the standard deprecation requirements. In terms of neutron development, this means:

• A launchpad bug to track the deprecation.
• A patch to mark the deprecated items. If the deprecation affects users (config items, API changes) then a release note must be included.
• Wait at least one cycle and at least three months linear time.
• A patch that removes the deprecated items. Make sure to refer to the original launchpad bug in the commit message of this patch.

Scalability issues

Document common pitfalls as well as good practices done when writing code that needs to process a lot of data.

Translation and logging

Document common pitfalls as well as good practices done when instrumenting your code.

• Make yourself familiar with OpenStack logging guidelines to avoid littering the logs with traces logged at inappropriate levels.
• The logger should only be passed unicode values. For example, do not pass it exceptions or other objects directly (LOG.error(exc), LOG.error(port), etc.). See http://docs.openstack.org/developer/oslo.log/usage.html#no-more-implicit-conversion-to-unicode-str for more details.
• Don't pass exceptions into LOG.exception: it is already implicitly included in the log message by Python logging module.
• Don't use LOG.exception when there is no exception registered in current thread context: Python 3.x versions before 3.5 are known to fail on it.

Project interfaces

Document common pitfalls as well as good practices done when writing code that is used to interface with other projects, like Keystone or Nova.

Documenting your code

Document common pitfalls as well as good practices done when writing docstrings.

Landing patches more rapidly

Scoping your patch appropriately

• Do not make multiple changes in one patch unless absolutely necessary. Cleaning up nearby functions or fixing a small bug you noticed while working on something else makes the patch very difficult to review. It also makes cherry-picking and reverting very difficult. Even apparently minor changes such as reformatting whitespace around your change can burden reviewers and cause merge conflicts.
• If a fix or feature requires code refactoring, submit the refactoring as a separate patch than the one that changes the logic. Otherwise it's difficult for a reviewer to tell the difference between mistakes in the refactor and changes required for the fix/feature. If it's a bug fix, try to implement the fix before the refactor to avoid making cherry-picks to stable branches difficult.
• Consider your reviewers' time before submitting your patch. A patch that requires many hours or days to review will sit in the "todo" list until someone has many hours or days free (which may never happen.) If you can deliver your patch in small but incrementally understandable and testable pieces you will be more likely to attract reviewers.

Nits and pedantic comments

Document common nits and pedantic comments to watch out for.

• Make sure you spell correctly, the best you can, no-one wants rebase generators at the end of the release cycle!
• The odd pep8 error may cause an entire CI run to be wasted. Consider running validation (pep8 and/or tests) before submitting your patch. If you keep forgetting consider installing a git hook so that Git will do it for you.
• Sometimes, new contributors want to dip their toes with trivial patches, but we at OpenStack love bike shedding and their patches may sometime stall. In some extreme cases, the more trivial the patch, the higher the chances it fails to merge. To ensure we as a team provide/have a frustration-free experience new contributors should be redirected to fixing low-hanging-fruit bugs that have a tangible positive impact to the codebase. Spelling mistakes, and docstring are fine, but there is a lot more that is relatively easy to fix and has a direct impact to Neutron users.

Reviewer comments

• Acknowledge them one by one by either clicking 'Done' or by replying extensively. If you do not, the reviewer won't know whether you thought it was not important, or you simply forgot. If the reply satisfies the reviewer, consider capturing the input in the code/document itself so that it's for reviewers of newer patchsets to see (and other developers when the patch merges).
• Watch for the feedback on your patches. Acknowledge it promptly and act on it quickly, so that the reviewer remains engaged. If you disappear for a week after you posted a patchset, it is very likely that the patch will end up being neglected.
• Do not take negative feedback personally. Neutron is a large project with lots of contributors with different opinions on how things should be done. Many come from widely varying cultures and languages so the English, text-only feedback can unintentionally come across as harsh. Getting a -1 means reviewers are trying to help get the patch into a state that can be merged, it doesn't just mean they are trying to block it. It's very rare to get a patch merged on the first iteration that makes everyone happy.

Code Review

• You should visit OpenStack How To Review wiki
• Stay focussed and review what matters for the release. Please check out the Neutron section for the Gerrit dashboard. The output is generated by this tool.

IRC

• IRC is a place where you can speak with many of the Neutron developers and core reviewers. For more information you should visit OpenStack IRC wiki Neutron IRC channel is #openstack-neutron
• There are weekly IRC meetings related to many different projects/teams in Neutron. A full list of these meetings and their date/time can be found in OpenStack IRC Meetings. It is important to attend these meetings in the area of your contribution and possibly mention your work and patches.
• When you have questions regarding an idea or a specific patch of yours, it can be helpful to find a relevant person in IRC and speak with them about it. You can find a user's IRC nickname in their launchpad account.
• Being available on IRC is useful, since reviewers can contact you directly to quickly clarify a review issue. This speeds up the feedback loop.
• Each area of Neutron or sub-project of Neutron has a specific lieutenant in charge of it. You can most likely find these lieutenants on IRC, it is advised however to try and send public questions to the channel rather then to a specific person if possible. (This increase the chances of getting faster answers to your questions). A list of the areas and lieutenants nicknames can be found at Core Reviewers.

Commit messages

Document common pitfalls as well as good practices done when writing commit messages. For more details see Git commit message best practices. This is the TL;DR version with the important points for committing to Neutron.

• One liners are bad, unless the change is trivial.
• Remember to use DocImpact, APIImpact, UpgradeImpact appropriately.
• Make sure the commit message doesn't have any spelling/grammar errors. This is the first thing reviewers read and they can be distracting enough to invite -1's.
• Describe what the change accomplishes. If it's a bug fix, explain how this code will fix the problem. If it's part of a feature implementation, explain what component of the feature the patch implements. Do not just describe the bug, that's what launchpad is for.
• Use the "Closes-Bug: #BUG-NUMBER" tag if the patch addresses a bug. Submitting a bugfix without a launchpad bug reference is unacceptable, even if it's trivial. Launchpad is how bugs are tracked so fixes without a launchpad bug are a nightmare when users report the bug from an older version and the Neutron team can't tell if/why/how it's been fixed. Launchpad is also how backports are identified and tracked so patches without a bug report cannot be picked to stable branches.
• Use the "Implements: blueprint NAME-OF-BLUEPRINT" or "Partially-Implements: blueprint NAME-OF-BLUEPRINT" for features so reviewers can determine if the code matches the spec that was agreed upon. This also updates the blueprint on launchpad so it's easy to see all patches that are related to a feature.
• If it's not immediately obvious, explain what the previous code was doing that was incorrect. (e.g. code assumed it would never get 'None' from a function call)
• Be specific in your commit message about what the patch does and why it does this. For example, "Fixes incorrect logic in security groups" is not helpful because the code diff already shows that you are modifying security groups. The message should be specific enough that a reviewer looking at the code can tell if the patch does what the commit says in the most appropriate manner. If the reviewer has to guess why you did something, lots of your time will be wasted explaining why certain changes were made.

Dealing with Zuul

Document common pitfalls as well as good practices done when dealing with OpenStack CI.

• When you submit a patch, consider checking its status in the queue. If you see a job failures, you might as well save time and try to figure out in advance why it is failing.
• Excessive use of 'recheck' to get test to pass is discouraged. Please examine the logs for the failing test(s) and make sure your change has not tickled anything that might be causing a new failure or race condition. Getting your change in could make it even harder to debug what is actually broken later on.