Node.js - Quality with Speed

By Michael Dawson and Myles Borins,

Node.js - Quality with Speed

One of the key tenets of the Node.js community is to allow change at a rapid pace in order to foster innovation and to allow Node.js to be used in a growing number of use cases.

At the same time the community values quality. Newer versions of the runtime must be as good or better than earlier versions and must not un-intentionally break existing applications.

Instead of trading off one for the other, the community looks for the path that allows us to maintain our rate of change while ensuring the required level of quality.

Many of the activities undertaken by the community over the last year are in support of this goal.

This is our take on how these activities fit together.

Key strategies

Several key strategies are in place to build the safety net in order to enable change/innovation while maintaining quality. These include:

  • Different release types
  • Change flow processes
  • Enhancement Proposal process
  • Automation and Testing
    • Functional Tests
    • Module Testing
    • Stress Testing
    • Platform/OS coverage
    • Development Workflows
  • Performance Benchmarks
  • Tools

Release Types

The Node.js project maintains 3 key types of releases

  • Nightlies
  • Current
  • LTS

Having different release types allows innovation/change to flow rapidly into Nightly builds where we can get early feedback on upcoming changes. When ready, these changes then transition to Current and LTS releases in a more controlled manner, such that the level of quality/stability increases at each level.

Nightlies

These are built from master and contain the very latest changes that have been landed. If you want to try out the bleeding edge these are the binaries to use. There is no additional testing on these releases, however, the standard Node.js unit tests are run for each change landed so these will most often be usable.

Current

Changes which have landed in master are backported to Current on a regular basis. In general all changes that land in master will be backported to Current, however there may be a lag if there are specific concerns or for larger changes where the community believes more soak time is required. One key exception is that semver breaking changes will not be backported until the next major version (ex 5 -> 6). This includes V8 and other components such that the expectation is that an application/module written to run on a major level will continue to do so.

These releases are documented in the changelog so there is more visibility with respect to the changes in each release. Current releases are created roughly every 1-2 weeks.

In addition to the regular Node.js unit tests, CITGM (see later sections) is run on Current releases.

If you want to try out the latest with a reasonable expectation that your application will continue to run, these are the releases to use.

LTS

Once changes have been proven in the Current stream, they are candidates for the LTS streams. In the first stage of LTS (Active) changes are limited to:

  • Bug fixes
  • Security updates
  • Non-semver-major npm updates
  • Relevant documentation updates
  • Certain performance improvements where the risk of breaking existing applications is minimal
  • Changes that introduce large amount of code churn where the risk of breaking existing applications is low and where the change in question may significantly ease the ability to backport future changes due to the reduction in diff noise.

Further, in the second stage of an LTS release (Maintenance), only critical bugs and critical security fixes will be included.

Like Current releases, CITGM (see later sections) is run on LTS releases. In addition we also track performance through nightly benchmarks reported on benchmarking.nodejs.org (See later sections).

You can read more about the LTS releases here.

If you want the best level of stability/quality for your production applications these are the releases to use.

Change flow processes

We've already touched on this in the discussion on the different release types but we'll expand on this strategy here.

The main idea is that as changes flow from Nightlies, to Stable, to LTS Active, to LTS Maintenance we increase the following:

  • scrutiny
  • time

Changes going into master are well reviewed and time is allowed (minimum 48 to 72 hours) for as many community members as possible to comment and review. However, as we all know, some problems will still get through.

Before changes are pulled into Current from the Nightly builds, they will have spent more time in master where intermittent issues may surface in the ongoing regressions runs and this provides time where users may more fully exercise the release and report issues. Further, there is an additional review/sanity check that they are not breaking as they are pulled over to Current.

Similarly, before changes are pulled into an LTS update release, they must have been in a Current release version for at least a week, and are often left longer. This provides additional time where users may more fully exercise the changes and report issues. In addition, changes are more carefully reviewed as they are pulled into LTS, again reducing the chance that unintentional breaking changes make it through. As an LTS release ages, particularly once it reaches maintenance, the scope of changes that will be pulled in narrows, further reducing the risk.

When it comes to new LTS versions, changes will have soaked in the latest release for up to 6 months. In particular, larger changes like an upgrade to V8 are done early in the lifespan of the stream such that they will have significant soaking and usage in the Current stream before they make it into an LTS release.

This strategy allows for rapid innovation/change, with releases being available where those changes can be used/validated and a funnel through which these can flow in an appropriate manner into releases used by more risk-averse community members.

Enhancement Proposal Process

Some changes are of such scope that they cannot simply be reviewed in a pull request. There are often larger questions that will factor into the decision as to whether the change proposed is desirable or appropriate for the Node.js runtime.

The strategy for these changes is the "enhancement proposal" process. The proposed change is documented, discussed and moves through a number of stages including DRAFT and ACCEPTED or REJECTED. You can read more on the process here.

This process ensures that larger changes can be discussed in advance and agreed by the community, allowing the final review of the pull request to focus on implementation. The result being that the merits of the concept can be discussed at the appropriate level of abstraction without having to review all of the technical details.

Automation and Testing

Automation and Testing are key strategies that go hand in hand in allowing rapid change in a safe manner.

Automation avoids error-prone manual steps. Once you have a task automated the likelihood of errors is orders of magnitude smaller than doing those tasks by hand, particularly when those tasks are done by different individuals.

One of our key tenets is to automate as much as we can. This ranges all the way from the configuration of the machines in our build infrastructure using Ansible, to automated jobs that build/sign/and release our binaries.

Automated Testing allows tests to be run often enough to catch regressions quickly and reliably. Given a good set of tests, we can make changes confident that we'll catch changes which introduce regressions.

There are many levels of testing and the strategy is to build our way up the levels until we have as complete coverage as is reasonable.

These levels include:

  • Functional Tests
  • Platform/OS Coverage
  • Dependency Testing
  • Module Testing
  • Stress Testing
  • Development Workflows
  • Use Case Testing

Functional Tests

Functional tests are the first level of defense. Our collaborator guidelines require test cases for all new features added, and our collaborators set a high standard in respect to requiring tests.

It is not enough to simply have tests, those tests must be effective at exercising the runtime. We measure code coverage nightly and publish the results at coverage.nodejs.org. This allows us to ensure our tests remain effective and provides the data necessary to further improve our tests.

You'll also notice that there has been a lot of effort put into making sure the tests pass reliably and consistently. If you watch the continuous integration (ci) runs you will see that they are mostly green and intermittent failures are rare.

Platform/OS Coverage

This is not a type of test by itself. But by applying the strategy of running tests across a broad range of platforms and OS types and levels it multiplies the effectiveness of the existing tests.

Issues which surface on a particular platform or OS often are not specific to that platform or OS but instead are uncovered because of different timing, default configuration or general environment. They could have occurred on any of the other platforms.

Our strategy is to test on a broad range of platforms both to ensure Node.js works on our supported platforms, but also to leverage the diversity to uncover as many problems as early as possible.

Dependency Testing

Node.js has a number of key dependencies. It's important that we ensure that any changes we apply to those dependencies don't have a negative effect.

To this end we have a job which runs the V8 tests on the V8 tree within the Node.js repo. This job runs nightly and on request for PRs that are making changes to the V8 tree.

We don't currently run the tests for other dependencies, but the delta in the Node.js tree for the dependencies other than V8 is more limited.

Module Tests

Module tests are the next level of defense. They help to validate that changes are not going to break for end users. Most applications use a number of modules, and many of the most popular modules are extensively used. Any changes that impact those modules would have a significant community impact.

Our strategy is to run the module's own unit tests on a set of key modules and to run these as often as possible. Currently they are run for Current and LTS releases and we are working to increase that frequency.

You can read more about our module testing efforts in https://github.com/nodejs/citgm/.

Stress Tests

Some problems only surface after running for a long time. Stress tests help to flush those out by running certain scenarios over a prolonged period of time.

We don't have any stress tests running at this point but it will be our next priority after we have module testing running at an appropriate frequency.

Development Workflows

Development Workflows is another level up from Module Testing. It aims to test common development workflows to ensure changes will not introduce any regressions to those flows.

These are more work to put in place and run but they will be next on our list after getting stress tests in place.

Use Case Testing

This would be the next logical step after Development Workflows, testing for the common use cases for Node.js.

Our current strategy is to get some of this coverage through the benchmarking that we put in place, but it is another area we can work on once we have the other levels of testing in place.

Performance Benchmarks

While ensuring functional stability is good, its not enough. We also need to make sure that performance is not degraded as changes flow in.

Our strategy is to define the common use cases for Node.js and then build up a set of benchmarks that we run and publish results for on a regular basis. This work is ongoing in the Benchmarking Working Group, but we already have a number of key benchmarks being run nightly across the major Node.js versions. You can view this data at:

https://benchmarking.nodejs.org/.

This data allows us to ensure we avoid performance regressions as changes flow in.

In Summary

This may have been a bit of a long read but I hope it has put a number of the activities you may have seen in the Node.js community over the last year into context. If you ever wondered "Why are they doing that?", the answer is:

Node.js - Quality with Speed

Retorne ao começo