Transferring Zend Framework to Laminas: A Retrospective.

Back in October 2018, Rogue Wave Software announced reorganization of its Zend portfolio, including Zend Framework. The Zend Framework community was understandably shaken, and many were concerned about the future of the framework. Six months later, in April of 2019, Rogue Wave announced they would be transferring the project to the Linux Foundation as the Laminas Project.

A couple months before this announcement, in February 2019, we started working on a tool for transferring the project repositories to their new homes. At the beginning, everyone thought the tool should be relatively simple, as we "just" needed to change namespaces. But our goal was much greater: we wanted to provide packages fully compatible with legacy Zend Framework components. We wanted our new components to replace legacy components (yes, we wanted to use replace in composer.json).

We needed a tool capable of much more than just "rewriting namespaces".

Lifetime decisions

Rewrite whole history or just tags?

We had over 150 components to move, many of them huge components with 8 years of history and thousands of commits. Some were much smaller — particularly the Expressive components — with only a few years of history and a few hundred commits.

The first plan was to rewrite the whole history, every commit by using the git filter-branch command. This command allows applying arbitrary actions on every revision of the repository history. Unfortunately, as the transfer tool's functionality and scope grew, this operation became prohibitively slow; rewriting even a single component took several hours.

As such, we decided on a different approach: we decided to rewrite only tags. Since users would not be able to pin to existing commits anyways (as filter-branch creates new signatures, and thus new commit identifiers), we really only needed to worry about specific tags, which then get translated into installable releases. This also ensures it's easier to identify the full history of any given commit, which is useful to maintainers.

So we changed our process: we checked out each tag, performed our rewrite operations, dropped the original tag, and re-issued the tag with the rewritten code.

Splitting projects

Apigility had always been a separate project under a separate organization. But Expressive was bundled under the same organization as the MVC and general components. We decided it should have its own organization, just like Apigility already did.

Additionally, we felt the organization names for these subprojects should match their new names: Laminas API Tools and Mezzio.

As such, we now have the following GitHub organisations representing the Laminas Project:

  • laminas, representing components and the MVC.
  • laminas-api-tools, representing the Laminas API Tools components (formerly Apigility).
  • mezzio, representing the Mezzio middleware runtime and components (formerly Expressive).

This makes it simpler to find code related specifically to each subproject.

Unification

With the new project, we wanted to keep consistent namespaces for all components.

As an example, here are some of the previous namespaces represented in Zend Framework packages:

  • Zend
  • Zend\Expressive
  • ZendXml
  • ZendOAuth
  • ZendService\{service component namespace}
  • ZF
  • ZF\Apigility

This situation was often confusing even to the maintainers! As such, we standardized on two top-level namespaces:

  • Laminas
  • Mezzio

We consider Laminas API Tools a subnamespace, as it builds on the Laminas MVC, so all components under that subproject now have the namespace Laminas\ApiTools. Service components such as ZendOAuth and ZendService\Twitter are now Laminas components, so they get namespaces such as Laminas\OAuth and Laminas\Twitter, which are now in line with other components.

Deprecation and abandoning packages

Hard as the decision was, we also decided to abandon some packages. These include:

We simply did not have resources to update them to the latest changes of their respective APIs. We would be happy to bring them back to the Laminas Project if we find people who want to maintain them (though any such effort would have to provide features beyond what the official SDK libraries provide).

Additionally, we decided to deprecate some other minor packages, not used by our other components. These include:

Making it all work

Bridge for all components

The final result is the laminas-zendframework-bridge component required by all migrated components.

Firstly, we needed to provide a compatibility layer to allow third-party components to work with Zend Framework and Laminas components at the same time. Our goal was that switching to Laminas should not require a BC break in third-party libraries. There were a couple challenges:

  1. Loading the appropriate Laminas class if the requested Zend Framework class does not exist. This was pretty easy: we achieved it by creating an autoloader that changes the namespace on the fly, and simultaneously creates an alias for the legacy class using class_alias. This approach assures that subsequent requests for the same class use the Laminas replacement instead.

    However, autoloading is not triggered for classes referenced by typehints, which created the second challenge.

  2. Ensuring typehints for legacy classes work with Laminas replacements. As an example, please consider the following code:

namespace ThirdParty\Component;

use Zend\ServiceManager\ServiceLocatorInterface;

interface MyInterface
{
    public function run(ServiceLocatorInterface $sm);
}

We wanted to ensure code such as this works when the Laminas component replacing the ZF component is installed. To accomplish this, we created an additional autoloader, that creates a class_alias of the legacy class to the Laminas equivalent any time a Laminas class is autoloaded.

There was also additional difficulty: we had some integration classes with "Zend" in the name (see for example LaminasRouter). We were able to resolve these by including class maps to use when resolving classes with the above autoloaders.

Custom functions

In several libraries we ship, we have defined namespaced functions. These pose a problem as there is no equivalent of class_alias for functions. We needed to keep the previous function in its legacy namespace, and have it delegate to its equivalent in the new namespace.

To accomplish this, the tooling duplicates each function file using the suffix .legacy.php (e.g., normalize_server.php would get duplicated to normalize_server.legacy.php), and we use the legacy namespace in that file, and modify the function to proxy to the function in the new namespace. We then add these legacy function files to the autoloader, as additions to the existing ones. This approach allows using the legacy functions side-by-side with the new versions.

For a more complete examples:

See the transfer tool code: FunctionAliasFixture.

Custom constants

In a similar vein to namespaced functions, we ran into an issue with namespaced constants. The solution for these was the same as for functions, fortunately.

For example, see constants.php and constants.legacy.php from the mezzio/mezzio package.

See the transfer tool code: NamespacedConstantFixture.

Container / Service Manager keys

Many components provide configuration for the DI container (generally our Service Manager or a PSR-11 Container). This allows retrieving services using code like:

$serviceManager->get(ClassName::class);

or:

$container->get(ClassName::class);

Interestingly, the ::class notation does not trigger autoloading; even worse - the class name before ::class does not even need to exist! PHP expands the string according to the current namespace and imports, without validating it exists.

What we wanted to accomplish is to have:

$container->get(\Zend\ClassName::class);

and:

$container->get(\Laminas\ClassName::class);

produce the same result; exactly the same instance, not just a new instance.

Why? Because while a user might migrate their project to Laminas, some third-party libraries they use might still use the legacy names.

If you know a bit of our Service Manager configuration, the solution is relatively easy: you provide aliases mapping the legacy classes to their Laminas equivalent.

However, it's not quite that simple in practice. Container configuration can come from a variety of sources, even with a given package:

We needed to ensure each of these dependency configuration locations would get rewritten.

Delegator Factories

Unfortunately we were not able to do the same for delegator factory configuration. Delegators must be defined on the original class, not on an alias.

If a library provides a delegator for \Zend\ClassName but you are using \Laminas\ClassName, the legacy delegator will not be triggered. You will need to update your own configuration to add it.

See the transfer tool code: DIAliasFixture.

Plugin Managers

Related to the previous point, many components provide plugin managers, which your own code or third-party libraries can provide configuration for as well. The configuration is the same, but done in different locations. Adding more difficulty, components providing a plugin manager often define the plugin manager configuration directly in the plugin manager definition.

Our solution here was to alter the plugin manager classes during rewrite to alias the legacy ZF classes to their Laminas equivalents. Doing so allows them to work without any further changes, fortunately!

For an example, you can inspect the additional aliases in the FilterPluginManager.

Adding difficulty to this scenario is the fact that plugin manager definitions changed from their first introduction to their latest release, particularly as they were updated to target the version 3 release of our service manager. Our tool had to accommodate these changes!

See the transfer tool code: PluginManagerFixture.

Factories

The next challenge posed was with factory classes. Many components provide factories for services for use with the service manager. Often these factories are using other services configured in the DI Container, as well as the configuration service itself. Consider the following example:

class ExampleFactory
{
    public function __invoke(ContainerInterface $container)
    {
        $otherService = $container->get(\Zend\OtherService::class);

        return new Example($otherService);
    }
}

While our work to provide aliases means that this code should continue to work, there's one catch: resolving aliases is the slowest operation the service manager performs. As such, our migration tooling rewrites these references:

class ExampleFactory
{
    public function __invoke(ContainerInterface $container)
    {
        $otherService = $container->get(\Laminas\OtherService::class);

        return new Example($otherService);
    }
}

But what about more complex scenarios, like this one:

class ExampleFactory
{
    public function __invoke(ContainerInterface $container)
    {
        if (! $container->has(\Zend\OtherService::class)) {
            throw new MissingDependencyException();
        }

        return new Example($container->get(\Zend\OtherService::class));
    }
}

Here, we also want to be able to use the equivalent Laminas service, if defined, falling back to the legacy service if not. The rewrite tooling thus produces:

class ExampleFactory
{
    public function __invoke(ContainerInterface $container)
    {
        if (! $container->has(\Laminas\OtherService::class)
            && ! $container->has(\Zend\OtherService::class)
        ) {
            throw new MissingDependencyException();
        }

        return new Example(
            $container->has(\Laminas\OtherService::class)
                ? $container->get(\Laminas\OtherService::class)
                : $container->get(\Zend\OtherService::class)
        );
    }
}

We had even more complicated examples; look at the SwooleRequestHandlerRunnerFactory or HalResponseFactoryFactory to see how complicated it got!

While we don't like nesting ternaries, in many cases, it was the most consistent way to accomplish our ends.

See the transfer tool code: LegacyFactoriesFixture.

Configuration keys - config postprocessor and configuration merge listener

Many factories also consume and reference specific configuration. Usually components provide default configuration, and the user must adjust it for a specific case. Default configuration is usually provided under a key named after the component itself. As an example:

// Default module configuration:
return [
    'zend-expressive-hal' => [
        'metadata-factories' => [
            ResourceMetadata::class => ResourceMetadataFactory::class,
        ],
    ],
];
// Custom user configuration:
return [
    'zend-expressive-hal' => [
        'metadata-factories' => [
            CustomCollectionMetadata::class => CustomCollectionMetadataFactory::class,
        ],
    ],
];

A factory which consumes the above configuration might look like this:

class MetadataMapFactory
{
    public function __invoke(ContainerInterface $container)
    {
        $config = $container->has('config') ? $container->get('config') : [];

        $metadataMapConfig = $config[\Zend\Expressive\Hal\MetadataMap::class] ?? [];
        $metadataFactories = $config['zend-expressive-hal']['metadata-factories'] ?? [];

        return new \Zend\Expressive\Hal\MetadataMap($metadataMapConfig, $metadataFactories);
    }
}

As you can see here, we have two strings in the configuration we want to change: \Zend\Expressive\Hal\MetadataMap::class and zend-expressive-hal.

And here is the problem: we can rename them here in the factories, but then all configuration provided by third-parties or the application that use the legacy keys will be ignored.

To address this issue, we introduced a Config Post Processor for Mezzio applications and a Configuration merge listener for MVC applications.

Under the hood, each does the same thing: they intercept legacy configuration keys and merge their values with the default configuration provided under the new keys.

The result is that component configuration can reference the new keys:

return [
    'mezzio-hal' => [
        'metadata-factories' => [
            ResourceMetadata::class => ResourceMetadataFactory::class,
        ],
    ],
];

and factories can reference only the new keys as well:

class MetadataMapFactory
{
    public function __invoke(ContainerInterface $container)
    {
        $config = $container->has('config') ? $container->get('config') : [];

        $metadataMapConfig = $config[\Mezzio\Hal\MetadataMap::class] ?? [];
        $metadataFactories = $config['mezzio-hal']['metadata-factories'] ?? [];

        return new \Mezzio\Hal\MetadataMap($metadataMapConfig, $metadataFactories);
    }
}

When a configuration post processor is in play, third party configuration referencing the old keys will have its own configuration merged under the new keys, keeping backwards compatibility.

While the configuration post processing works, it is highly reliant on the idea that application-specific configuration is merged last. As such, we recommend that third-party library providers update their libraries. Until then, however, the configuration post processors provide a solution that does not incur a BC break.

Custom Request Attributes in Middleware

When using PSR-15 middleware, such as with Mezzio (formerly Expressive), we pass information between middleware using request attributes. We have standardized on using class names for these attributes, raising another migration problem. As an example, consider the following routing middleware:

class RouteMiddleware implements MiddlewareInterface
{
    public function process(
        ServerRequestInterface $request,
        RequestHandlerInterface $handler
    ) : ResponseInterface {
        $result = $this->router->match($request);

        $request = $request->withAttribute(
            \Zend\Expressive\Router\RouteResult::class,
            $result
        );

        return $handler->handle($request);
    }
}

This middleware injects an attribute containing the results of routing, so that users can later access them. This also means that users are using the legacy class name, Zend\Expressive\Router\RouteResult, in order to retrieve those values.

To provide backwards compatibility, we decided to inject two attributes, one under both the current class name, and one under the legacy class name:

class RouteMiddleware implements MiddlewareInterface
{
    public function process(
        ServerRequestInterface $request,
        RequestHandlerInterface $handler
    ) : ResponseInterface {
        $result = $this->router->match($request);

        $request = $request
            ->withAttribute(\Mezzio\Router\RouteResult::class, $result)
            ->withAttribute(\Zend\Expressive\Router\RouteResult::class, $result);

        return $handler->handle($request);
    }
}

This allows pulling using either name, ensuring your code, or third-party code, continues to work without changes.

See the transfer tool code: MiddlewareAttributesFixture.

Methods with "Zend" in the name

In several places, we have defined method names containing the word "Zend". We provided a solution here mimicing what we did for namespaced functions: we renamed the existing function using "Laminas" in the name, and then added a new function using the old name that proxies to the original. For an example, you can review the Psr7Response class.

See the transfer tool code: SourceFixture.

Still not everything

The above sections detail the various common problems we encountered across multiple repositories. Unfortunately, we encountered many edge cases, and ended up with custom rules for 30 components.

One example worth noting: we needed to keep all references to the Zend Server product, but change references to the zend-server component (which has nothing to do with the Zend Server product). See the ZendServerDisk class and ZendMonitor class for examples.

As another example, "Zend" or "Expressive" are used as subnamespace names in several third-party libraries (see container-auryn), and we needed to ensure these were left unchanged.

How we tested

Before we were able to launch, we needed to test that everything would work. We tried the rewrite tool on multiple libraries, and they looked fine - but we needed something more useful than looking over the code.

Rewrite packages in vendor

The first approach we took was to rewrite all packages under the vendor/ directory of a ZF-based project. We added a command to our transfer tool to do this, and had some immediate successes.The approach gave us confidence that what we had written would likely work.

The problem, however, is that it didn't allow us to test every possible component or combination. We needed something more robust.

See the transfer tool code: VendorCommand.

Local testing

The second idea was to rewrite all components locally, create a local Composer repository from them, add this repository to each component, and then install dependencies and run tests for each.

This worked well, and gave us a lot of useful information. It helped us to find some failing cases, but we felt it still was not sufficient: we were generally testing only against one PHP version; we were only testing against the latest versions of dependencies; there were many a variety of configuration issue, and we were not able to run all tests due to missing dependencies, extensions, and/or services (e.g., mongodb, database, swoole, ...).

See the transfer tool code: LocalTestCommand.

Composer repository

The next thing we did was to rewrite all tags of all components we planned to migrate, and use these to create a proper Composer repository using Satis that we could expose publicly. This would allow us to test any project or library against any version available by adding the repository to the package configuration.

At this point, we opened testing to the public, and asked the community to help. We released the very first version of Laminas Migration Tool and started testing ZF-based projects.

During this phase, we identified and resolved a number of edge cases we would never have found otherwise.

But it was still not enough.

Running unit tests on each component

As a final effort, we decided to do full continuous integration on each component.

(You would probably think that we should have started from that, but you'd be wrong! We couldn't take this step until we had a public Composer repository.)

To do this, we created test organisations for each of our projects, pushed all components to these organisations, and enabled Travis CI on each.

We modified the configuration to also run the php linter on the source code, as our rewrite tool was heavily using regular expressions. This helped identify some edge cases with rewriting, but also reported a lot of false-positives. As one example of a false positive, we have a number of classes that are enabled only under specific PHP versions. We also have classes named after PHP keywords that were later reserved; in those cases, we have replacements, but the linter would flag the legacy classes as invalid.

We anticipated this phase would take quite some time, due to the fact that Travis CI for open source limits the number of parallel operations that can be run for any given account. Considering we were testing over 150 repositories on each of PHP 5.6, 7.0, 7.1, 7.2, and 7.3 (and also some on 7.4!), each against both lowest and latest dependencies, the number of jobs was enormous! In reality, this went much faster than expected, but nevertheless, we were often fixing issues during the day, and waiting for tests to run overnight.

This approach allowed us to identify a ton of issues, and we ended up rebuilding repositories and our Composer repository multiple times during the process, until we were satisfied with the results. In the end, the remaining issues we had were cases of test expectations that needed to change due to renamed classes and/or configuration, and a few minor ones that we were unable to reproduce anywhere except on the Travis CI platform itself.

Not everything was perfect

While we were constantly improving the tooling, we knew that it would never likely be perfect, and we would have issues to resolve after the migration was over.

For example we dropped Code Style checks in our tests as many of them were failing due to line lengths.

We were not able to rewrite images. Regenerating all images just to keep references to new libraries was not possible at this time, so we decided to leave them as "post-migration" manual operations.

Nevertheless, we used the migration tooling as a chance to resolve a number of long-standing issues. To list a few:

  1. Alphabetising import statements (as namespaces were changed we, want to keep alphabetical order). We were able to do this via a PHP Code Sniffer on rewritten PHP files.

  2. In some old version of packages, we were using grouped import statements so we ran another CS fix to split them.

  3. Copyright headers. We have completely changed copyright headers in all of the files. Before, we kept the copyright year as part of the header, and it was inconsistently updated (we had some rules when the year should be updated, but very often we forgot to do it). Now we have a much simpler copyright header with references to other files in the repository (LICENSE.md and COPYRIGHT.md).

  4. Updating spacing to follow PSR-12. As the PSR-12 Coding Standard was already approved, we decided to add a blank line after the opening <?php tag so we wouldn't need to to it later on.

  5. Multiple QA unification and improvements: Travis CI configuration, entries in .gitattributes, .gitignore, PHPUnit and PHPSpec configurations, etc. were all made consistent.

  6. Documentation and its configuration unification. Some documentation was under the legacy doc/ subdirectory, while others used docs/ (which is a recommended path for GitHub community support documents as well). The mkdocs.yml file used to govern how documentation is rendered has also been updated over time. We used this as a chance to make these consistent across all repositories.

  7. Consistent Github templates for Pull Requests and Issues. In fact, we ended up moving these into organization-level .github repositories, so they can be updated all at once, instead of having to update them across all repositories.

  8. Updated all support files (e.g., CONTRIBUTING.md, CODE_OF_CONDUCT.md, SUPPORT.md).

Finally: the transition

Finally, after over a year from the first announcement, and 10 months of working on the transfer tool, we decided we were ready to launch Laminas, and we chose to do so on the last day of 2019.

The days and weeks before, normally holidays for most of us, were spent polishing and prepping the tools for the transfer. We even had some "last minute changes" which surprised us and prevented us from starting as early as we wanted, but managed to migrate everything by around noon UTC time.

We knew not everything was perfect, and there was still a lot left to do, but we'd managed to deliver what we'd been promising: we'd deprecated and archived all Zend Framework repositories, and created all new components under three brand new organisations: Laminas, Mezzio and Laminas API Tools.

We were right: not everything was perfect

Shortly after completing the migration, we received user reports of issues.

The first and most serious issue was with namespaced function. We somehow missed including a "return" statement in the legacy functions when calling then new variants.

Because of that we had to issue patch versions for:

We noticed also that some versions of laminas-view had not been transferred correctly (all tags between 2.2.4 and 2.5.3) and we have also issued patch version for these.

The Laminas API Tool skeleton application had an incorrect module registered (due to renaming from ZendDeveloperTools to Laminas\DeveloperTools), and it, too had to be patched.

Patched version

In the above repositories we released new tags with the p1 suffix, which denotes a "patch" or "build" version. Composer will prefer these over the tag they patch when performing a composer update operation.

Thankfully that was everything. We've had no blocking issues reported since late January, while we continue to get reports of successful migrations.

In the last two months, we've also seen many third party repositories migrate to Laminas.

If you still have not updated your application or your company is still using Zend Framework components, we recommend you migrate, so that you continue to get security updates. Please see our migration guide.

Life after life. What next?

We'e had our first Technical Steering Committee meeting, and started planning how we want to maintain and expand the project.

Please follow us on Twitter, join our chat, and visit our forums today!

Stay tuned!