docs/injection.txt

   1 injection.txt
   2
   3 This is an overview of how MediaWiki makes use of dependency injection.
   4 The design described here grew from the discussion of RFC T384.
   5
   6
   7 The term "dependency injection" (DI) refers to a pattern on object oriented
   8 programming that tries to improve modularity by reducing strong coupling
   9 between classes. In practical terms, this means that anything an object needs
  10 to operate should be injected from the outside, the object itself should only
  11 know narrow interfaces, no concrete implementation of the logic it relies on.
  12
  13 The requirement to inject everything typically results in an architecture that
  14 based on two main types of objects: simple value objects with no business logic
  15 (and often immutable), and essentially stateless service objects that use
  16 other service objects to operate on the value objects.
  17
  18 As of the beginning of 2016 (MW version 1.27), MediaWiki is only starting to
  19 use the DI approach. Much of the code still relies on global state or direct
  20 instantiation, resulting in a highly cyclical dependency graph.
  21
  22
  23 == Overview ==
  24 The heart of the DI in MediaWiki is the central service locator,
  25 MediaWikiServices, which acts as the top level factory for services in
  26 MediaWiki. MediaWikiServices::getInstance() returns the default service
  27 locator instance, which can be used to gain access to default instances of
  28 various services. MediaWikiServices however also allows new services to be
  29 defined and default services to be redefined. Services are defined or
  30 redefined by providing a callback function, the "instantiator" function,
  31 that will return a new instance of the service.
  32
  33 When MediaWikiServices::getInstance() is first called, it will create an
  34 instance of MediaWikiServices and populate it with the services defined
  35 in the files listed by $wgServiceWiringFiles, thereby "bootstrapping" the
  36 DI framework. Per default, $wgServiceWiringFiles lists
  37 includes/ServiceWiring.php, which defines all default service
  38 implementations, and specifies how they depend on each other ("wiring").
  39
  40 When a new service is added to MediaWiki core, an instantiator function
  41 that will create the appropriate default instance for that service must
  42 be added to ServiceWiring.php. This makes the service available through
  43 the generic getService() method on the service locator returned by
  44 MediaWikiServices::getInstance().
  45
  46 Extensions can add their own wiring files to $wgServiceWiringFiles, in order
  47 to define their own service. Extensions may also use the 'MediaWikiServices'
  48 hook to define or redefined services by calling methods on the default
  49 MediaWikiServices instance.
  50
  51
  52 It should be noted that the term "service locator" is often used to refer to a
  53 top level factory that is accessed directly, throughout the code, to avoid
  54 explicit dependency injection. In contrast, the term "DI container" is often
  55 used to describe a top level factory that is only accessed when services
  56 are created. We use the term "service locator" for the top level factory
  57 because it is more descriptive than "DI container", even though application
  58 logic is strongly discouraged from accessing MediaWikiServices directly.
  59 MediaWikiServices::getInstance() should ideally be accessed only in "static
  60 entry points" such as hook handler functions. See "Migration" below.
  61
  62
  63 == Service Reset ==
  64
  65 Services get their configuration injected, and changes to global
  66 configuration variables will not have any effect on services that were already
  67 instantiated. This would typically be the case for low level services like
  68 the ConfigFactory or the ObjectCacheManager, which are used during extension
  69 registration. To address this issue, Setup.php resets the global service
  70 locator instance by calling MediaWikiServices::resetGlobalInstance() once
  71 configuration and extension registration is complete.
  72
  73 Note that "unmanaged" legacy services services that manage their own singleton
  74 must not keep references to services managed by MediaWikiServices, to allow a
  75 clean reset. After the global MediaWikiServices instance got reset, any such
  76 references would be stale, and using a stale service will result in an error.
  77
  78 Services should either have all dependencies injected and be themselves managed
  79 by MediaWikiServices, or they should use the Service Locator pattern, accessing
  80 service instances via the global MediaWikiServices instance state when needed.
  81 This ensures that no stale service references remain after a reset.
  82
  83
  84 == Configuration ==
  85
  86 When the default MediaWikiServices instance is created, a Config object is
  87 provided to the constructor. This Config object represents the "bootstrap"
  88 configuration which will become available as the 'BootstrapConfig' service.
  89 As of MW 1.27, the bootstrap config is a GlobalVarConfig object providing
  90 access to the $wgXxx configuration variables.
  91
  92 The bootstrap config is then used to construct a 'ConfigFactory' service,
  93 which in turn is used to construct the 'MainConfig' service. Application
  94 logic should use the 'MainConfig' service (or a more specific configuration
  95 object). 'BootstrapConfig' should only be used for bootstrapping basic
  96 services that are needed to load the 'MainConfig'.
  97
  98
  99 Note: Several well known services in MediaWiki core act as factories
 100 themselves, e.g. ApiModuleManager, ObjectCache, SpecialPageFactory, etc.
 101 The registries these factories are based on are currently managed as part of
 102 the configuration. This may however change in the future.
 103
 104
 105 == Migration ==
 106
 107 This section provides some recipes for improving code modularity by reducing
 108 strong coupling. The dependency injection mechanism described above is an
 109 essential tool in this effort.
 110
 111 Migrate access to global service instances and config variables:
 112 Assume Foo is a class that uses the $wgScriptPath global and calls
 113 wfGetDB() to get a database connection, in non-static methods.
 114 * Add $scriptPath as a constructor parameter and use $this->scriptPath
 115   instead of $wgScriptPath.
 116 * Add LoadBalancer $dbLoadBalancer as a constructor parameter. Use
 117   $this->dbLoadBalancer->getConnection() instead of wfGetDB().
 118 * Any code that calls Foo's constructor would now need to provide the
 119   $scriptPath and $dbLoadBalancer. To avoid this, avoid direct instantiation
 120   of services all together - see below.
 121
 122 Migrate class-level singleton getters:
 123 Assume class Foo has mostly non-static methods, and provides a static
 124 getInstance() method that returns a singleton (or default instance).
 125 * Add an instantiator function for Foo into ServiceWiring.php. The instantiator
 126   would do exactly what Foo::getInstance() did. However, it should
 127   replace any access to global state with calls to $services->getXxx() to get a
 128   service, or $services->getMainConfig()->get() to get a configuration setting.
 129 * Add a getFoo() method to MediaWikiServices. Don't forget to add the
 130   appropriate test cases in MediaWikiServicesTest.
 131 * Turn Foo::getInstance() into a deprecated alias for
 132   MediaWikiServices::getInstance()->getFoo(). Change all calls to
 133   Foo::getInstance() to use injection (see above).
 134
 135 Migrate direct service instantiation:
 136 Assume class Bar calls new Foo().
 137 * Add an instantiator function for Foo into ServiceWiring.php and add a getFoo()
 138   method to MediaWikiServices. Don't forget to add the appropriate test cases
 139   in MediaWikiServicesTest.
 140 * In the instantiator, replace any access to global state with calls
 141   to $services->getXxx() to get a service, or $services->getMainConfig()->get()
 142   to get a configuration setting.
 143 * The code in Bar that calls Foo's constructor should be changed to have a Foo
 144   instance injected; Eventually, the only code that instantiates Foo is the
 145   instantiator in ServiceWiring.php.
 146 * As an intermediate step, Bar's constructor could initialize the $foo member
 147   variable by calling MediaWikiServices::getInstance()->getFoo(). This is
 148   acceptable as a stepping stone, but should be replaced by proper injection
 149   via a constructor argument. Do not however inject the MediaWikiServices
 150   object!
 151
 152 Migrate parameterized helper instantiation:
 153 Assume class Bar creates some helper object by calling new Foo( $x ),
 154 and Foo uses a global singleton of the Xyzzy service.
 155 * Define a FooFactory class (or a FooFactory interface along with a MyFooFactory
 156   implementation). FooFactory defines the method newFoo( $x ) or getFoo( $x ),
 157   depending on the desired semantics (newFoo would guarantee a fresh instance).
 158   When Foo gets refactored to have Xyzzy injected, FooFactory will need a
 159   Xyzzy instance, so newFoo() can pass it to new Foo().
 160 * Add an instantiator function for FooFactory into ServiceWiring.php and add a
 161   getFooFactory() method to MediaWikiServices. Don't forget to add the
 162   appropriate test cases in MediaWikiServicesTest.
 163 * The code in Bar that calls Foo's constructor should be changed to have a
 164   FooFactory instance injected; Eventually, the only code that instantiates
 165   Foo are implementations of FooFactory, and the only code that instantiates
 166   FooFactory is the instantiator in ServiceWiring.php.
 167 * As an intermediate step, Bar's constructor could initialize the $fooFactory
 168   member variable by calling MediaWikiServices::getInstance()->getFooFactory().
 169   This is acceptable as a stepping stone, but should be replaced by proper
 170   injection via a constructor argument. Do not however inject the
 171   MediaWikiServices object!
 172
 173 Migrate a handler registry:
 174 Assume class Bar calls FooRegistry::getFoo( $x ) to get a specialized Foo
 175 instance for handling $x.
 176 * Turn getFoo into a non-static method.
 177 * Add an instantiator function for FooRegistry into ServiceWiring.php and add
 178   a getFooRegistry() method to MediaWikiServices. Don't forget to add the
 179   appropriate test cases in MediaWikiServicesTest.
 180 * Change all code that calls FooRegistry::getFoo() statically to call this
 181   method on a FooRegistry instance. That is, Bar would have a $fooRegistry
 182   member, initialized from a constructor parameter.
 183 * As an intermediate step, Bar's constructor could initialize the $fooRegistry
 184   member variable by calling MediaWikiServices::getInstance()->
 185   getFooRegistry(). This is acceptable as a stepping stone, but should be
 186   replaced by proper injection via a constructor argument. Do not however
 187   inject the MediaWikiServices object!
 188
 189 Migrate deferred service instantiation:
 190 Assume class Bar calls new Foo(), but only when needed, to avoid the cost of
 191 instantiating Foo().
 192 * Define a FooFactory interface and a MyFooFactory implementation of that
 193   interface. FooFactory defines the method getFoo() with no parameters.
 194 * Precede as for the "parameterized helper instantiation" case described above.
 195
 196 Migrate a class with only static methods:
 197 Assume Foo is a class with only static methods, such as frob(), which
 198 interacts with global state or system resources.
 199 * Introduce a FooService interface and a DefaultFoo implementation of that
 200   interface. FooService contains the public methods defined by Foo.
 201 * Add an instantiator function for FooService into ServiceWiring.php and
 202   add a getFooService() method to MediaWikiServices. Don't forget to
 203   add the appropriate test cases in MediaWikiServicesTest.
 204 * Add a private static getFooService() method to Foo. That method just
 205   calls MediaWikiServices::getInstance()->getFooService().
 206 * Make all methods in Foo delegate to the FooService returned by
 207   getFooService(). That is, Foo::frob() would do self::getFooService()->frob().
 208 * Deprecate Foo. Inject a FooService into all code that calls methods
 209   on Foo, and change any calls to static methods in foo to the methods
 210   provided by the FooService interface.
 211
 212 Migrate static hook handler functions (to allow unit testing):
 213 Assume MyExtHooks::onFoo is a static hook handler function that is called with
 214 the parameter $x; Further assume MyExt::onFoo needs service Bar, which is
 215 already known to MediaWikiServices (if not, see above).
 216 * Create a non-static doFoo( $x ) method in MyExtHooks that has the same
 217   signature as onFoo( $x ). Move the code from onFoo() into doFoo(), replacing
 218   any access to global or static variables with access to instance member
 219   variables.
 220 * Add a constructor to MyExtHooks that takes a Bar service as a parameter.
 221 * Add a static method called newFromGlobalState() with no parameters. It should
 222   just return new MyExtHooks( MediaWikiServices::getInstance()->getBar() ).
 223 * The original static handler method onFoo( $x ) is then implemented as
 224   self::newFromGlobalState()->doFoo( $x ).
 225
 226 Migrate a "smart record":
 227 Assume Thingy is a "smart record" that "knows" how to load and store itself.
 228 For this purpose, Thingy uses wfGetDB().
 229 * Create a "dumb" value class ThingyRecord that contains all the information
 230   that Thingy represents (e.g. the information from a database row). The value
 231   object should not know about any service.
 232 * Create a DAO-style service for loading and storing ThingyRecords, called
 233   ThingyStore. It may be useful to split the interfaces for reading and
 234   writing, with a single class implementing both interfaces, so we in the
 235   end have the ThingyLookup and ThingyStore interfaces, and a SqlThingyStore
 236   implementation.
 237 * Add instantiator functions for ThingyLookup and ThingyStore in
 238   ServiceWiring.php. Since we want to use the same instance for both service
 239   interfaces, the instantiator for ThingyLookup would return
 240   $services->getThingyStore().
 241 * Add getThingyLookup() and getThingyStore methods to MediaWikiServices.
 242   Don't forget to add the appropriate test cases in MediaWikiServicesTest.
 243 * In the old Thingy class, replace all member variables that represent the
 244   record's data with a single ThingyRecord object.
 245 * In the old Thingy class, replace all calls to static methods or functions,
 246   such as wfGetDB(), with calls to the appropriate services, such as
 247   LoadBalancer::getConnection().
 248 * In Thingy's constructor, pull in any services needed, such as the
 249   LoadBalancer, by using MediaWikiServices::getInstance(). These services
 250   cannot be injected without changing the constructor signature, which
 251   is often impractical for "smart records" that get instantiated directly
 252   in many places in the code base.
 253 * Deprecate the old Thingy class. Replace all usages of it with one of the
 254   three new classes: loading needs a ThingyLookup, storing needs a ThingyStore,
 255   and reading data needs a ThingyRecord.
 256
 257 Migrate lazy loading:
 258 Assume Thingy is a "smart record" as described above, but requires lazy loading
 259 of some or all the data it represents.
 260 * Instead of a plain object, define ThingyRecord to be an interface. Provide a
 261   "simple" and "lazy" implementations, called SimpleThingyRecord and
 262   LazyThingyRecord. LazyThingyRecord knows about some lower level storage
 263   interface, like a LoadBalancer, and uses it to load information on demand.
 264 * Any direct instantiation of a ThingyRecord would use the SimpleThingyRecord
 265   implementation.
 266 * SqlThingyStore however creates instances of LazyThingyRecord, and injects
 267   whatever storage layer service LazyThingyRecord needs to perform lazy loading.
 268
 269