[
https://issues.apache.org/jira/browse/LOG4J2-2803?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17358216#comment-17358216
]
Matt Sicker edited comment on LOG4J2-2803 at 6/6/21, 8:23 PM:
--------------------------------------------------------------
Yes, the idea is to keep backward compatibility for configuration files and
existing plugins generally speaking. As for using an existing framework, we're
too low in the stack to rely on one (plus it would add a dependency as Ralph
explained). Also, most (if not all) existing frameworks are far more powerful,
flexible, and large than we need. The DI framework I started last year is based
on the business logic from Weld, OpenWebBeans (both CDI implementations), and
Guice (a similar framework but using a different API), but without all the
JavaEE stuff, no observers, no generic event system, no interceptors, no
decorators, and no other fancy DI features. The goal here is to simplify our
own code while making it easier to write plugins that need to make use of more
complex programmatic configuration (e.g., defining your own database connection
beans to add to your config).
Now of course, if someone knows of an existing DI framework that is
ALv2-compatible that does this, I'd love to just shade it in like we do with
picocli, but I don't think such a thing exists. I came across
[https://pf4j.org|https://pf4j.org/] in Spinnaker as a general plugin
framework, but even that is a bit heavy for our use case (it offers things like
ClassLoader plugin isolation like how Jenkins plugins work or web application
containers do). Plus, like I said about the stack, even pf4j's minimal
dependency footprint includes... a Logging API.
was (Author: jvz):
Yes, the idea is to keep backward compatibility for configuration files and
existing plugins generally speaking. As for using an existing framework, we're
too low in the stack to rely on one (plus it would add a dependency as Ralph
explained). Also, most (if not all) existing frameworks are far more powerful,
flexible, and large than we need. The DI framework I started last year is based
on the business logic from Weld, OpenWebBeans (both CDI implementations), and
Guice (a similar framework but using a different API), but without all the
JavaEE stuff, no observers, no generic event system, no interceptors, no
decorators, and no other fancy DI features. The goal here is to simplify our
own code while making it easier to write plugins that need to make use of more
complex programmatic configuration (e.g., defining your own database connection
beans to add to your config).
Now of course, if someone knows of an existing DI framework that is
ALv2-compatible that does this, I'd love to just shade it in like we do with
picocli, but I don't think such a thing exists. I came across
[https://pf4j.org|https://pf4j.org/] in Spinnaker as a general plugin
framework, but even that is a bit heavy for our use case (it offers things like
ClassLoader plugin isolation like how Jenkins plugins work or web application
containers do).
> Create standardized scopes and dependency injection API
> -------------------------------------------------------
>
> Key: LOG4J2-2803
> URL: https://issues.apache.org/jira/browse/LOG4J2-2803
> Project: Log4j 2
> Issue Type: Epic
> Components: Configuration, Core, Plugins
> Reporter: Matt Sicker
> Assignee: Matt Sicker
> Priority: Major
> Fix For: 3.0.0
>
>
> h2. Context
> The existing plugin system revolves around {{@Plugin}} annotations which
> group together configurable interfaces into categories. The category of a
> plugin typically determines the configuration and instantiation strategy for
> the plugin classes involved which leads to some ad hoc methods for different
> plugins to obtain or configure different aspects of the system. All classes
> that are included in the plugin system can be queried at runtime by
> constructing a {{PluginManager}} with the category name as its constructor
> argument. To date, there are a few plugin categories (case-insensitive):
> * {{Core}}: used for configuration elements in config files for general core
> plugins like appenders, filters, layouts, etc. These are provided values from
> parsed configuration nodes and can also obtain the {{Configuration}} instance
> to access other services.
> * {{Level}}: used for specifying custom log levels so that they can be
> instantiated early enough for configurations to make use of it.
> * {{ConfigurationFactory}}: used for parsing and validating configuration
> sources in some format into a configuration node tree which is transformed
> into a {{Configuration}}. Also useful for programmatic configuration.
> * {{Lookup}}: used for variable interpolation in strings via {{StrLookup}}
> classes.
> * {{TypeConverter}}: used for converting strings into other objects. Mostly
> useful for conversion of plugin attributes into common types along with use
> in certain core plugins such as database column mappers.
> * {{Converter}}: used for pattern layout conversion keys. These are used for
> formatting log event info into a layout.
> * {{Watcher}}: used for watching configuration sources for changes in order
> to allow for reconfiguration.
> Then there are some system-wide plugin-like classes that can be specified via
> system properties including:
> * {{ClockFactory}}: controls {{Clock}} instances for obtaining the current
> time. Mostly useful in testing and scenarios where approximate time is more
> useful than exact time.
> * {{LogEventFactory}}: constructs {{LogEvent}} instances for a given log
> message and parameters. Useful for implementing different allocation
> strategies for log events.
> * {{MergeStrategy}}: used for composite configurations.
> * {{ContextSelector}}: used for obtaining a {{LoggerContext}} for an
> invocation context when initializing or obtaining {{Logger}} instances. By
> default, this maps contexts to class loaders, and additional strategies are
> available including async loggers, a singleton context, a JNDI-lookup-based
> context, and an OSGi bundle context.
> * {{LoggerContextFactory}}: used for binding a logging implementation to the
> Logging API.
> * {{ShutdownCallbackRegistry}}: used for registering a logging system
> cleanup shutdown callback in various systems. The default strategy hooks in
> to the {{Runtime}} shutdown threads API.
> * {{ContextDataInjector}}: used for adding initial context data into the
> {{ThreadContext}} of a {{LogEvent}}.
> h2. Proposal
> This system grew organically over time, and after identifying various
> duplicated plugin functionality around the codebase, the plugin system should
> be overhauled along the lines of a more standard {{@Inject}}-style dependency
> injection framework. By adopting more conventional DI patterns, this will
> remove the need for ad hoc plugin mechanisms in different subsystems, allow
> for more flexible declarations of plugins by specifying direct classes needed
> by plugins rather than doing manual lookups on the {{Configuration}}
> instance, allow for smarter initialization strategies to minimize startup
> time and resource usage based on configurations rather than eagerly loading
> plugins, make it simpler to write unit tests, and make it generally simpler
> to write and maintain plugins without requiring deep expertise of esoteric
> subsystems.
> Taking inspiration from [CDI|https://docs.jboss.org/cdi/api/2.0/],
> [@Inject|https://docs.oracle.com/javaee/6/api/javax/inject/Inject.html], and
> [Guice|https://github.com/google/guice], the plugin system should be updated
> to a new API that is broken up into a few useful concepts:
> * Beans: beans are analogous to Java classes with additional metadata to
> enable lifecycle management and dependency injection. Unlike a normal Java
> class, a bean has injection points (like constructor arguments, methods, and
> fields) where other managed instances can be injected at runtime based on the
> currently configured beans. As these are managed, they can have
> post-construct and pre-destruct callbacks for further lifecycle customization.
> ** Scopes: scopes control where the instances are stored or to what
> lifecycle they're tied to. At minimum, this includes dependent-scoped
> (creates new instances for every injection point) and singleton-scoped
> (creates and maintains a single instance for all injection points). In
> general DI frameworks, other scopes include things like HTTP sessions and
> HTTP requests. For this framework, the {{ContextSelector}} configuration API
> corresponds to a {{Configuration}} scope.
> ** Qualifiers: qualifiers add metadata to a bean besides its type to match
> on for injection. Typical qualifiers are names. Plugin configuration
> attribute names would likely make a good match here as another example.
> * Injection points: in order to obtain injectable values into a bean
> instance, injection points control how these instances are bound to the bean.
> This is essentially the strategy for binding data to fields, constructors,
> and methods.
> * Providers: providers are analogous to plugin builders in that they specify
> how to create an instance of type {{T}} given some plugin builder class
> {{Builder<T> implements Provider<T>}}.
> * Producer and disposer methods: beans can specify producer and disposer
> methods as alternative ways to specify how to construct a bean instance.
> These are analogous to plugin factory methods.
> Adopting such a plugin system based on beans, all the plugins that currently
> require either a {{Configuration}} or {{Node}} instance to do anything should
> be refactored to rely on proper dependency injection. This should be used to
> replace as many disparate dependency injection and configuration systems as
> possible to reduce the complexity required to implement plugins that rely on
> shared components.
> h2. Additional Details
>
> Beyond the dependency injection API itself, this epic is motivated by the
> following goals:
> * Simplify plugin classes that currently require manual method calls to
> {{Configuration}} or {{Node}}.
> * Unify the various ad hoc dependency injection for configuration-scoped or
> singleton-scoped classes (the latter which are overridden via system
> properties) including:
> ** {{StrSubstitutor}}
> ** {{ConfigurationScheduler}} (or scheduled tasks in general)
> ** {{ScriptManager}}
> ** {{WatchManager}}
> ** {{NanoClock}}
> ** {{ShutdownCallbackRegistry}}
> ** {{Clock}}
> ** {{ContextSelector}}
> ** {{ConfigurationFactory}}
> ** {{LogEventFactory}}
> ** {{ContextDataFactory}}
> * Make dependencies between classes more explicit via inversion of control
> which allows for easier testing and modularity.
> * Avoid the use of manual configuration handling at the API level of any
> plugins including basic string parsing (as currently supported through the
> {{TypeConverter}} API), class loading from strings, and ad hoc reflection.
> This relates to the changes made in LOG4J2-2621 and LOG4J2-1917.
> * Provide smart initialization logic to continue (or maintain) improvements
> to startup time and avoid loading plugins that aren't referenced in the
> loaded configuration.
> * Provide scopes for other ad hoc scopes/contexts used in various plugins
> such as HTTP request, HTTP session, servlet context, etc.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
