spring5的那些事：@Indexed 解密

随着云原生的发展，很多技术会被重新掂量，重新定义，历来技术的发展也是遵循天时地利，以其势尽享其利。再云原生下，jdk的最大的问题在于笨重（几百mb），启动慢，而像Serverless架构，NodeJS技术栈可谓更完美。其实在jdk9中倡导模块化本质在于减少JVM的体积，不需要资源（Jar）不用再加载，而启动慢的问题其实也有解决方案GraalVM （一款类似于HotSpot VM），它的先进之处在于缩短运行的成本将.java文件直接编译成native code，而jvm则多了一个环节，首先将.java文件编译成字节码(.class)，再借助JVM运行时JIT技术编译成native code。

spring5.0开始支持@Indexed来提升进应用启动速度，通过Annotation Processing Tools API在编译时来构建索引文件，本质是通过静态化来解决启动时Bean扫描加载的时间长的问题。

what is Annotation Processing Tools API？

不是什么黑科技，之前的系列也讲过，有点类似lombok。

哪些资源会被索引？

默认支持标记为Component及其派生注解（Controller、Repository、Service、Configuration等）的类，当然也可以是非spring bean（@Indexed修饰的类）。

注：如果已经是spring bean（Component修饰的类，并且Component已经被标记为@Indexed）了就没必要再标记@Indexed，否则索引文件会再追加一个相同的，感觉这是个bug

如何使用？

使用非常讲的，添加依赖就可以了，install后默认会生成一个META-INF/spring.components。

<dependency>
    <groupId>org.springframework</groupId>
    <artifactId>spring-context-indexer</artifactId>
    <optional>true</optional>
</dependency>

#spring.components
com.yh.rfe.lucky.day.service.impl.BasCostReportServiceImpl=org.springframework.stereotype.Component
com.yh.rfe.lucky.day.service.impl.BasShopRuleDetailServiceImpl=org.springframework.stereotype.Component

而CandidateComponentsIndexer负责对符合条件的注解生成索引文件，整个源码也不是特别复杂，通过三个组件：StereotypesProvider、MetadataCollector、MetadataStore来完成。

public class CandidateComponentsIndexer implements Processor {
	@Override
	public synchronized void init(ProcessingEnvironment env) {
		this.stereotypesProviders = getStereotypesProviders(env);
		this.typeHelper = new TypeHelper(env);
		this.metadataStore = new MetadataStore(env);
		this.metadataCollector = new MetadataCollector(env, this.metadataStore.readMetadata());
	}

	@Override
	public boolean process(Set<? extends TypeElement> annotations, RoundEnvironment roundEnv) {
		this.metadataCollector.processing(roundEnv);
		roundEnv.getRootElements().forEach(this::processElement);
		if (roundEnv.processingOver()) {
			writeMetaData();
		}
		return false;
	}
}
//定义了哪些注解需要被索引
interface StereotypesProvider {
	/**
	 * Return the stereotypes that are present on the given {@link Element}.
	 * @param element the element to handle
	 * @return the stereotypes or an empty set if none were found
	 */
	Set<String> getStereotypes(Element element);
}
//获取需要被索引的CandidateComponentsMetadata（元数据）
class MetadataCollector {
	public CandidateComponentsMetadata getMetadata() {
		CandidateComponentsMetadata metadata = new CandidateComponentsMetadata();
		for (ItemMetadata item : this.metadataItems) {
			metadata.add(item);
		}
		if (this.previousMetadata != null) {
			List<ItemMetadata> items = this.previousMetadata.getItems();
			for (ItemMetadata item : items) {
				if (shouldBeMerged(item)) {
					metadata.add(item);
				}
			}
		}
		return metadata;
	}    
}
//将上面的结果输出到spring.components中
class MetadataStore {
	static final String METADATA_PATH = "META-INF/spring.components";    
	public void writeMetadata(CandidateComponentsMetadata metadata) throws IOException {
		if (!metadata.getItems().isEmpty()) {
			try (OutputStream outputStream = createMetadataResource().openOutputStream()) {
				PropertiesMarshaller.write(metadata, outputStream);
			}
		}
	}    
}    

原理

其实在spring boot项目中绝对存在ComponentScan(在SpringBootApplication中)，而传统的spring项目中xml中对应<context:component-scan>，通过指定的 package（路径）来扫描注入spring bean，在扫描时通过读取spring.components文件来读取class（类全路径）从而达到提升速度的目的。

CandidateComponentsIndex存储了spring.components文件的内容

public class CandidateComponentsIndex {
	private static final AntPathMatcher pathMatcher = new AntPathMatcher(".");
	private final MultiValueMap<String, Entry> index;
	/*返回指定的注解类型和包路径相关候选类型
	* Set<String> candidates = index.getCandidateTypes("com.example", "org.springframework.stereotype.Component");
	*/
	public Set<String> getCandidateTypes(String basePackage, String stereotype) {
		List<Entry> candidates = this.index.get(stereotype);
		if (candidates != null) {
			return candidates.parallelStream()
					.filter(t -> t.match(basePackage))
					.map(t -> t.type)
					.collect(Collectors.toSet());
		}
		return Collections.emptySet();
	}    
}

CandidateComponentsIndexLoader从classloader中读取，可以从多个jar中读取多个索引文件。

public final class CandidateComponentsIndexLoader {
	public static final String COMPONENTS_RESOURCE_LOCATION = "META-INF/spring.components";
	private static final ConcurrentMap<ClassLoader, CandidateComponentsIndex> cache =
			new ConcurrentReferenceHashMap<>();
	@Nullable
	public static CandidateComponentsIndex loadIndex(@Nullable ClassLoader classLoader) {
		ClassLoader classLoaderToUse = classLoader;
		if (classLoaderToUse == null) {
			classLoaderToUse = CandidateComponentsIndexLoader.class.getClassLoader();
		}
		return cache.computeIfAbsent(classLoaderToUse, CandidateComponentsIndexLoader::doLoadIndex);
	}
	@Nullable
	private static CandidateComponentsIndex doLoadIndex(ClassLoader classLoader) {
		if (shouldIgnoreIndex) {
			return null;
		}
		try {
			Enumeration<URL> urls = classLoader.getResources(COMPONENTS_RESOURCE_LOCATION);
			if (!urls.hasMoreElements()) {
				return null;
			}
			List<Properties> result = new ArrayList<>();
			while (urls.hasMoreElements()) {
				URL url = urls.nextElement();
				Properties properties = PropertiesLoaderUtils.loadProperties(new UrlResource(url));
				result.add(properties);
			}
			if (logger.isDebugEnabled()) {
				logger.debug("Loaded " + result.size() + "] index(es)");
			}
			int totalCount = result.stream().mapToInt(Properties::size).sum();
			return (totalCount > 0 ? new CandidateComponentsIndex(result) : null);
		}
		catch (IOException ex) {
			throw new IllegalStateException("Unable to load indexes from location [" +
					COMPONENTS_RESOURCE_LOCATION + "]", ex);
		}
	}
}    

ClassPathBeanDefinitionScanner非常重要，它就是spring 中scan时干最脏最累的活的终结者。而ClassPathScanningCandidateComponentProvider非常重要可以视为scan的顶级实现类。

其中ClassPathMapperScanner是mybatis的mapper扫描类。

public class ClassPathBeanDefinitionScanner extends ClassPathScanningCandidateComponentProvider {
	public int scan(String... basePackages) {
		int beanCountAtScanStart = this.registry.getBeanDefinitionCount();

		doScan(basePackages);

		// Register annotation config processors, if necessary.
		if (this.includeAnnotationConfig) {
			AnnotationConfigUtils.registerAnnotationConfigProcessors(this.registry);
		}
		return (this.registry.getBeanDefinitionCount() - beanCountAtScanStart);
	}    
	protected Set<BeanDefinitionHolder> doScan(String... basePackages) {
		Assert.notEmpty(basePackages, "At least one base package must be specified");
		Set<BeanDefinitionHolder> beanDefinitions = new LinkedHashSet<>();
		for (String basePackage : basePackages) {
			Set<BeanDefinition> candidates = findCandidateComponents(basePackage);//看这里吧
			for (BeanDefinition candidate : candidates) {
				ScopeMetadata scopeMetadata = this.scopeMetadataResolver.resolveScopeMetadata(candidate);
				candidate.setScope(scopeMetadata.getScopeName());
				String beanName = this.beanNameGenerator.generateBeanName(candidate, this.registry);
				if (candidate instanceof AbstractBeanDefinition) {
					postProcessBeanDefinition((AbstractBeanDefinition) candidate, beanName);
				}
				if (candidate instanceof AnnotatedBeanDefinition) {
					AnnotationConfigUtils.processCommonDefinitionAnnotations((AnnotatedBeanDefinition) candidate);
				}
				if (checkCandidate(beanName, candidate)) {
					BeanDefinitionHolder definitionHolder = new BeanDefinitionHolder(candidate, beanName);
					definitionHolder =
							AnnotationConfigUtils.applyScopedProxyMode(scopeMetadata, definitionHolder, this.registry);
					beanDefinitions.add(definitionHolder);
					registerBeanDefinition(definitionHolder, this.registry);
				}
			}
		}
		return beanDefinitions;
	}
}
public class ClassPathScanningCandidateComponentProvider implements EnvironmentCapable, ResourceLoaderAware {
	private MetadataReaderFactory metadataReaderFactory;//这个之前讲过类元数据读取
	private CandidateComponentsIndex componentsIndex;//前面讲过
	public Set<BeanDefinition> findCandidateComponents(String basePackage) {
		if (this.componentsIndex != null && indexSupportsIncludeFilters()) {
			return addCandidateComponentsFromIndex(this.componentsIndex, basePackage);
		}
		else {
			return scanCandidateComponents(basePackage);
		}
	}
	private Set<BeanDefinition> addCandidateComponentsFromIndex(CandidateComponentsIndex index, String basePackage) {
		Set<BeanDefinition> candidates = new LinkedHashSet<>();
		try {
			Set<String> types = new HashSet<>();
			for (TypeFilter filter : this.includeFilters) {
				String stereotype = extractStereotype(filter);
				if (stereotype == null) {
					throw new IllegalArgumentException("Failed to extract stereotype from "+ filter);
				}
				types.addAll(index.getCandidateTypes(basePackage, stereotype));
			}
			boolean traceEnabled = logger.isTraceEnabled();
			boolean debugEnabled = logger.isDebugEnabled();
			for (String type : types) {
				MetadataReader metadataReader = getMetadataReaderFactory().getMetadataReader(type);
				if (isCandidateComponent(metadataReader)) {
					AnnotatedGenericBeanDefinition sbd = new AnnotatedGenericBeanDefinition(
							metadataReader.getAnnotationMetadata());
					if (isCandidateComponent(sbd)) {
						if (debugEnabled) {
							logger.debug("Using candidate component class from index: " + type);
						}
						candidates.add(sbd);
					}
					else {
						if (debugEnabled) {
							logger.debug("Ignored because not a concrete top-level class: " + type);
						}
					}
				}
				else {
					if (traceEnabled) {
						logger.trace("Ignored because matching an exclude filter: " + type);
					}
				}
			}
		}
		catch (IOException ex) {
			throw new BeanDefinitionStoreException("I/O failure during classpath scanning", ex);
		}
		return candidates;
	}    
}

AnnotationConfigApplicationContext#scan你一定不陌生吧，这可是开发用户级的API，其实它的scanner就是ClassPathBeanDefinitionScanner

public class AnnotationConfigApplicationContext extends GenericApplicationContext implements AnnotationConfigRegistry {
	private final AnnotatedBeanDefinitionReader reader;
	private final ClassPathBeanDefinitionScanner scanner;
	public AnnotationConfigApplicationContext() {
		this.reader = new AnnotatedBeanDefinitionReader(this);
		this.scanner = new ClassPathBeanDefinitionScanner(this);
	}
	public AnnotationConfigApplicationContext(String... basePackages) {
		this();
		scan(basePackages);
		refresh();
	}  
	public void register(Class<?>... annotatedClasses) {
		Assert.notEmpty(annotatedClasses, "At least one annotated class must be specified");
		this.reader.register(annotatedClasses);
	}
}   

其实关于@Indexed个人觉得实现上还是有一定局限性（只是针对当前maven的一个module，换言之是基于jar的），要基于当前整个工程文件特别是org.springframework包（这个下面有很多待加载到ioc的bean的jar）工作量还是不少的，官方还没考虑吧。