Nutch 2.0 之 抓取流程简单分析

Nutch 2.0 抓取流程介绍

---------------------

1. 整体流程

InjectorJob => GeneratorJob => FetcherJob => ParserJob => DbUpdaterJob => SolrIndexerJob

InjectorJob : 从文件中得到一批种子网页，把它们放到抓取数据库中去

GeneratorJob: 从抓取数据库中产生要抓取的页面放到抓取队列中去

FetcherJob: 对抓取队列中的网页进行抓取,在reducer中使用了生产/消费者模型

ParserJob: 对抓取完成的网页进行解析，产生一些新的链接与网页内容的解析结果

DbUpdaterJob: 把新产生的链接更新到抓取数据库中去

SolrIndexerJob: 对解析后的内容进行索引建立

2. InjectorJob分析

下面是InjectorJob的启动函数，代码如下

public Map<String,Object> run(Map<String,Object> args) throws Exception {
    getConf().setLong("injector.current.time", System.currentTimeMillis());
    Path input;
    Object path = args.get(Nutch.ARG_SEEDDIR);
    if (path instanceof Path) {
      input = (Path)path;
    } else {
      input = new Path(path.toString());
    }
    numJobs = 2;
    currentJobNum = 0;
    status.put(Nutch.STAT_PHASE, "convert input");
    currentJob = new NutchJob(getConf(), "inject-p1 " + input);
    FileInputFormat.addInputPath(currentJob, input);
	// mapper方法，从文件中解析出url，写入数据库
    currentJob.setMapperClass(UrlMapper.class);
    currentJob.setMapOutputKeyClass(String.class);
	// map 的输出为WebPage，它是用Gora compile生成的，可以通过Gora把它映射到不同的数据库中，
    currentJob.setMapOutputValueClass(WebPage.class);
	// 输出到GoraOutputFormat
    currentJob.setOutputFormatClass(GoraOutputFormat.class);
    DataStore<String, WebPage> store = StorageUtils.createWebStore(currentJob.getConfiguration(),
        String.class, WebPage.class);
    GoraOutputFormat.setOutput(currentJob, store, true);
    currentJob.setReducerClass(Reducer.class);
    currentJob.setNumReduceTasks(0);
    currentJob.waitForCompletion(true);
    ToolUtil.recordJobStatus(null, currentJob, results);
    currentJob = null;

    status.put(Nutch.STAT_PHASE, "merge input with db");
    status.put(Nutch.STAT_PROGRESS, 0.5f);
    currentJobNum = 1;
    currentJob = new NutchJob(getConf(), "inject-p2 " + input);
    StorageUtils.initMapperJob(currentJob, FIELDS, String.class,
        WebPage.class, InjectorMapper.class);
    currentJob.setNumReduceTasks(0);
    ToolUtil.recordJobStatus(null, currentJob, results);
    status.put(Nutch.STAT_PROGRESS, 1.0f);
    return results;
  }

因为InjectorJob扩展自NutchTool，实现了它的run方法。

我们可以看到，这里有两个MR任务，第一个主要是从文件中读入种子网页，写到DataStore数据库中，第二个MR任务主要是对数据库中的WebPage对象做一个分数与抓取间隔的设置。它使用到一个initMapperJob方法，代码如下

public static <K, V> void initMapperJob(Job job,
      Collection<WebPage.Field> fields,
      Class<K> outKeyClass, Class<V> outValueClass,
      Class<? extends GoraMapper<String, WebPage, K, V>> mapperClass,
      Class<? extends Partitioner<K, V>> partitionerClass, boolean reuseObjects)
  throws ClassNotFoundException, IOException {
	  // 这里是生成一个DataStore的抽象，这里的DataStore用户可以不同的模块，如Hbase，MySql等
    DataStore<String, WebPage> store = createWebStore(job.getConfiguration(),
        String.class, WebPage.class);
    if (store==null) throw new RuntimeException("Could not create datastore");
    Query<String, WebPage> query = store.newQuery();
    query.setFields(toStringArray(fields));
    GoraMapper.initMapperJob(job, query, store,
        outKeyClass, outValueClass, mapperClass, partitionerClass, reuseObjects);
    GoraOutputFormat.setOutput(job, store, true);
  }

3. GeneratorJob 源代码分析

下面是GeneratorJob的run方法代码

public Map<String,Object> run(Map<String,Object> args) throws Exception {
    // map to inverted subset due for fetch, sort by score
    Long topN = (Long)args.get(Nutch.ARG_TOPN);
    Long curTime = (Long)args.get(Nutch.ARG_CURTIME);
    if (curTime == null) {
      curTime = System.currentTimeMillis();
    }
    Boolean filter = (Boolean)args.get(Nutch.ARG_FILTER);
    Boolean norm = (Boolean)args.get(Nutch.ARG_NORMALIZE);
    // map to inverted subset due for fetch, sort by score
    getConf().setLong(GENERATOR_CUR_TIME, curTime);
    if (topN != null)
      getConf().setLong(GENERATOR_TOP_N, topN);
    if (filter != null)
      getConf().setBoolean(GENERATOR_FILTER, filter);
    int randomSeed = Math.abs(new Random().nextInt());
    batchId = (curTime / 1000) + "-" + randomSeed;
    getConf().setInt(GENERATOR_RANDOM_SEED, randomSeed);
    getConf().set(BATCH_ID, batchId);
    getConf().setLong(Nutch.GENERATE_TIME_KEY, System.currentTimeMillis());
    if (norm != null)
      getConf().setBoolean(GENERATOR_NORMALISE, norm);
    String mode = getConf().get(GENERATOR_COUNT_MODE, GENERATOR_COUNT_VALUE_HOST);
    if (GENERATOR_COUNT_VALUE_HOST.equalsIgnoreCase(mode)) {
      getConf().set(URLPartitioner.PARTITION_MODE_KEY, URLPartitioner.PARTITION_MODE_HOST);
    } else if (GENERATOR_COUNT_VALUE_DOMAIN.equalsIgnoreCase(mode)) {
        getConf().set(URLPartitioner.PARTITION_MODE_KEY, URLPartitioner.PARTITION_MODE_DOMAIN);
    } else {
      LOG.warn("Unknown generator.max.count mode '" + mode + "', using mode=" + GENERATOR_COUNT_VALUE_HOST);
      getConf().set(GENERATOR_COUNT_MODE, GENERATOR_COUNT_VALUE_HOST);
      getConf().set(URLPartitioner.PARTITION_MODE_KEY, URLPartitioner.PARTITION_MODE_HOST);
    }

	// 上面是设置一些要使用要的常量
    numJobs = 1;
    currentJobNum = 0;
	// 生成一个job
    currentJob = new NutchJob(getConf(), "generate: " + batchId);
	// 初始化Map,这里的Map的输出类型为<SelectorEntry,WebPage>, 使用 SelectorEntryPartitioner来进行切分
    StorageUtils.initMapperJob(currentJob, FIELDS, SelectorEntry.class,
        WebPage.class, GeneratorMapper.class, SelectorEntryPartitioner.class, true);
	// 初始化Reducer, 使用了generatorReducer来进行聚合处理
    StorageUtils.initReducerJob(currentJob, GeneratorReducer.class);
    currentJob.waitForCompletion(true);
    ToolUtil.recordJobStatus(null, currentJob, results);
    results.put(BATCH_ID, batchId);
    return results;
  }

好像比原来的Generate简单很多，这里的GeneratorMapper完成的工作与之前的版本是一样的，如url的正规化，过滤，分数的设置，而GeneratorReducer完成的工作也和之前差不多，只是输出变成了DataStore，如HBase，完成以后会每个WebPage进行打标记，表示当前WebPage所完成的一个状态。

4. FetcherJob 源代码分析

使用了Gora的 fetcher比原来简单了很多，下面是其run的源代码

public Map<String,Object> run(Map<String,Object> args) throws Exception {
    checkConfiguration();
    String batchId = (String)args.get(Nutch.ARG_BATCH);
    Integer threads = (Integer)args.get(Nutch.ARG_THREADS);
    Boolean shouldResume = (Boolean)args.get(Nutch.ARG_RESUME);
    Integer numTasks = (Integer)args.get(Nutch.ARG_NUMTASKS);
 
    if (threads != null && threads > 0) {
      getConf().setInt(THREADS_KEY, threads);
    }
    if (batchId == null) {
      batchId = Nutch.ALL_BATCH_ID_STR;
    }
    getConf().set(GeneratorJob.BATCH_ID, batchId);
    if (shouldResume != null) {
      getConf().setBoolean(RESUME_KEY, shouldResume);
    }
    
    LOG.info("FetcherJob : timelimit set for : " + getConf().getLong("fetcher.timelimit", -1));
    LOG.info("FetcherJob: threads: " + getConf().getInt(THREADS_KEY, 10));
    LOG.info("FetcherJob: parsing: " + getConf().getBoolean(PARSE_KEY, false));
    LOG.info("FetcherJob: resuming: " + getConf().getBoolean(RESUME_KEY, false));

    // set the actual time for the timelimit relative
    // to the beginning of the whole job and not of a specific task
    // otherwise it keeps trying again if a task fails
    long timelimit = getConf().getLong("fetcher.timelimit.mins", -1);
    if (timelimit != -1) {
      timelimit = System.currentTimeMillis() + (timelimit * 60 * 1000);
      getConf().setLong("fetcher.timelimit", timelimit);
    }
    numJobs = 1;
    currentJob = new NutchJob(getConf(), "fetch");
	// 得到它过滤的字段
    Collection<WebPage.Field> fields = getFields(currentJob);
	// 初始化mapper, 其输出为<IntWritable,FetchEntry>
	// 在mapper中输入数据进行过滤，主要是对不是同一个batch与已经fetch的数据进行过滤
    StorageUtils.initMapperJob(currentJob, fields, IntWritable.class,
        FetchEntry.class, FetcherMapper.class, FetchEntryPartitioner.class, false);
	// 初始化reducer
    StorageUtils.initReducerJob(currentJob, FetcherReducer.class);
    if (numTasks == null || numTasks < 1) {
      currentJob.setNumReduceTasks(currentJob.getConfiguration().getInt("mapred.map.tasks",
          currentJob.getNumReduceTasks()));
    } else {
      currentJob.setNumReduceTasks(numTasks);
    }
    currentJob.waitForCompletion(true);
    ToolUtil.recordJobStatus(null, currentJob, results);
    return results;
  }

这里把原来在Mapper中使用到的生产者与消费者模型用到了reducer中，重写了reducer的run方法，在其中打开多个抓取线程，对url进行多线程抓取，有兴趣可以看一下FetcherReducer这个类。

5. ParserJob 代码分析

下面是ParserJob.java中的run代码

@Override
  public Map<String,Object> run(Map<String,Object> args) throws Exception {
    String batchId = (String)args.get(Nutch.ARG_BATCH);
    Boolean shouldResume = (Boolean)args.get(Nutch.ARG_RESUME);
    Boolean force = (Boolean)args.get(Nutch.ARG_FORCE);
    
    if (batchId != null) {
      getConf().set(GeneratorJob.BATCH_ID, batchId);
    }
    if (shouldResume != null) {
      getConf().setBoolean(RESUME_KEY, shouldResume);
    }
    if (force != null) {
      getConf().setBoolean(FORCE_KEY, force);
    }
    LOG.info("ParserJob: resuming:\t" + getConf().getBoolean(RESUME_KEY, false));
    LOG.info("ParserJob: forced reparse:\t" + getConf().getBoolean(FORCE_KEY, false));
    if (batchId == null || batchId.equals(Nutch.ALL_BATCH_ID_STR)) {
      LOG.info("ParserJob: parsing all");
    } else {
      LOG.info("ParserJob: batchId:\t" + batchId);
    }
    currentJob = new NutchJob(getConf(), "parse");
    
    Collection<WebPage.Field> fields = getFields(currentJob);
	// 初始化mapper，输出类型为<String,WebPage>, 解析全部在maper完成
    StorageUtils.initMapperJob(currentJob, fields, String.class, WebPage.class,
        ParserMapper.class);
	// 初始化reducer,这里是支持把<key,values>写到数据库中
    StorageUtils.initReducerJob(currentJob, IdentityPageReducer.class);
    currentJob.setNumReduceTasks(0);

    currentJob.waitForCompletion(true);
    ToolUtil.recordJobStatus(null, currentJob, results);
    return results;
  }

6. DbUpdaterJob 代码分析

下面是DbUpdaterjob的run方法代码

public Map<String,Object> run(Map<String,Object> args) throws Exception {
    String crawlId = (String)args.get(Nutch.ARG_CRAWL);
    numJobs = 1;
    currentJobNum = 0;
    currentJob = new NutchJob(getConf(), "update-table");
    if (crawlId != null) {
      currentJob.getConfiguration().set(Nutch.CRAWL_ID_KEY, crawlId);
    }
    //job.setBoolean(ALL, updateAll);
    ScoringFilters scoringFilters = new ScoringFilters(getConf());
    HashSet<WebPage.Field> fields = new HashSet<WebPage.Field>(FIELDS);
    fields.addAll(scoringFilters.getFields());
    
    // Partition by {url}, sort by {url,score} and group by {url}.
    // This ensures that the inlinks are sorted by score when they enter
    // the reducer.
    
    currentJob.setPartitionerClass(UrlOnlyPartitioner.class);
    currentJob.setSortComparatorClass(UrlScoreComparator.class);
    currentJob.setGroupingComparatorClass(UrlOnlyComparator.class);
    
	// 这里的maper读取webpage中的outlinks字段值，对每个外链接计算分数
    StorageUtils.initMapperJob(currentJob, fields, UrlWithScore.class,
        NutchWritable.class, DbUpdateMapper.class);
	// 对新生成的外链接设置一些分数，状态等信息，再把新的WebPage写回数据库
    StorageUtils.initReducerJob(currentJob, DbUpdateReducer.class);
    currentJob.waitForCompletion(true);
    ToolUtil.recordJobStatus(null, currentJob, results);
    return results;
  }

7. SolrIndexerJob 代码分析

下面是其run方法的源代码

@Override
  public Map<String,Object> run(Map<String,Object> args) throws Exception {
    String solrUrl = (String)args.get(Nutch.ARG_SOLR);
    String batchId = (String)args.get(Nutch.ARG_BATCH);
    NutchIndexWriterFactory.addClassToConf(getConf(), SolrWriter.class);
    getConf().set(SolrConstants.SERVER_URL, solrUrl);

// 初始化 job
    currentJob = createIndexJob(getConf(), "solr-index", batchId);
    Path tmp = new Path("tmp_" + System.currentTimeMillis() + "-"
                + new Random().nextInt());
// 设置输出索引到文件，输出格式使用IndexeroutputFormat, 其默认调用Solr的API把数据传给Solr建立索引
    FileOutputFormat.setOutputPath(currentJob, tmp);
    currentJob.waitForCompletion(true);
    ToolUtil.recordJobStatus(null, currentJob, results);
    return results;
  }

有兴趣可以看一下SolrWriter，它实现了NutchIndexerWriter这个接口，来把数据写到不同的后台搜索引擎中，这里默认使用了Solr，当然你也可以通过实现它来扩展你自己的搜索引擎，当然nutch还提供了插件来自定义索引的字段值,也就是IndexingFilter.java这个接口。

8. 总结

Nutch 2.0个人感觉现在还是不成熟的，有很多功能还没有完成，主要的改变还是在它的数据存储层，把原来的数据存储进行了抽象，使其可以更好的运行在大规模数据抓取中，而且可以让用户来扩展具体的数据存储。当然数据存储层的变化带来了一些流程上的变化，有一些操作可以支持使用数据库操作来完成，这也大大减少了一些原来要MR任务来完成的代码。总之nutch 2.0 还是让我们看到了nutch的一个发展方向。希望它发现的越来越好吧。