ZipCodePageProcessor.java

package us.codecraft.webmagic.samples.scheduler;

import org.apache.commons.lang3.StringUtils;
import us.codecraft.webmagic.Page;
import us.codecraft.webmagic.Request;
import us.codecraft.webmagic.Site;
import us.codecraft.webmagic.Spider;
import us.codecraft.webmagic.processor.PageProcessor;
import us.codecraft.webmagic.scheduler.PriorityScheduler;

import java.util.List;
import java.util.regex.Matcher;
import java.util.regex.Pattern;

import static us.codecraft.webmagic.selector.Selectors.xpath;

/**
 * @author code4crafter@gmail.com
 */
public class ZipCodePageProcessor implements PageProcessor {

    private Site site = Site.me().setCharset("gb2312")
            .setSleepTime(100);

    @Override
    public void process(Page page) {
        if (page.getUrl().toString().equals("http://www.ip138.com/post/")) {
            processCountry(page);
        } else if (page.getUrl().regex("http://www\\.ip138\\.com/\\d{6}[/]?$").toString() != null) {
            processDistrict(page);
        } else {
            processProvince(page);
        }

    }

    private void processCountry(Page page) {
        List<String> provinces = page.getHtml().xpath("//*[@id=\"newAlexa\"]/table/tbody/tr/td").all();
        for (String province : provinces) {
            String link = xpath("//@href").select(province);
            String title = xpath("/text()").select(province);
            Request request = new Request(link).setPriority(0).putExtra("province", title);
            page.addTargetRequest(request);
        }
    }

    private void processProvince(Page page) {
        //这里仅靠xpath没法精准定位,所以使用正则作为筛选,不符合正则的会被过滤掉
        List<String> districts = page.getHtml().xpath("//body/table/tbody/tr[@bgcolor=\"#ffffff\"]").all();
        Pattern pattern = Pattern.compile("<td>([^<>]+)</td>.*?href=\"(.*?)\"",Pattern.DOTALL);
        for (String district : districts) {
            Matcher matcher = pattern.matcher(district);
            while (matcher.find()) {
                String title = matcher.group(1);
                String link = matcher.group(2);
                Request request = new Request(link).setPriority(1).putExtra("province", page.getRequest().getExtra("province")).putExtra("district", title);
                page.addTargetRequest(request);
            }
        }
    }

    private void processDistrict(Page page) {
        String province = page.getRequest().getExtra("province").toString();
        String district = page.getRequest().getExtra("district").toString();
        String zipCode = page.getHtml().regex("<h2>邮编:(\\d+)</h2>").toString();
        page.putField("result", StringUtils.join(new String[]{province, district,
                zipCode}, "\t"));
        List<String> links = page.getHtml().links().regex("http://www\\.ip138\\.com/\\d{6}[/]?$").all();
        for (String link : links) {
            page.addTargetRequest(new Request(link).setPriority(2).putExtra("province", province).putExtra("district", district));
        }

    }

    @Override
    public Site getSite() {
        return site;
    }

    public static void main(String[] args) {
        Spider spider = Spider.create(new ZipCodePageProcessor()).scheduler(new PriorityScheduler()).addUrl("http://www.ip138.com/post/");

        spider.run();
    }
}