xin.bluesky.leiothrix.server.action.allocate.SequencePartitionAllocator Maven / Gradle / Ivy
package xin.bluesky.leiothrix.server.action.allocate;
import xin.bluesky.leiothrix.model.task.partition.PartitionTask;
import xin.bluesky.leiothrix.server.action.exception.NoTaskException;
import java.util.ArrayList;
import java.util.List;
/**
* 按照table+range作为维度,依次分配.
*
* eg:假设两个表(table-1,table-2),各自range情况如下:
*
* -
* table-1:0-1000,1001-2000两个range
*
* -
* table-2:0-3000,3001-6000,6001-9000三个range
*
*
* 那么分配顺序依次是:table-1:0-1000,table-1:1001-2000,table-2:0-3000,table-2:3001-6000,table-3:6001-9000.
* 当然这是理论上的划分,实际上由于多线程的关系和网络时延区别,这个顺序会略有颠倒,并不保证严格有序,但总体来说,是按照从前到后的顺序.
*
*
* 在这种情况下,会出现同一时刻,两个表的range在同时处理的情况,比如table-1:1001-2000和table-2:0-3000会分别被两个worker线程执行.
* 在实际场景中,可能有这样的业务:需要等待table-1全部处理完毕,再处理table-2的数据(比如数据清洗),此时需要保证table-2严格在table-1
* 完全结束后再执行,此时应该使用{@link TableOneByOnePartitionAllocator}
*
* @author 张轲
*/
public class SequencePartitionAllocator extends AbstractPartitionAllocator {
@Override
public List findRange(final String taskId, List tableNameList) throws NoTaskException {
List result = new ArrayList();
for (final String tableName : tableNameList) {
if (isTableProcessingOrFinished(taskId, tableName)) {
continue;
}
RangeScanner scanner = new RangeScanner(taskId, tableName);
RangeScanResult scanResult = scanner.scan();
if (scanResult.hasUnallocatedRange()) {
result = preAllocate(taskId, tableName, scanResult.getUnallocatedRangeNameList());
break;
} else {
continue;
}
}
if (result.isEmpty()) {
throw new NoTaskException(String.format("任务[taskId=%s]下没有未被分配的任务片", taskId));
}
return result;
}
}