最近一个群友的boss让研究hbase,让hbase的入库速度达到5w+/s,这可愁死了,4台个人电脑组成的集群,多线程入库调了好久,速度也才1w左右,都没有达到理想的那种速度,然后就想到了这种方式,但是网上多是用mapreduce来实现入库,而现在的需求是实时入库,不生成文件了,所以就只能自己用代码实现了,但是网上查了很多资料都没有查到,最后在一个网友的指引下,看了源码,最后找到了生成Hfile的方式,实现了之后,发现单线程入库速度才达到1w4左右,和之前的多线程的全速差不多了,百思不得其解之时,调整了一下代码把列的Byte.toBytes(cols)这个方法调整出来只做一次,速度立马就到3w了,提升非常明显,这是我的电脑上的速度,估计在它的集群上能更快一点吧,下面把代码和大家分享一下。
1 String tableName = "taglog"; 2 byte[] family = Bytes.toBytes("logs"); 3 //配置文件设置 4 Configuration conf = HBaseConfiguration.create(); 5 conf.set("hbase.master", "192.168.1.133:60000"); 6 conf.set("hbase.zookeeper.quorum", "192.168.1.135"); 7 //conf.set("zookeeper.znode.parent", "/hbase"); 8 conf.set("hbase.metrics.showTableName", "false"); 9 //conf.set("io.compression.codecs", "org.apache.hadoop.io.compress.SnappyCodec"); 10 11 String outputdir = "hdfs://hadoop.Master:8020/user/SEA/hfiles/"; 12 Path dir = new Path(outputdir); 13 Path familydir = new Path(outputdir, Bytes.toString(family)); 14 FileSystem fs = familydir.getFileSystem(conf); 15 BloomType bloomType = BloomType.NONE; 16 final HFileDataBlockEncoder encoder = NoOpDataBlockEncoder.INSTANCE; 17 int blockSize = 64000; 18 Configuration tempConf = new Configuration(conf); 19 tempConf.set("hbase.metrics.showTableName", "false"); 20 tempConf.setFloat(HConstants.HFILE_BLOCK_CACHE_SIZE_KEY, 1.0f); 21 //实例化HFile的Writer,StoreFile实际上只是HFile的轻量级的封装 22 StoreFile.Writer writer = new StoreFile.WriterBuilder(conf, new CacheConfig(tempConf), 23 fs, blockSize) 24 .withOutputDir(familydir) 25 .withCompression(Compression.Algorithm.NONE) 26 .withBloomType(bloomType).withComparator(KeyValue.COMPARATOR) 27 .withDataBlockEncoder(encoder).build(); 28 long start = System.currentTimeMillis(); 29 30 DecimalFormat df = new DecimalFormat("0000000"); 31 32 33 34 KeyValue kv1 = null; 35 KeyValue kv2 = null; 36 KeyValue kv3 = null; 37 KeyValue kv4 = null; 38 KeyValue kv5 = null; 39 KeyValue kv6 = null; 40 KeyValue kv7 = null; 41 KeyValue kv8 = null; 42 43 //这个是耗时操作,只进行一次 44 byte[] cn = Bytes.toBytes("cn"); 45 byte[] dt = Bytes.toBytes("dt"); 46 byte[] ic = Bytes.toBytes("ic"); 47 byte[] ifs = Bytes.toBytes("if"); 48 byte[] ip = Bytes.toBytes("ip"); 49 byte[] le = Bytes.toBytes("le"); 50 byte[] mn = Bytes.toBytes("mn"); 51 byte[] pi = Bytes.toBytes("pi"); 52 53 int maxLength = 3000000; 54 for(int i=0;i<maxLength;i++){ 55 String currentTime = ""+System.currentTimeMillis() + df.format(i); 56 long current = System.currentTimeMillis(); 57 //rowkey和列都要按照字典序的方式顺序写入,否则会报错的 58 kv1 = new KeyValue(Bytes.toBytes(currentTime), 59 family, cn,current,KeyValue.Type.Put,Bytes.toBytes("3")); 60 61 kv2 = new KeyValue(Bytes.toBytes(currentTime), 62 family, dt,current,KeyValue.Type.Put,Bytes.toBytes("6")); 63 64 kv3 = new KeyValue(Bytes.toBytes(currentTime), 65 family, ic,current,KeyValue.Type.Put,Bytes.toBytes("8")); 66 67 kv4 = new KeyValue(Bytes.toBytes(currentTime), 68 family, ifs,current,KeyValue.Type.Put,Bytes.toBytes("7")); 69 70 kv5 = new KeyValue(Bytes.toBytes(currentTime), 71 family, ip,current,KeyValue.Type.Put,Bytes.toBytes("4")); 72 73 kv6 = new KeyValue(Bytes.toBytes(currentTime), 74 family, le,current,KeyValue.Type.Put,Bytes.toBytes("2")); 75 76 kv7 = new KeyValue(Bytes.toBytes(currentTime), 77 family, mn,current,KeyValue.Type.Put,Bytes.toBytes("5")); 78 79 kv8 = new KeyValue(Bytes.toBytes(currentTime), 80 family,pi,current,KeyValue.Type.Put,Bytes.toBytes("1")); 81 82 writer.append(kv1); 83 writer.append(kv2); 84 writer.append(kv3); 85 writer.append(kv4); 86 writer.append(kv5); 87 writer.append(kv6); 88 writer.append(kv7); 89 writer.append(kv8); 90 } 91 92 93 writer.close(); 94 95 //把生成的HFile导入到hbase当中 96 HTable table = new HTable(conf,tableName); 97 LoadIncrementalHFiles loader = new LoadIncrementalHFiles(conf); 98 loader.doBulkLoad(dir, table);
最后再附上查看hfile的方式,查询正确的hfile和自己生成的hfile,方便查找问题。
hbase org.apache.hadoop.hbase.io.hfile.HFile -p -f hdfs://hadoop.Master:8020/user/SEA/hfiles/logs/51aa97b2a25446f89d5c870af92c9fc1