kudu系列: Java API使用和效率测试

Kudu+Impala很适合数据分析, 但直接使用Insert values语句往Kudu表插入数据, 效率实在不好, 测试下来insert的速度仅为80笔/秒. 原因也是显然的, Kudu本身写入效率很高, 但是Impala并没有做这方面优化, 观察下来每次Impala语句执行的overhead都太大了, 导致频繁小批次写入效率非常差, Kudu官方推荐使用Java API或Python API完成数据写入工作. 下面是使用Java API的测试用例, 也可以看出Kudu API的大致用法.

=========================
准备测试Table
=========================

-- kudu table
CREATE TABLE kudu_testdb.tmp_test_perf
(
    id string ENCODING PLAIN_ENCODING COMPRESSION SNAPPY,
    int_value int ,
    bigint_value bigint  ,
    timestamp_value  timestamp ,
    boolean_value  int,
    PRIMARY KEY (id)
)
PARTITION BY HASH (id) PARTITIONS 6
STORED AS KUDU
TBLPROPERTIES (
'kudu.table_name' = 'testdb.tmp_test_perf',
'kudu.master_addresses' = '10.0.0.100:7051,10.0.0.101:7051,10.0.0.101:7051',
'kudu.num_tablet_replicas' = '1'
)
;

=========================
编写测试java程序
=========================
Kudu API 编码注意事项:

1. 尽管建表Impala DDL中,kudu表字段名大小写不敏感, 但在kudu层面, 字段名称已经转成为小写形式, 在Kudu API中, 字段名称必须是小写字母.
2. 建表Impala DDL表名称大小写会被完整地保留下来, 并没有被转成小写, 而且在Kudu API使用中, 表名是大小写敏感的, 必须和建表DDL完全一致.
3. Kudu API给字段赋值函数是不接受传入null, 所以如果在为字段赋值之前, 最好先判断一下取值是否为null. 例如下面两行代码会报错.

Long longTmp=null;
row.addLong("bigint_value",longTmp);

package kudu_perf_test;

import java.sql.Timestamp;
import java.util.UUID;
import org.apache.kudu.client.*;

public class Test {
    private final static int OPERATION_BATCH = 500;

    //同时支持三个模式的测试用例
    public static void insertTestGeneric(KuduSession session, KuduTable table, SessionConfiguration.FlushMode mode,
            int recordCount) throws Exception {
        // SessionConfiguration.FlushMode.AUTO_FLUSH_BACKGROUND
        // SessionConfiguration.FlushMode.AUTO_FLUSH_SYNC
        // SessionConfiguration.FlushMode.MANUAL_FLUSH
        session.setFlushMode(mode);
        if (SessionConfiguration.FlushMode.AUTO_FLUSH_SYNC != mode) {
            session.setMutationBufferSpace(OPERATION_BATCH);
        }
        int uncommit = 0;

        for (int i = 0; i < recordCount; i++) {
            Insert insert = table.newInsert();
            PartialRow row = insert.getRow();
            UUID uuid = UUID.randomUUID();
            row.addString("id", uuid.toString());
            row.addInt("int_value", 100);
            row.addLong("bigint_value", 10000L);

            Long gtmMillis;
            /* System.currentTimeMillis() 是从1970-01-01开始算的毫秒数(GMT), kudu API是采用纳秒数, 所以需要*1000
             另外, 考虑到我们是东8区时间, 所以转成Long型需要再加8个小时, 否则存到Kudu的时间是GTM, 比东8区晚8个小时
             */

            //方法1: 获取当前时间对应的GTM时区unix毫秒数
            gtmMillis=System.currentTimeMillis(); 

            //方法2: 将timestamp转成对应的GTM时区unix毫秒数
            Timestamp localTimestamp = new Timestamp(System.currentTimeMillis());
            gtmMillis=localTimestamp.getTime();   

            //将GTM的毫秒数转成东8区的毫秒数量
            Long shanghaiTimezoneMillis=gtmMillis+8*3600*1000;
            row.addLong("timestamp_value", shanghaiTimezoneMillis*1000);

            session.apply(insert);

            // 对于手工提交, 需要buffer在未满的时候flush,这里采用了buffer一半时即提交
            if (SessionConfiguration.FlushMode.MANUAL_FLUSH == mode) {
                uncommit = uncommit + 1;
                if (uncommit > OPERATION_BATCH / 2) {
                    session.flush();
                    uncommit = 0;
                }
            }
        }

        // 对于手工提交, 保证完成最后的提交
        if (SessionConfiguration.FlushMode.MANUAL_FLUSH == mode && uncommit > 0) {
            session.flush();
        }

        // 对于后台自动提交, 必须保证完成最后的提交, 并保证有错误时能抛出异常
        if (SessionConfiguration.FlushMode.AUTO_FLUSH_BACKGROUND == mode) {
            session.flush();
            RowErrorsAndOverflowStatus error = session.getPendingErrors();
            if (error.isOverflowed() || error.getRowErrors().length > 0) {
                if (error.isOverflowed()) {
                    throw new Exception("Kudu overflow exception occurred.");
                }
                StringBuilder errorMessage = new StringBuilder();
                if (error.getRowErrors().length > 0) {
                    for (RowError errorObj : error.getRowErrors()) {
                        errorMessage.append(errorObj.toString());
                        errorMessage.append(";");
                    }
                }
                throw new Exception(errorMessage.toString());
            }
        }

    }

    //仅支持手动flush的测试用例
    public static void insertTestManual(KuduSession session, KuduTable table, int recordCount) throws Exception {
        // SessionConfiguration.FlushMode.AUTO_FLUSH_BACKGROUND
        // SessionConfiguration.FlushMode.AUTO_FLUSH_SYNC
        // SessionConfiguration.FlushMode.MANUAL_FLUSH
        SessionConfiguration.FlushMode mode = SessionConfiguration.FlushMode.MANUAL_FLUSH;
        session.setFlushMode(mode);
        session.setMutationBufferSpace(OPERATION_BATCH);

        int uncommit = 0;
        for (int i = 0; i < recordCount; i++) {
            Insert insert = table.newInsert();
            PartialRow row = insert.getRow();
            UUID uuid = UUID.randomUUID();
            row.addString("id", uuid.toString());
            row.addInt("int_value", 100);
            row.addLong("bigint_value", 10000L);

            Long gtmMillis;
            /* System.currentTimeMillis() 是从1970-01-01开始算的毫秒数(GMT), kudu API是采用纳秒数, 所以需要*1000
             另外, 考虑到我们是东8区时间, 所以转成Long型需要再加8个小时, 否则存到Kudu的时间是GTM, 比东8区晚8个小时
             */

            //方法1: 获取当前时间对应的GTM时区unix毫秒数
            gtmMillis=System.currentTimeMillis(); 

            //方法2: 将timestamp转成对应的GTM时区unix毫秒数
            Timestamp localTimestamp = new Timestamp(System.currentTimeMillis());
            gtmMillis=localTimestamp.getTime();   

            //将GTM的毫秒数转成东8区的毫秒数量
            Long shanghaiTimezoneMillis=gtmMillis+8*3600*1000;
            row.addLong("timestamp_value", shanghaiTimezoneMillis*1000);

            session.apply(insert);

            // 对于手工提交, 需要buffer在未满的时候flush,这里采用了buffer一半时即提交
            uncommit = uncommit + 1;
            if (uncommit > OPERATION_BATCH / 2) {
                session.flush();
                uncommit = 0;
            }
        }

        // 对于手工提交, 保证完成最后的提交
        if (uncommit > 0) {
            session.flush();
        }
    }

    //仅支持自动flush的测试用例
    public static void insertTestInAutoSync(KuduSession session, KuduTable table, int recordCount) throws Exception {
        // SessionConfiguration.FlushMode.AUTO_FLUSH_BACKGROUND
        // SessionConfiguration.FlushMode.AUTO_FLUSH_SYNC
        // SessionConfiguration.FlushMode.MANUAL_FLUSH
        SessionConfiguration.FlushMode mode = SessionConfiguration.FlushMode.AUTO_FLUSH_SYNC;
        session.setFlushMode(mode);        

        for (int i = 0; i < recordCount; i++) {
            Insert insert = table.newInsert();
            PartialRow row = insert.getRow();
            UUID uuid = UUID.randomUUID();
            row.addString("id", uuid.toString());
            row.addInt("int_value", 100);
            row.addLong("bigint_value", 10000L);

            Long gtmMillis;
            /* System.currentTimeMillis() 是从1970-01-01开始算的毫秒数(GMT), kudu API是采用纳秒数, 所以需要*1000
             另外, 考虑到我们是东8区时间, 所以转成Long型需要再加8个小时, 否则存到Kudu的时间是GTM, 比东8区晚8个小时
             */

            //方法1: 获取当前时间对应的GTM时区unix毫秒数
            gtmMillis=System.currentTimeMillis(); 

            //方法2: 将timestamp转成对应的GTM时区unix毫秒数
            Timestamp localTimestamp = new Timestamp(System.currentTimeMillis());
            gtmMillis=localTimestamp.getTime();   

            //将GTM的毫秒数转成东8区的毫秒数量
            Long shanghaiTimezoneMillis=gtmMillis+8*3600*1000;
            row.addLong("timestamp_value", shanghaiTimezoneMillis*1000);

            //对于AUTO_FLUSH_SYNC模式, apply()将立即完成kudu写入
            session.apply(insert);
        }
    }

    public static void test() throws KuduException {
        KuduClient client = new KuduClient.KuduClientBuilder("10.0.0.100:7051,10.0.0.101:7051,10.0.0.101:7051")
                .build();
        KuduSession session = client.newSession();
        KuduTable table = client.openTable("testdb.tmp_test_perf");

        SessionConfiguration.FlushMode mode;
        Timestamp d1 = null;
        Timestamp d2 = null;
        long millis;
        long seconds;
        int recordCount = 0;

        try {
            mode = SessionConfiguration.FlushMode.AUTO_FLUSH_BACKGROUND;
            d1 = new Timestamp(System.currentTimeMillis());
            insertTestGeneric(session, table, mode, recordCount);
            d2 = new Timestamp(System.currentTimeMillis());
            millis = d2.getTime() - d1.getTime();
            seconds = millis / 1000 % 60;
            System.out.println(mode.name() + "耗时秒数:" + seconds);

            mode = SessionConfiguration.FlushMode.AUTO_FLUSH_SYNC;
            d1 = new Timestamp(System.currentTimeMillis());
            insertTestInAutoSync(session, table,  recordCount);
            d2 = new Timestamp(System.currentTimeMillis());
            millis = d2.getTime() - d1.getTime();
            seconds = millis / 1000 % 60;
            System.out.println(mode.name() + "耗时秒数:" + seconds);

            mode = SessionConfiguration.FlushMode.MANUAL_FLUSH;
            d1 = new Timestamp(System.currentTimeMillis());
            insertTestManual(session, table,  recordCount);
            d2 = new Timestamp(System.currentTimeMillis());
            millis = d2.getTime() - d1.getTime();
            seconds = millis / 1000 % 60;
            System.out.println(mode.name() + "耗时秒数:" + seconds);            

        } catch (Exception e) {
            // TODO Auto-generated catch block
            e.printStackTrace();
        } finally {
            if (!session.isClosed()) {
                session.close();
            }
        }

    }

    public static void main(String[] args) {
        try {
            test();
        } catch (KuduException e) {
            // TODO Auto-generated catch block
            e.printStackTrace();
        }
        System.out.println("Done");

    }
}

=========================
性能测试结果
=========================
MANUAL_FLUSH 模式:8000 row/second
AUTO_FLUSH_BACKGROUND 模式:8000 row/second
AUTO_FLUSH_SYNC 模式:1000 row/second
Impala SQL Insert 语句:80 row/second

=========================
Kudu API 使用总结
=========================
1. 尽量采用 MANUAL_FLUSH, 性能最好, 如果有写入kudu错误, flush()函数就会抛出异常, 逻辑非常清晰.
2. 在性能要求不高的情况下, AUTO_FLUSH_SYNC 也是一个好的选择.
3. 仅仅在demo场景下使用 AUTO_FLUSH_BACKGROUND, 不考虑异常处理时候代码可以很简单, 性能也很好. 在生产环境下, 不推荐的 原因是: 插入数据可能会是乱序, 一旦考虑捕获异常代码就很拖沓.