在过去多个版本中,Apache Doris 持续加深与数据湖的融合,当前已演进出一套成熟的湖仓一体解决方案。
Apache Hudi 是目前最主流的开放数据湖格式之一,也是事务性的数据湖管理平台,支持包括 Apache Doris 在内的多种主流查询引擎。Apache Doris 同样对 Apache Hudi 数据表的读取能力进行了增强:
凭借 Apache Doris 的高性能查询执行以及 Apache Hudi 的实时数据管理能力,可以实现高效、灵活、低成本的数据查询和分析,同时也提供了强大的数据回溯、审计和增量处理功能,当前基于 Apache Doris 和 Apache Hudi 的组合已经在多个社区用户的真实业务场景中得到验证和推广:
本文将在 Docker 环境下,为读者介绍如何快速搭建 Apache Doris + Apache Hudi 的测试及演示环境,并对各功能操作进行演示,帮助读者快速入门。
本文涉及所有脚本和代码可以从该地址获取:https://github.com/apache/doris/tree/master/samples/datalake/hudi
本文示例采用 Docker Compose 部署,组件及版本号如下:
1. 创建 Docker 网络
sudo docker network create -d bridge hudi-net
2. 启动所有组件
sudo ./start-hudi-compose.sh
3. 启动后,可以使用如下脚本,登陆 Spark 命令行或 Doris 命令行:
sudo ./login-spark.sh
sudo ./login-doris.sh
03 数据准备
接下来先通过 Spark 生成 Hudi 的数据。如下方代码所示,集群中已经包含一张名为customer
的 Hive 表,可以通过这张 Hive 表,创建一个 Hudi 表:
-- ./login-spark.sh
spark-sql> use default;
-- create a COW table
spark-sql> CREATE TABLE customer_cow
USING hudi
TBLPROPERTIES (
type = 'cow',
primaryKey = 'c_custkey',
preCombineField = 'c_name'
)
PARTITIONED BY (c_nationkey)
AS SELECT * FROM customer;
-- create a MOR table
spark-sql> CREATE TABLE customer_mor
USING hudi
TBLPROPERTIES (
type = 'mor',
primaryKey = 'c_custkey',
preCombineField = 'c_name'
)
PARTITIONED BY (c_nationkey)
AS SELECT * FROM customer;
04 数据查询
如下所示,Doris 集群中已经创建了名为hudi
的 Catalog(可通过SHOW CATALOGS
查看)。以下为该 Catalog 的创建语句:
-- 已经创建,无需再次执行
CREATE CATALOG `hive` PROPERTIES (
"type"="hms",
'hive.metastore.uris' = 'thrift://hive-metastore:9083',
"s3.access_key" = "minio",
"s3.secret_key" = "minio123",
"s3.endpoint" = "http://minio:9000",
"s3.region" = "us-east-1",
"use_path_style" = "true"
);
1. 手动刷新该 Catalog,对创建的 Hudi 表进行同步:
-- ./login-doris.sh
doris> REFRESH CATALOG hive;
2. 使用 Spark 操作 Hudi 中的数据,都可以在 Doris 中实时可见,不需要再次刷新 Catalog。我们通过 Spark 分别给 COW 和 MOR 表插入一行数据:
spark-sql> insert into customer_cow values (100, "Customer#000000100", "jD2xZzi", "25-430-914-2194", 3471.59, "BUILDING", "cial ideas. final, furious requests", 25);
spark-sql> insert into customer_mor values (100, "Customer#000000100", "jD2xZzi", "25-430-914-2194", 3471.59, "BUILDING", "cial ideas. final, furious requests", 25);
3. 通过 Doris 可以直接查询到最新插入的数据:
doris> use hive.default;
doris> select * from customer_cow where c_custkey = 100;
doris> select * from customer_mor where c_custkey = 100;
4. 再通过 Spark 插入c_custkey=32
已经存在的数据,即覆盖已有数据:
spark-sql> insert into customer_cow values (32, "Customer#000000032_update", "jD2xZzi", "25-430-914-2194", 3471.59, "BUILDING", "cial ideas. final, furious requests", 15);
spark-sql> insert into customer_mor values (32, "Customer#000000032_update", "jD2xZzi", "25-430-914-2194", 3471.59, "BUILDING", "cial ideas. final, furious requests", 15);
5. 通过 Doris 可以查询更新后的数据:
doris> select * from customer_cow where c_custkey = 32;
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
| c_custkey | c_name | c_address | c_phone | c_acctbal | c_mktsegment | c_comment | c_nationkey |
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
| 32 | Customer#000000032_update | jD2xZzi | 25-430-914-2194 | 3471.59 | BUILDING | cial ideas. final, furious requests | 15 |
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
doris> select * from customer_mor where c_custkey = 32;
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
| c_custkey | c_name | c_address | c_phone | c_acctbal | c_mktsegment | c_comment | c_nationkey |
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
| 32 | Customer#000000032_update | jD2xZzi | 25-430-914-2194 | 3471.59 | BUILDING | cial ideas. final, furious requests | 15 |
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
05 Incremental Read
Incremental Read 是 Hudi 提供的功能特性之一,通过 Incremental Read,用户可以获取指定时间范围的增量数据,从而实现对数据的增量处理。对此, Doris 可对插入c_custkey=100
后的变更数据进行查询。如下所示,我们插入了一条c_custkey=32
的数据:
doris> select * from customer_cow@incr('beginTime'='20240603015018572');
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
| c_custkey | c_name | c_address | c_phone | c_acctbal | c_mktsegment | c_comment | c_nationkey |
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
| 32 | Customer#000000032_update | jD2xZzi | 25-430-914-2194 | 3471.59 | BUILDING | cial ideas. final, furious requests | 15 |
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
spark-sql> select * from hudi_table_changes('customer_cow', 'latest_state', '20240603015018572');
doris> select * from customer_mor@incr('beginTime'='20240603015058442');
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
| c_custkey | c_name | c_address | c_phone | c_acctbal | c_mktsegment | c_comment | c_nationkey |
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
| 32 | Customer#000000032_update | jD2xZzi | 25-430-914-2194 | 3471.59 | BUILDING | cial ideas. final, furious requests | 15 |
+-----------+---------------------------+-----------+-----------------+-----------+--------------+-------------------------------------+-------------+
spark-sql> select * from hudi_table_changes('customer_mor', 'latest_state', '20240603015058442');
06 TimeTravel
Doris 支持查询指定快照版本的 Hudi 数据,从而实现对数据的 Time Travel 功能。首先,可以通过 Spark 查询两张 Hudi 表的提交历史:
spark-sql> call show_commits(table => 'customer_cow', limit => 10);
20240603033556094 20240603033558249 commit 448833 0 1 1 183 0 0
20240603015444737 20240603015446588 commit 450238 0 1 1 202 1 0
20240603015018572 20240603015020503 commit 436692 1 0 1 1 0 0
20240603013858098 20240603013907467 commit 44902033 100 0 25 18751 0 0
spark-sql> call show_commits(table => 'customer_mor', limit => 10);
20240603033745977 20240603033748021 deltacommit 1240 0 1 1 0 0 0
20240603015451860 20240603015453539 deltacommit 1434 0 1 1 1 1 0
20240603015058442 20240603015100120 deltacommit 436691 1 0 1 1 0 0
20240603013918515 20240603013922961 deltacommit 44904040 100 0 25 18751 0 0
接着,可通过 Doris 执行c_custkey=32
,查询数据插入之前的数据快照。如下可看到c_custkey=32
的数据还未更新:
注:Time Travel 语法暂时不支持新优化器,需要先执行
set enable_nereids_planner=false;
关闭新优化器,该问题将会在后续版本中修复。
doris> select * from customer_cow for time as of '20240603015018572' where c_custkey = 32 or c_custkey = 100;
+-----------+--------------------+---------------------------------------+-----------------+-----------+--------------+--------------------------------------------------+-------------+
| c_custkey | c_name | c_address | c_phone | c_acctbal | c_mktsegment | c_comment | c_nationkey |
+-----------+--------------------+---------------------------------------+-----------------+-----------+--------------+--------------------------------------------------+-------------+
| 32 | Customer#000000032 | jD2xZzi UmId,DCtNBLXKj9q0Tlp2iQ6ZcO3J | 25-430-914-2194 | 3471.53 | BUILDING | cial ideas. final, furious requests across the e | 15 |
| 100 | Customer#000000100 | jD2xZzi | 25-430-914-2194 | 3471.59 | BUILDING | cial ideas. final, furious requests | 25 |
+-----------+--------------------+---------------------------------------+-----------------+-----------+--------------+--------------------------------------------------+-------------+
-- compare with spark-sql
spark-sql> select * from customer_mor timestamp as of '20240603015018572' where c_custkey = 32 or c_custkey = 100;
doris> select * from customer_mor for time as of '20240603015058442' where c_custkey = 32 or c_custkey = 100;
+-----------+--------------------+---------------------------------------+-----------------+-----------+--------------+--------------------------------------------------+-------------+
| c_custkey | c_name | c_address | c_phone | c_acctbal | c_mktsegment | c_comment | c_nationkey |
+-----------+--------------------+---------------------------------------+-----------------+-----------+--------------+--------------------------------------------------+-------------+
| 100 | Customer#000000100 | jD2xZzi | 25-430-914-2194 | 3471.59 | BUILDING | cial ideas. final, furious requests | 25 |
| 32 | Customer#000000032 | jD2xZzi UmId,DCtNBLXKj9q0Tlp2iQ6ZcO3J | 25-430-914-2194 | 3471.53 | BUILDING | cial ideas. final, furious requests across the e | 15 |
+-----------+--------------------+---------------------------------------+-----------------+-----------+--------------+--------------------------------------------------+-------------+
spark-sql> select * from customer_mor timestamp as of '20240603015058442' where c_custkey = 32 or c_custkey = 100;
查询优化
Apache Hudi 中的数据大致可以分为两类 —— 基线数据和增量数据。基线数据通常是已经经过合并的 Parquet 文件,而增量数据是指由 INSERT、UPDATE 或 DELETE 产生的数据增量。基线数据可以直接读取,增量数据需要通过 Merge on Read 的方式进行读取。
对于 Hudi COW 表的查询或者 MOR 表的 Read Optimized 查询而言,其数据都属于基线数据,可直接通过 Doris 原生的 Parquet Reader 读取数据文件,且可获得极速的查询响应。而对于增量数据,Doris 需要通过 JNI 调用 Hudi 的 Java SDK 进行访问。为了达到最优的查询性能,Apache Doris 在查询时,会将一个查询中的数据分为基线和增量数据两部分,并分别使用上述方式进行读取。
为验证该优化思路,我们通过EXPLAIN
语句来查看一个下方示例的查询中,分别有多少基线数据和增量数据。对于 COW 表来说,所有 101 个数据分片均为是基线数据(hudiNativeReadSplits=101/101
),因此 COW 表全部可直接通过 Doris Parquet Reader 进行读取,因此可获得最佳的查询性能。对于 ROW 表,大部分数据分片是基线数据(hudiNativeReadSplits=100/101
),一个分片数为增量数据,基本也能够获得较好的查询性能。
-- COW table is read natively
doris> explain select * from customer_cow where c_custkey = 32;
| 0:VHUDI_SCAN_NODE(68) |
| table: customer_cow |
| predicates: (c_custkey[#5] = 32) |
| inputSplitNum=101, totalFileSize=45338886, scanRanges=101 |
| partition=26/26 |
| cardinality=1, numNodes=1 |
| pushdown agg=NONE |
| hudiNativeReadSplits=101/101 |
-- MOR table: because only the base file contains `c_custkey = 32` that is updated, 100 splits are read natively, while the split with log file is read by JNI.
doris> explain select * from customer_mor where c_custkey = 32;
| 0:VHUDI_SCAN_NODE(68) |
| table: customer_mor |
| predicates: (c_custkey[#5] = 32) |
| inputSplitNum=101, totalFileSize=45340731, scanRanges=101 |
| partition=26/26 |
| cardinality=1, numNodes=1 |
| pushdown agg=NONE |
| hudiNativeReadSplits=100/101 |
可以通过 Spark 进行一些删除操作,进一步观察 Hudi 基线数据和增量数据的变化:
-- Use delete statement to see more differences
spark-sql> delete from customer_cow where c_custkey = 64;
doris> explain select * from customer_cow where c_custkey = 64;
spark-sql> delete from customer_mor where c_custkey = 64;
doris> explain select * from customer_mor where c_custkey = 64;
此外,还可以通过分区条件进行分区裁剪,从而进一步减少数据量,以提升查询速度。如下示例中,通过分区条件c_nationkey = 15
进行分区裁减,使得查询请求只需要访问一个分区(partition=1/26
)的数据即可。
-- customer_xxx is partitioned by c_nationkey, we can use the partition column to prune data
doris> explain select * from customer_mor where c_custkey = 64 and c_nationkey = 15;
| 0:VHUDI_SCAN_NODE(68) |
| table: customer_mor |
| predicates: (c_custkey[#5] = 64), (c_nationkey[#12] = 15) |
| inputSplitNum=4, totalFileSize=1798186, scanRanges=4 |
| partition=1/26 |
| cardinality=1, numNodes=1 |
| pushdown agg=NONE |
| hudiNativeReadSplits=3/4 |
结束语
以上是基于 Apache Doris 与 Apache Hudi 快速搭建测试 / 演示环境的详细指南,后续我们还将陆续推出 Apache Doris 与各类主流数据湖格式及存储系统构建湖仓一体架构的系列指南,欢迎持续关注。