实时数仓： Hudi 表管理、Flink 性能调优或治理工具脚本

1. Hudi 表管理

1.1 Hudi 表基础管理

创建 Hudi 表
在 HDFS 上创建一个 Hudi 表（以 Merge-on-Read 为例）：

CREATE TABLE real_time_dw.dwd_order_fact (order_id STRING,user_id STRING,product_id STRING,amount DOUBLE,order_date STRING,update_time TIMESTAMP
)
PARTITIONED BY (order_date)
STORED AS PARQUET
TBLPROPERTIES ('type'='MERGE_ON_READ','primaryKey'='order_id','preCombineField'='update_time'
);

1.2 数据操作

插入/更新数据
利用 Hudi 写入工具（如 Spark）进行批量或实时插入更新：

from pyspark.sql import SparkSession
from datetime import datetimespark = SparkSession.builder \.appName("Hudi Example") \.config("spark.serializer", "org.apache.spark.serializer.KryoSerializer") \.getOrCreate()# 加载数据
data = [{"order_id": "1", "user_id": "101", "product_id": "201", "amount": 99.99, "order_date": "2025-01-01", "update_time": datetime.now()},{"order_id": "2", "user_id": "102", "product_id": "202", "amount": 199.99, "order_date": "2025-01-01", "update_time": datetime.now()}
]
df = spark.createDataFrame(data)# 写入 Hudi
hudi_options = {"hoodie.table.name": "dwd_order_fact","hoodie.datasource.write.recordkey.field": "order_id","hoodie.datasource.write.precombine.field": "update_time","hoodie.datasource.write.partitionpath.field": "order_date","hoodie.datasource.write.operation": "upsert","hoodie.datasource.write.table.type": "MERGE_ON_READ","hoodie.datasource.hive.sync.enable": "true","hoodie.datasource.hive.database": "real_time_dw","hoodie.datasource.hive.table": "dwd_order_fact","hoodie.datasource.hive.partition_fields": "order_date"
}df.write.format("hudi").options(**hudi_options).mode("append").save("hdfs://path/to/hudi/dwd_order_fact")

1.3 Hudi 表维护

表清理

• 配置清理策略，清理过期版本：

  hoodie.cleaner.commits.retained=10
  hoodie.cleaner.policy=KEEP_LATEST_COMMITS
  

  保留最近 10 个提交版本。

表压缩

• 针对 MOR 表，定期运行 compaction 任务：

  spark-submit --class org.apache.hudi.utilities.HoodieCompactor \--master yarn \--table-path hdfs://path/to/hudi/dwd_order_fact \--table-name dwd_order_fact
  

元数据管理

• 更新 Hive 元数据：

  MSCK REPAIR TABLE real_time_dw.dwd_order_fact;
  

---

2. Flink 性能调优

2.1 Checkpoint 性能优化

增量 Checkpoint
启用 RocksDB 增量检查点，减少状态存储大小：

env.getCheckpointConfig().enableIncrementalCheckpoints(true);

异步快照
减少 Checkpoint 对性能的影响：

env.getCheckpointConfig().setCheckpointingMode(CheckpointingMode.EXACTLY_ONCE);
env.getCheckpointConfig().setCheckpointTimeout(60000); // 60秒超时
env.getCheckpointConfig().setPreferCheckpointForRecovery(true); // 优先使用Checkpoint恢复

2.2 Watermark 优化

如果数据有延迟，可以允许一定的 out-of-order 数据处理：

WatermarkStrategy.forBoundedOutOfOrderness(Duration.ofSeconds(5)) // 最大延迟5秒.withTimestampAssigner((event, timestamp) -> event.getEventTime());

2.3 状态管理优化

状态后端选择

• 优先选择 RocksDB 状态后端，支持更大的状态数据：

  env.setStateBackend(new RocksDBStateBackend("hdfs://path/to/checkpoints", true));
  

TTL（Time-to-Live）设置

• 自动清理无用状态：

  stateDescriptor.enableTimeToLive(StateTtlConfig.newBuilder(Time.hours(1)).setUpdateType(StateTtlConfig.UpdateType.OnCreateAndWrite).build());
  

2.4 Task Slot 配置

根据并发优化 TaskManager：

• 每个 TaskManager 提供更多 slots：

  taskmanager.numberOfTaskSlots: 4
  

---

3. 治理工具脚本

3.1 数据质量治理（Great Expectations）

脚本自动化
以下 Python 脚本可以实现自动化数据校验（如字段非空和值域校验）：

from great_expectations.core.batch import BatchRequest
from great_expectations.data_context import DataContextcontext = DataContext()batch_request = BatchRequest(datasource_name="my_s3_datasource",data_connector_name="default_runtime_data_connector_name",data_asset_name="dwd_order_fact",runtime_parameters={"path": "s3://path/to/hudi/dwd_order_fact/"},batch_identifiers={"default_identifier_name": "2025-01-01"}
)validator = context.get_validator(batch_request=batch_request)# 非空校验
validator.expect_column_values_to_not_be_null("order_id")
# 值域校验
validator.expect_column_values_to_be_in_set("order_status", ["CREATED", "PAID", "SHIPPED", "CANCELLED"])
# 保存结果
validator.save_expectation_suite("order_fact_suite")context.run_validation_operator("action_list_operator",assets_to_validate=[validator]
)

3.2 数据权限管理（Apache Ranger）

策略 JSON 配置
以下为权限策略 JSON 文件的示例，适用于 Ranger API 批量添加策略：

{"policyName": "dwd_order_fact_policy","serviceType": "hive","resources": {"database": {"values": ["real_time_dw"],"isExcludes": false,"isRecursive": false},"table": {"values": ["dwd_order_fact"],"isExcludes": false,"isRecursive": false}},"policyItems": [{"accesses": [{"type": "select", "isAllowed": true}],"users": ["bi_user"],"groups": ["BI_Group"]},{"accesses": [{"type": "select", "isAllowed": true}, {"type": "insert", "isAllowed": true}],"users": ["etl_user"],"groups": ["ETL_Team"]}]
}

通过 Ranger REST API 部署该策略：

curl -u admin:admin -H "Content-Type: application/json" -X POST -d @policy.json http://<RANGER_HOST>:6080/service/public/v2/api/policy

3.3 数据血缘治理（Apache Atlas）

Flink 血缘注册脚本
通过 REST API 自动将 Flink 作业的输入输出血缘关系上传到 Atlas：

curl -X POST http://<ATLAS_HOST>:21000/api/atlas/v2/entity \
-H "Content-Type: application/json" \
-d '{"entity": {"typeName": "process","attributes": {"name": "flink_order_job","inputs": [{"typeName": "kafka_topic", "uniqueAttributes": {"qualifiedName": "order_topic"}}],"outputs": [{"typeName": "hdfs_path", "uniqueAttributes": {"qualifiedName": "hdfs://path/to/hudi/dwd_order_fact"}}]}}
}'

---