[大数据]Apache Hudi 在 AWS Glue 中的实践

AWS Glue 实践

Glue Job 配置

ps：IAM的权限配置请自行摸索，权限上基本是缺啥补啥，如果只是作为测试，直接给所有的权限（action和resource都是*）就可以一劳永逸了

Job Pramater

Dependent jars path

s3://gavin-test2/dependency_jars/hudi/spark-avro_2.11-2.4.3.jar,s3://gavin-test2/dependency_jars/hudi/hudi-spark-bundle_2.11-0.8.0.jar

jar 的下载路径：

存储为非分区表

Python Code

import sys
from awsglue.transforms import *
from awsglue.utils import getResolvedOptions
from pyspark.context import SparkContext
from awsglue.context import GlueContext
from awsglue.job import Job
from pyspark.sql.types import *


args = getResolvedOptions(sys.argv, ['JOB_NAME'])
sc = SparkContext()
glueContext = GlueContext(sc)
spark = glueContext.spark_session

basePath = 's3://gavin-test2/tables/hudi/table1/'
table_name = 'table1'
database = 'default'
data = [('Alice', 1, '2022/02/28'), ('Jhone', 2, '2022/03/01')]
rdd = sc.parallelize(data)
schema = StructType(
    [
        StructField("name", StringType(), True),
        StructField("age", IntegerType(), True),
        StructField("partitioin_path", StringType(), True),
    ]
)
src_df = spark.createDataFrame(rdd, schema)

hudi_options = {
            'hoodie.table.name': table_name,
            'hoodie.datasource.write.operation': 'insert',
            'hoodie.datasource.write.recordkey.field': 'name',
            'hoodie.datasource.write.table.name': table_name,
            'hoodie.datasource.hive_sync.enable': 'true',
            'hoodie.datasource.hive_sync.database': 'default',
            'hoodie.datasource.hive_sync.use_jdbc': 'false',
            'hoodie.datasource.hive_sync.table': table_name,
            'hoodie.datasource.write.partitionpath.field': '',
            'hoodie.datasource.hive_sync.partition_fields': '',
            'hoodie.datasource.write.keygenerator.class': 'org.apache.hudi.keygen.NonpartitionedKeyGenerator',
            'hoodie.datasource.hive_sync.partition_extractor_class': 'org.apache.hudi.hive.NonPartitionedExtractor'
        }

src_df.write.format("hudi").options(**hudi_options).mode("overwrite").save(basePath)

AWS Glue Catalog

Query in Athena

Files in S3 Bucket

存储为分区表

分区表可以选择时间字段作为分区字段，也可以选择非时间字段作为分区字段，本文使用形如「yyyy/mm/dd」的时间字符串作为分区字段；

Python Code

import sys
from awsglue.transforms import *
from awsglue.utils import getResolvedOptions
from pyspark.context import SparkContext
from awsglue.context import GlueContext
from awsglue.job import Job
from pyspark.sql.types import *


args = getResolvedOptions(sys.argv, ['JOB_NAME'])
sc = SparkContext()
glueContext = GlueContext(sc)
spark = glueContext.spark_session

basePath = 's3://gavin-test2/tables/hudi/table2/'
table_name = 'table2'
database = 'default'
data = [('Alice', 1, '2022/02/28'), ('Jhone', 2, '2022/03/01')]
rdd = sc.parallelize(data)
schema = StructType(
    [
        StructField("name", StringType(), True),
        StructField("age", IntegerType(), True),
        StructField("partitioin_path", StringType(), True),
    ]
)
src_df = spark.createDataFrame(rdd, schema)

hudi_options = {
            'hoodie.table.name': table_name,
            'hoodie.datasource.write.operation': 'insert',
            'hoodie.datasource.write.recordkey.field': 'name',
            'hoodie.datasource.write.table.name': table_name,
            'hoodie.datasource.hive_sync.enable': 'true',
            'hoodie.datasource.hive_sync.database': 'default',
            'hoodie.datasource.hive_sync.use_jdbc': 'false',
            'hoodie.datasource.hive_sync.table': table_name,
            'hoodie.datasource.write.partitionpath.field': 'partitioin_path',
            'hoodie.datasource.hive_sync.partition_fields': 'partitioin_path',
            'hoodie.datasource.write.keygenerator.class': 'org.apache.hudi.keygen.SimpleKeyGenerator',
            'hoodie.datasource.hive_sync.partition_extractor_class': 'org.apache.hudi.hive.SlashEncodedDayPartitionValueExtractor'
        }

src_df.write.format("hudi").options(**hudi_options).mode("overwrite").save(basePath)

AWS Glue Catalog

Query in Athena

Files in S3 Bucket

FAQ

ClassNotFoundException: org.apache.calcite.rel.type.RelDataTypeSystem

Error Info

在Hudi中设置了：DataSourceWriteOptions.HIVE_USE_JDBC_OPT_KEY -> “false”，具体错误为：

java.lang.ClassNotFoundException: org.apache.calcite.rel.type.RelDataTypeSystem

这是由于Hiv3/Spark3移除了对于calcite包的依赖引起的

Solution:

我偷了个懒，将spark 的版本降到2.X

IllegalArgumentException: Partition path default is not in the form yyyy/mm/dd

Error Info

Caused by: java.lang.IllegalArgumentException: Partition path default is not in the form yyyy/mm/dd 
	at org.apache.hudi.hive.SlashEncodedDayPartitionValueExtractor.extractPartitionValuesInPath(SlashEncodedDayPartitionValueExtractor.java:55)
	at org.apache.hudi.hive.HoodieHiveClient.getPartitionEvents(HoodieHiveClient.java:220)
	at org.apache.hudi.hive.HiveSyncTool.syncPartitions(HiveSyncTool.java:221)
	... 42 more

Solution

由于配置'hoodie.datasource.hive_sync.partition_extractor_class':'org.apache.hudi.hive.SlashEncodedDayPartitionValueExtractor', 而org.apache.hudi.hive.SlashEncodedDayPartitionValueExtractor 中要求时间格式必须是这个类型

org.apache.hudi.hive.SlashEncodedDayPartitionValueExtractor#extractPartitionValuesInPath

参考

[1] Hudi 实践 | 在 AWS Glue 中使用 Apache Hudi: https://jishuin.proginn.com/p/763bfbd56de6

[2] 详解Apache Hudi如何配置各种类型分区: https://www.cnblogs.com/leesf456/p/13521694.html

[3] EMR + Hudi报ClassNotFoundException: RelDataTypeSystem错误的解决方法: https://blog.csdn.net/bluishglc/article/details/117441071