29、Spark写数据到Hudi时，同步hive表的一些坑

1.hudi的同步hive表没有comment

原以为hudi同步的hive表是根据数据写入的dataframe的schema创建的。就和spark write hive时类似，查看源码后发现不是。

1.1 hudi同步hive的模式

HMS , JDBC , HIVESQL。我这儿常用的是HMS和JDBC
在这里插入图片描述
各个同步模式对应的执行器：

1.2 schema生成

我们可以看到schema生成的代码块。先从提交的commit中获取元数据信息，没有的话则从数据文件中获取schema。两种方式获取到的schema都是没有comment信息的。
org.apache.hudi.common.table.TableSchemaResolver#getTableParquetSchema
在这里插入图片描述

  /**
   * Gets the schema for a hoodie table. Depending on the type of table, read from any file written in the latest
   * commit. We will assume that the schema has not changed within a single atomic write.
   *
   * @return Parquet schema for this table
   * @throws Exception
   */
  private MessageType getTableParquetSchemaFromDataFile() throws Exception {
    HoodieActiveTimeline activeTimeline = metaClient.getActiveTimeline();

    try {
      switch (metaClient.getTableType()) {
        case COPY_ON_WRITE:
          // If this is COW, get the last commit and read the schema from a file written in the
          // last commit
          HoodieInstant lastCommit =
              activeTimeline.getCommitsTimeline().filterCompletedInstants().lastInstant().orElseThrow(() -> new InvalidTableException(metaClient.getBasePath()));
          HoodieCommitMetadata commitMetadata = HoodieCommitMetadata
              .fromBytes(activeTimeline.getInstantDetails(lastCommit).get(), HoodieCommitMetadata.class);
          String filePath = commitMetadata.getFileIdAndFullPaths(metaClient.getBasePath()).values().stream().findAny()
              .orElseThrow(() -> new IllegalArgumentException("Could not find any data file written for commit "
                  + lastCommit + ", could not get schema for table " + metaClient.getBasePath() + ", Metadata :"
                  + commitMetadata));
          return readSchemaFromBaseFile(new Path(filePath));
        case MERGE_ON_READ:
          // If this is MOR, depending on whether the latest commit is a delta commit or
          // compaction commit
          // Get a datafile written and get the schema from that file
          Option<HoodieInstant> lastCompactionCommit =
              metaClient.getActiveTimeline().getCommitTimeline().filterCompletedInstants().lastInstant();
          LOG.info("Found the last compaction commit as " + lastCompactionCommit);

          Option<HoodieInstant> lastDeltaCommit;
          if (lastCompactionCommit.isPresent()) {
            lastDeltaCommit = metaClient.getActiveTimeline().getDeltaCommitTimeline().filterCompletedInstants()
                .findInstantsAfter(lastCompactionCommit.get().getTimestamp(), Integer.MAX_VALUE).lastInstant();
          } else {
            lastDeltaCommit =
                metaClient.getActiveTimeline().getDeltaCommitTimeline().filterCompletedInstants().lastInstant();
          }
          LOG.info("Found the last delta commit " + lastDeltaCommit);

          if (lastDeltaCommit.isPresent()) {
            HoodieInstant lastDeltaInstant = lastDeltaCommit.get();
            // read from the log file wrote
            commitMetadata = HoodieCommitMetadata.fromBytes(activeTimeline.getInstantDetails(lastDeltaInstant).get(),
                HoodieCommitMetadata.class);
            Pair<String, HoodieFileFormat> filePathWithFormat =
                commitMetadata.getFileIdAndFullPaths(metaClient.getBasePath()).values().stream()
                    .filter(s -> s.contains(HoodieLogFile.DELTA_EXTENSION)).findAny()
                    .map(f -> Pair.of(f, HoodieFileFormat.HOODIE_LOG)).orElseGet(() -> {
                      // No Log files in Delta-Commit. Check if there are any parquet files
                      return commitMetadata.getFileIdAndFullPaths(metaClient.getBasePath()).values().stream()
                          .filter(s -> s.contains((metaClient.getTableConfig().getBaseFileFormat().getFileExtension())))
                          .findAny().map(f -> Pair.of(f, HoodieFileFormat.PARQUET)).orElseThrow(() ->
                              new IllegalArgumentException("Could not find any data file written for commit "
                              + lastDeltaInstant + ", could not get schema for table " + metaClient.getBasePath()
                              + ", CommitMetadata :" + commitMetadata));
                    });
            switch (filePathWithFormat.getRight()) {
              case HOODIE_LOG:
                return readSchemaFromLogFile(lastCompactionCommit, new Path(filePathWithFormat.getLeft()));
              case PARQUET:
                return readSchemaFromBaseFile(new Path(filePathWithFormat.getLeft()));
              default:
                throw new IllegalArgumentException("Unknown file format :" + filePathWithFormat.getRight()
                    + " for file " + filePathWithFormat.getLeft());
            }
          } else {
            return readSchemaFromLastCompaction(lastCompactionCommit);
          }
        default:
          LOG.error("Unknown table type " + metaClient.getTableType());
          throw new InvalidTableException(metaClient.getBasePath());
      }
    } catch (IOException e) {
      throw new HoodieException("Failed to read data schema", e);
    }
  }

1.3建表DDL

获取到schema后，我们再看建表行为。
org.apache.hudi.hive.ddl.DDLExecutor#createTable 定义了这个接口建表方法。有两个实现类，一个是
org.apache.hudi.hive.ddl.HMSDDLExecutor。另一个是 org.apache.hudi.hive.ddl.QueryBasedDDLExecutor
在这里插入图片描述
首先，看org.apache.hudi.hive.ddl.HMSDDLExecutor#createTable方法：
ddl操作中使用的字段信息在HiveSchemaUtil.convertMapSchemaToHiveFieldSchema生成，可以直接在这个方法里看到字段的comment信息是直接写死为空字符串的。

再看，org.apache.hudi.hive.ddl.QueryBasedDDLExecutor#createTable方法。
方法里是通过HiveSchemaUtil.generateCreateDDL方法直接生成的ddl建表语句的。这个方法里generateSchemaString方法来生成字段信息的。在这个方法里，也是没有涉及comment信息的。
在这里插入图片描述

1.4结论

同步hive表是在数据写入hudi目录后，根据目录里的schema来创建的hive表，所以创建的hive表没有带着dataframe的comment信息。需要手动执行修改字段comment。

2.追加comment

2.1.使用spark.sql的方式修改comment

用spark.sql(）的方式执行修改comment的sql语句，会调用hudi里的AlterHoodieTableChangeColumnCommand类。这个里面会比较schema，刷新sparksession里的catalog信息，会让任务hang住。（为什么hang住没去排查）大概操作就是写一个使用新的schema的空数据集到hudi来实现schema更新。
org.apache.spark.sql.hudi.command.AlterHoodieTableChangeColumnCommand。
在这里插入图片描述

2.2使用hive-sql的方式修改comment

用hive-jdbc的方式执行修改sql语句。这个方式不会更新hive表里的 TBLPROPERTIES 的 'spark.sql.sources.schema.part.0’信息。
使用dataframe的schame.tojson ，去修改 ‘spark.sql.sources.schema.part.0’ 信息

  /**
   * 将 dataframe 中的comment加到 hudi的hive表中
   *
   * @param df      dataframe
   * @param dbTable hive表
   * @param spark   spark session
   */
  def addCommentForSyncHive(df: DataFrame, dbTable: String, spark: SparkSession, writeOptions: mutable.Map[String, String]): Unit = {
    val comment: Map[String, String] = df.schema.map(sf => (sf.name, sf.getComment().getOrElse(""))).toMap
    info(s"数据集的字段名->备注为：\n${comment.mkString("\n")}")


    val jdbcUrlOption = writeOptions.get(DataSourceWriteOptions.HIVE_URL.key())
    val jdbcUserOption = writeOptions.get(DataSourceWriteOptions.HIVE_USER.key())
    val jdbcPassOption = writeOptions.get(DataSourceWriteOptions.HIVE_PASS.key())
    assert(jdbcUrlOption.isDefined, s"${DataSourceWriteOptions.HIVE_URL.key()} 必须被指定")

    val connection = DbUtil.createHiveConnection(
      jdbcUrlOption.get, jdbcUserOption.getOrElse(""), jdbcPassOption.getOrElse("")
    )
    val stmt = connection.createStatement()
    //需要手动更新hive表中的spark.sql.sources.schema.part.0信息
    stmt.execute(s"ALTER TABLE $dbTable SET TBLPROPERTIES ('spark.sql.sources.schema.part.0' = '${df.schema.json}')")

    // 获取表字段和类型
    val tableSchema = spark.sql(s"DESCRIBE $dbTable")
      .select("col_name", "data_type")
      .collect()
      .map(row => (row.getString(0), row.getString(1)))

    tableSchema.foreach { case (column, dataType) =>
      if (comment.contains(column) && !Seq("ym", "ymd").contains(column)) {
        val newComment = comment.getOrElse(column, "")
        val sql = s"""ALTER TABLE $dbTable CHANGE COLUMN $column $column $dataType COMMENT '$newComment'"""
        info(s"添加备注执行sql：$sql")
        try {
          stmt.execute(sql)
        } catch {
          case e:Throwable =>
            warn("添加备注sql执行失败")
        }
      }
    }
    stmt.close()
    connection.close()
  }