Invalid Schema error in AWS Glue created via Terraform

Question

I have a Kinesis Firehose configuration in Terraform, which reads data from Kinesis stream in JSON, converts it to Parquet using Glue and writes to S3. There is something wrong with data format conversion and I am getting the below error(with some details removed):

{"attemptsMade":1,"arrivalTimestamp":1624541721545,"lastErrorCode":"DataFormatConversion.InvalidSchema","lastErrorMessage":"The schema is invalid. The specified table has no columns.","attemptEndingTimestamp":1624542026951,"rawData":"xx","sequenceNumber":"xx","subSequenceNumber":null,"dataCatalogTable":{"catalogId":null,"databaseName":"db_name","tableName":"table_name","region":null,"versionId":"LATEST","roleArn":"xx"}}

The Terraform configuration for Glue Table, I am using, is as follows:

resource "aws_glue_catalog_table" "stream_format_conversion_table" {
  name          = "${var.resource_prefix}-parquet-conversion-table"
  database_name = aws_glue_catalog_database.stream_format_conversion_db.name

  table_type = "EXTERNAL_TABLE"

  parameters = {
    EXTERNAL              = "TRUE"
    "parquet.compression" = "SNAPPY"
  }

  storage_descriptor {
    location      = "s3://${element(split(":", var.bucket_arn), 5)}/"
    input_format  = "org.apache.hadoop.hive.ql.io.parquet.MapredParquetInputFormat"
    output_format = "org.apache.hadoop.hive.ql.io.parquet.MapredParquetOutputFormat"

    ser_de_info {
      name                  = "my-stream"
      serialization_library = "org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe"

      parameters = {
        "serialization.format" = 1
      }
    }
    columns {
      name = "metadata"
      type = "struct<tenantId:string,env:string,eventType:string,eventTimeStamp:timestamp>"   
    }
    columns {
      name = "eventpayload"
      type = "struct<operation:string,timestamp:timestamp,user_name:string,user_id:int,user_email:string,batch_id:string,initiator_id:string,initiator_email:string,payload:string>"         
    }
  }
}

What needs to change here?

Enclose

Answer 1

I faced the "The schema is invalid. The specified table has no columns" with the following combination:

• avro schema in Glue schema registry,
• glue table created through console using "Add table from existing schema"
• kinesis data firehose configured with Parquet conversion and referencing the glue table created from the schema registry.

It turns out that KDF is unable to read table's schema if table is created from existing schema. Table have to be created from scratch (in opposition to "Add table from existing schema") This isn't documented ... for now.

Budweis

Do you by any chance know if this has been fixed or is documented anywhere now? This is a pretty severe limitation when you have a large body of existing schemas... – Nessy

Hello @Nessy basically AWS support's answer was: we re sorry for the limitation and it is unfortunately not in the upcoming service roadmap. – Budweis

Answer 2

In addition to the answer from mberchon I found that the default generated policy for the Kinesis Delivery Stream did not include the necessary IAM permissions to actually read the schema.

I had to manually modify the IAM policy to include glue:GetSchema and glue:GetSchemaVersion.

Raymer

Answer 3

Frustrated by having to manually define columns, wrote a little python tool that takes a pydantic class (could be made to work with json-schema too) and generated a json that can be used with terraform to create the table.

https://github.com/nanit/j2g

from pydantic import BaseModel
from typing import List

class Bar(BaseModel):
    name: str
    age: int

class Foo(BaseModel):
    nums: List[int]
    bars: List[Bar]
    other: str

get converted to

{
 "nums": "array<int>",
 "bars": "array<struct<name:string,age:int>>",
 "other": "string"
}

and can be used in terraform like so

locals {
  columns = jsondecode(file("${path.module}/glue_schema.json"))
}

resource "aws_glue_catalog_table" "table" {
  name          = "table_name"
  database_name = "db_name"

  storage_descriptor {
    dynamic "columns" {
      for_each = local.columns

      content {
        name = columns.key
        type = columns.value
      }
    }
  }
}

Soften

Answer 4

Thought id post here as i was facing the same problem and found a workaround for this that appears to work.

As is stated above AWS do not allow you to use tables generated from existing schema to convert data types using Firehose. That said if you are using terraform you can create the table using the existing schema, then use the columns attribute from the first table created to create another table and then use that second table as the table for data type conversion in the firehose config, i can confirm this works.

tables terraform:

resource "aws_glue_catalog_table" "aws_glue_catalog_table_from_schema" {
  name          = "first_table"
  database_name = "foo"
  storage_descriptor {
    schema_reference {
      schema_id {
        schema_arn = aws_glue_schema.your_glue_schema.arn
      }
      schema_version_number = aws_glue_schema.your_glue_schema.latest_schema_version
    }
  }
}

resource "aws_glue_catalog_table" "aws_glue_catalog_table_from_first_table" {
  name          = "second_table"
  database_name = "foo"
  storage_descriptor {
    dynamic "columns" {
      for_each = aws_glue_catalog_table.aws_glue_catalog_table_from_schema.storage_descriptor[0].columns
      content {
        name = columns.value.name
        type = columns.value.type
      }
    }
  }
}

firehose data format conversion configuration:

data_format_conversion_configuration {
  output_format_configuration{
    serializer {
      parquet_ser_de {}
    }
  }
  input_format_configuration {
    deserializer {
      hive_json_ser_de {}
    }
  }
  schema_configuration {
    database_name = aws_glue_catalog_table.aws_glue_catalog_table_from_first_table.database_name
    role_arn      = aws_iam_role.firehose_role.arn
    table_name    = aws_glue_catalog_table.aws_glue_catalog_table_from_first_table.name
  }
}

Hax