Skip to main content
Version: Next

S3 Data Lake

This connector ingests S3 datasets into DataHub. It allows mapping an individual file or a folder of files to a dataset in DataHub. To specify the group of files that form a dataset, use path_specs configuration in ingestion recipe. Refer section Path Specs for more details.

Concept Mapping​

This ingestion source maps the following Source System Concepts to DataHub Concepts:

Source ConceptDataHub ConceptNotes
"s3"Data Platform
s3 object / Folder containing s3 objectsDataset
s3 bucketContainerSubtype S3 bucket
s3 folderContainerSubtype Folder

This connector supports both local files as well as those stored on AWS S3 (which must be identified using the prefix s3://). [a]

Supported file types​

Supported file types are as follows:

  • CSV (*.csv)
  • TSV (*.tsv)
  • JSONL (*.jsonl)
  • JSON (*.json)
  • Parquet (*.parquet)
  • Apache Avro (*.avro)

Schemas for Parquet and Avro files are extracted as provided.

Schemas for schemaless formats (CSV, TSV, JSONL, JSON) are inferred. For CSV, TSV and JSONL files, we consider the first 100 rows by default, which can be controlled via the max_rows recipe parameter (see below) JSON file schemas are inferred on the basis of the entire file (given the difficulty in extracting only the first few objects of the file), which may impact performance. We are working on using iterator-based JSON parsers to avoid reading in the entire JSON object.

Profiling​

This plugin extracts:

  • Row and column counts for each dataset
  • For each column, if profiling is enabled:
    • null counts and proportions
    • distinct counts and proportions
    • minimum, maximum, mean, median, standard deviation, some quantile values
    • histograms or frequencies of unique values

Note that because the profiling is run with PySpark, we require Spark 3.0.3 with Hadoop 3.2 to be installed (see compatibility for more details). If profiling, make sure that permissions for s3a:// access are set because Spark and Hadoop use the s3a:// protocol to interface with AWS (schema inference outside of profiling requires s3:// access). Enabling profiling will slow down ingestion runs.Incubating

Important Capabilities​

CapabilityStatusNotes
Data Profiling✅Optionally enabled via configuration
Detect Deleted Entities✅Optionally enabled via stateful_ingestion.remove_stale_metadata
Extract Tags✅Can extract S3 object/bucket tags if enabled

CLI based Ingestion​

Install the Plugin​

pip install 'acryl-datahub[s3]'

Starter Recipe​

Check out the following recipe to get started with ingestion! See below for full configuration options.

For general pointers on writing and running a recipe, see our main recipe guide.

source:
  type: s3
  config:
    path_specs:
      -
        include: "s3://covid19-lake/covid_knowledge_graph/csv/nodes/*.*"

    aws_config:
      aws_access_key_id: *****
      aws_secret_access_key: *****
      aws_region: us-east-2
    env: "PROD"
    profiling:
      enabled: false

# sink configs

Config Details​

Note that a . is used to denote nested fields in the YAML recipe.

FieldDescription
path_specs ✅
array
List of PathSpec. See below the details about PathSpec
path_specs.PathSpec
PathSpec
path_specs.PathSpec.include ❓
string
Path to table. Name variable {table} is used to mark the folder with dataset. In absence of {table}, file level dataset will be created. Check below examples for more details.
path_specs.PathSpec.allow_double_stars
boolean
Allow double stars in the include path. This can affect performance significantly if enabled
Default: False
path_specs.PathSpec.default_extension
string
For files without extension it will assume the specified file type. If it is not set the files without extensions will be skipped.
path_specs.PathSpec.enable_compression
boolean
Enable or disable processing compressed files. Currently .gz and .bz files are supported.
Default: True
path_specs.PathSpec.sample_files
boolean
Not listing all the files but only taking a handful amount of sample file to infer the schema. File count and file size calculation will be disabled. This can affect performance significantly if enabled
Default: True
path_specs.PathSpec.table_name
string
Display name of the dataset.Combination of named variables from include path and strings
path_specs.PathSpec.exclude
array
list of paths in glob pattern which will be excluded while scanning for the datasets
path_specs.PathSpec.exclude.string
string
path_specs.PathSpec.file_types
array
Files with extenstions specified here (subset of default value) only will be scanned to create dataset. Other files will be omitted.
Default: ['csv', 'tsv', 'json', 'parquet', 'avro']
path_specs.PathSpec.file_types.string
string
add_partition_columns_to_schema
boolean
Whether to add partition fields to the schema.
Default: False
max_rows
integer
Maximum number of rows to use when inferring schemas for TSV and CSV files.
Default: 100
number_of_files_to_sample
integer
Number of files to list to sample for schema inference. This will be ignored if sample_files is set to False in the pathspec.
Default: 100
platform
string
The platform that this source connects to (either 's3' or 'file'). If not specified, the platform will be inferred from the path_specs.
Default:
platform_instance
string
The instance of the platform that all assets produced by this recipe belong to
sort_schema_fields
boolean
Whether to sort schema fields by fieldPath when inferring schemas.
Default: False
spark_config
object
Spark configuration properties to set on the SparkSession. Put config property names into quotes. For example: '"spark.executor.memory": "2g"'
Default: {}
spark_driver_memory
string
Max amount of memory to grant Spark.
Default: 4g
use_s3_bucket_tags
boolean
Whether or not to create tags in datahub from the s3 bucket
use_s3_object_tags
boolean
Whether or not to create tags in datahub from the s3 object
verify_ssl
One of boolean, string
Either a boolean, in which case it controls whether we verify the server's TLS certificate, or a string, in which case it must be a path to a CA bundle to use.
Default: True
env
string
The environment that all assets produced by this connector belong to
Default: PROD
aws_config
AwsConnectionConfig
AWS configuration
aws_config.aws_access_key_id
string
AWS access key ID. Can be auto-detected, see the AWS boto3 docs for details.
aws_config.aws_advanced_config
object
Advanced AWS configuration options. These are passed directly to botocore.config.Config.
aws_config.aws_endpoint_url
string
The AWS service endpoint. This is normally constructed automatically, but can be overridden here.
aws_config.aws_profile
string
Named AWS profile to use. Only used if access key / secret are unset. If not set the default will be used
aws_config.aws_proxy
map(str,string)
aws_config.aws_region
string
AWS region code.
aws_config.aws_secret_access_key
string
AWS secret access key. Can be auto-detected, see the AWS boto3 docs for details.
aws_config.aws_session_token
string
AWS session token. Can be auto-detected, see the AWS boto3 docs for details.
aws_config.read_timeout
number
The timeout for reading from the connection (in seconds).
Default: 60
aws_config.aws_role
One of string, array
AWS roles to assume. If using the string format, the role ARN can be specified directly. If using the object format, the role can be specified in the RoleArn field and additional available arguments are documented at https://boto3.amazonaws.com/v1/documentation/api/latest/reference/services/sts.html?highlight=assume_role#STS.Client.assume_role
aws_config.aws_role.union
One of string, AwsAssumeRoleConfig
aws_config.aws_role.union.RoleArn ❓
string
ARN of the role to assume.
aws_config.aws_role.union.ExternalId
string
External ID to use when assuming the role.
profile_patterns
AllowDenyPattern
regex patterns for tables to profile
Default: {'allow': ['.*'], 'deny': [], 'ignoreCase': True}
profile_patterns.ignoreCase
boolean
Whether to ignore case sensitivity during pattern matching.
Default: True
profile_patterns.allow
array
List of regex patterns to include in ingestion
Default: ['.*']
profile_patterns.allow.string
string
profile_patterns.deny
array
List of regex patterns to exclude from ingestion.
Default: []
profile_patterns.deny.string
string
profiling
DataLakeProfilerConfig
Data profiling configuration
Default: {'enabled': False, 'operation_config': {'lower_fre...
profiling.enabled
boolean
Whether profiling should be done.
Default: False
profiling.include_field_distinct_value_frequencies
boolean
Whether to profile for distinct value frequencies.
Default: True
profiling.include_field_histogram
boolean
Whether to profile for the histogram for numeric fields.
Default: True
profiling.include_field_max_value
boolean
Whether to profile for the max value of numeric columns.
Default: True
profiling.include_field_mean_value
boolean
Whether to profile for the mean value of numeric columns.
Default: True
profiling.include_field_median_value
boolean
Whether to profile for the median value of numeric columns.
Default: True
profiling.include_field_min_value
boolean
Whether to profile for the min value of numeric columns.
Default: True
profiling.include_field_null_count
boolean
Whether to profile for the number of nulls for each column.
Default: True
profiling.include_field_quantiles
boolean
Whether to profile for the quantiles of numeric columns.
Default: True
profiling.include_field_sample_values
boolean
Whether to profile for the sample values for all columns.
Default: True
profiling.include_field_stddev_value
boolean
Whether to profile for the standard deviation of numeric columns.
Default: True
profiling.max_number_of_fields_to_profile
integer
A positive integer that specifies the maximum number of columns to profile for any table. None implies all columns. The cost of profiling goes up significantly as the number of columns to profile goes up.
profiling.profile_table_level_only
boolean
Whether to perform profiling at table-level only or include column-level profiling as well.
Default: False
profiling.operation_config
OperationConfig
Experimental feature. To specify operation configs.
profiling.operation_config.lower_freq_profile_enabled
boolean
Whether to do profiling at lower freq or not. This does not do any scheduling just adds additional checks to when not to run profiling.
Default: False
profiling.operation_config.profile_date_of_month
integer
Number between 1 to 31 for date of month (both inclusive). If not specified, defaults to Nothing and this field does not take affect.
profiling.operation_config.profile_day_of_week
integer
Number between 0 to 6 for day of week (both inclusive). 0 is Monday and 6 is Sunday. If not specified, defaults to Nothing and this field does not take affect.
stateful_ingestion
StatefulStaleMetadataRemovalConfig
Base specialized config for Stateful Ingestion with stale metadata removal capability.
stateful_ingestion.enabled
boolean
Whether or not to enable stateful ingest. Default: True if a pipeline_name is set and either a datahub-rest sink or datahub_api is specified, otherwise False
Default: False
stateful_ingestion.remove_stale_metadata
boolean
Soft-deletes the entities present in the last successful run but missing in the current run with stateful_ingestion enabled.
Default: True

Path Specs​

Path Specs (path_specs) is a list of Path Spec (path_spec) objects where each individual path_spec represents one or more datasets. Include path (path_spec.include) represents formatted path to the dataset. This path must end with *.* or *.[ext] to represent leaf level. If *.[ext] is provided then files with only specified extension type will be scanned. ".[ext]" can be any of supported file types. Refer example 1 below for more details.

All folder levels need to be specified in include path. You can use /*/ to represent a folder level and avoid specifying exact folder name. To map folder as a dataset, use {table} placeholder to represent folder level for which dataset is to be created. For a partitioned dataset, you can use placeholder {partition_key[i]} to represent name of ith partition and {partition[i]} to represent value of ith partition. During ingestion, i will be used to match partition_key to partition. Refer example 2 and 3 below for more details.

Exclude paths (path_spec.exclude) can be used to ignore paths that are not relevant to current path_spec. This path cannot have named variables ( {} ). Exclude path can have ** to represent multiple folder levels. Refer example 4 below for more details.

Refer example 5 if your bucket has more complex dataset representation.

Additional points to note

  • Folder names should not contain {, }, *, / in their names.
  • Named variable {folder} is reserved for internal working. please do not use in named variables.

Path Specs - Examples​

Example 1 - Individual file as Dataset​

Bucket structure:

test-bucket
├── employees.csv
├── departments.json
└── food_items.csv

Path specs config to ingest employees.csv and food_items.csv as datasets:

path_specs:
    - include: s3://test-bucket/*.csv

This will automatically ignore departments.json file. To include it, use *.* instead of *.csv.

Example 2 - Folder of files as Dataset (without Partitions)​

Bucket structure:

test-bucket
└──  offers
     ├── 1.avro
     └── 2.avro

Path specs config to ingest folder offers as dataset:

path_specs:
    - include: s3://test-bucket/{table}/*.avro

{table} represents folder for which dataset will be created.

Example 3 - Folder of files as Dataset (with Partitions)​

Bucket structure:

test-bucket
├── orders
│   └── year=2022
│       └── month=2
│           ├── 1.parquet
│           └── 2.parquet
└── returns
    └── year=2021
        └── month=2
            └── 1.parquet

Path specs config to ingest folders orders and returns as datasets:

path_specs:
    - include: s3://test-bucket/{table}/{partition_key[0]}={partition[0]}/{partition_key[1]}={partition[1]}/*.parquet

One can also use include: s3://test-bucket/{table}/*/*/*.parquet here however above format is preferred as it allows declaring partitions explicitly.

Example 4 - Folder of files as Dataset (with Partitions), and Exclude Filter​

Bucket structure:

test-bucket
├── orders
│   └── year=2022
│       └── month=2
│           ├── 1.parquet
│           └── 2.parquet
└── tmp_orders
    └── year=2021
        └── month=2
            └── 1.parquet


Path specs config to ingest folder orders as dataset but not folder tmp_orders:

path_specs:
    - include: s3://test-bucket/{table}/{partition_key[0]}={partition[0]}/{partition_key[1]}={partition[1]}/*.parquet
      exclude: 
        - **/tmp_orders/**

Example 5 - Advanced - Either Individual file OR Folder of files as Dataset​

Bucket structure:

test-bucket
├── customers
│   ├── part1.json
│   ├── part2.json
│   ├── part3.json
│   └── part4.json
├── employees.csv
├── food_items.csv
├── tmp_10101000.csv
└──  orders
    └── year=2022
        └── month=2
            ├── 1.parquet
            ├── 2.parquet
            └── 3.parquet

Path specs config:

path_specs:
    - include: s3://test-bucket/*.csv
      exclude:
        - **/tmp_10101000.csv
    - include: s3://test-bucket/{table}/*.json
    - include: s3://test-bucket/{table}/{partition_key[0]}={partition[0]}/{partition_key[1]}={partition[1]}/*.parquet

Above config has 3 path_specs and will ingest following datasets

  • employees.csv - Single File as Dataset
  • food_items.csv - Single File as Dataset
  • customers - Folder as Dataset
  • orders - Folder as Dataset and will ignore file tmp_10101000.csv

Valid path_specs.include

s3://my-bucket/foo/tests/bar.avro # single file table   
s3://my-bucket/foo/tests/*.* # mulitple file level tables
s3://my-bucket/foo/tests/{table}/*.avro #table without partition
s3://my-bucket/foo/tests/{table}/*/*.avro #table where partitions are not specified
s3://my-bucket/foo/tests/{table}/*.* # table where no partitions as well as data type specified
s3://my-bucket/{dept}/tests/{table}/*.avro # specifying keywords to be used in display name
s3://my-bucket/{dept}/tests/{table}/{partition_key[0]}={partition[0]}/{partition_key[1]}={partition[1]}/*.avro # specify partition key and value format
s3://my-bucket/{dept}/tests/{table}/{partition[0]}/{partition[1]}/{partition[2]}/*.avro # specify partition value only format
s3://my-bucket/{dept}/tests/{table}/{partition[0]}/{partition[1]}/{partition[2]}/*.* # for all extensions
s3://my-bucket/*/{table}/{partition[0]}/{partition[1]}/{partition[2]}/*.* # table is present at 2 levels down in bucket
s3://my-bucket/*/*/{table}/{partition[0]}/{partition[1]}/{partition[2]}/*.* # table is present at 3 levels down in bucket

Valid path_specs.exclude

  • **/tests/**
  • s3://my-bucket/hr/**
  • */tests/.csv
  • s3://my-bucket/foo/*/my_table/**

If you would like to write a more complicated function for resolving file names, then a {transformer} would be a good fit.

caution

Specify as long fixed prefix ( with out /*/ ) as possible in path_specs.include. This will reduce the scanning time and cost, specifically on AWS S3

caution

Running profiling against many tables or over many rows can run up significant costs. While we've done our best to limit the expensiveness of the queries the profiler runs, you should be prudent about the set of tables profiling is enabled on or the frequency of the profiling runs.

caution

If you are ingesting datasets from AWS S3, we recommend running the ingestion on a server in the same region to avoid high egress costs.

Compatibility​

Profiles are computed with PyDeequ, which relies on PySpark. Therefore, for computing profiles, we currently require Spark 3.0.3 with Hadoop 3.2 to be installed and the SPARK_HOME and SPARK_VERSION environment variables to be set. The Spark+Hadoop binary can be downloaded here.

For an example guide on setting up PyDeequ on AWS, see this guide.

caution

From Spark 3.2.0+, Avro reader fails on column names that don't start with a letter and contains other character than letters, number, and underscore. [https://github.com/apache/spark/blob/72c62b6596d21e975c5597f8fff84b1a9d070a02/connector/avro/src/main/scala/org/apache/spark/sql/avro/AvroFileFormat.scala#L158] Avro files that contain such columns won't be profiled.

Code Coordinates​

  • Class Name: datahub.ingestion.source.s3.source.S3Source
  • Browse on GitHub

Questions

If you've got any questions on configuring ingestion for S3 Data Lake, feel free to ping us on our Slack.