DataFrames on AWS

Read the Docs!

Read the Tutorials: Catalog & Metadata | Athena Nested | S3 Write Modes

---

Contents: Use Cases | Installation | Examples | Diving Deep | Step By Step | Contributing

---

Use Cases

Pandas

• Pandas -> Parquet (S3) (Parallel)
• Pandas -> CSV (S3) (Parallel)
• Pandas -> Glue Catalog Table
• Pandas -> Athena (Parallel)
• Pandas -> Redshift (Append/Overwrite/Upsert) (Parallel)
• Pandas -> Aurora (MySQL/PostgreSQL) (Append/Overwrite) (Via S3) (NEW ⭐)
• Parquet (S3) -> Pandas (Parallel)
• CSV (S3) -> Pandas (One shot or Batching)
• Glue Catalog Table -> Pandas (Parallel)
• Athena -> Pandas (One shot, Batching or Parallel)
• Redshift -> Pandas (Parallel)
• CloudWatch Logs Insights -> Pandas
• Aurora -> Pandas (MySQL) (Via S3) (NEW ⭐)
• Encrypt Pandas Dataframes on S3 with KMS keys
• Glue Databases Metadata -> Pandas (Jupyter output compatible)
• Glue Table Metadata -> Pandas (Jupyter output compatible)

PySpark

• PySpark -> Redshift (Parallel)
• Register Glue table from Dataframe stored on S3
• Flatten nested DataFrames

General

• List S3 objects (Parallel)
• Delete S3 objects (Parallel)
• Delete listed S3 objects (Parallel)
• Delete NOT listed S3 objects (Parallel)
• Copy listed S3 objects (Parallel)
• Get the size of S3 objects (Parallel)
• Get CloudWatch Logs Insights query results
• Load partitions on Athena/Glue table (repair table)
• Create EMR cluster (For humans)
• Terminate EMR cluster
• Get EMR cluster state
• Submit EMR step(s) (For humans)
• Get EMR step state
• Get EMR step state
• Athena query to receive the result as python primitives (Iterable[Dict[str, Any])
• Load and Unzip SageMaker jobs outputs
• Load and Unzip SageMaker models
• Redshift -> Parquet (S3)
• Aurora -> CSV (S3) (MySQL) (NEW ⭐)
• Get Glue Metadata

Installation

pip install awswrangler

Runs only with Python 3.6 and 3.7.

Runs anywhere (AWS Lambda, AWS Glue Python Shell, EMR, EC2, on-premises, local, etc).

P.S. Lambda Layer's bundle and Glue's wheel/egg are available to download. Just upload it and run! 🚀

P.P.S. Have you never used Layers? Check the step-by-step guide.

P.P.P.S. AWS Data Wrangler counts on compiled dependencies (C/C++) so there is no support for Glue PySpark by now (Only Glue Python Shell).

Examples

Pandas

Writing Pandas Dataframe to S3 + Glue Catalog

import awswrangler as wr

wr.pandas.to_parquet(
    dataframe=df,
    database="database",
    path="s3://...",
    partition_cols=["col_name"],
)

If a Glue Database name is passed, all the metadata will be created in the Glue Catalog. If not, only the s3 data write will be done.

Writing Pandas Dataframe to S3 as Parquet encrypting with a KMS key

import awswrangler as wr

extra_args = {
    "ServerSideEncryption": "aws:kms",
    "SSEKMSKeyId": "YOUR_KMY_KEY_ARN"
}
sess = wr.Session(s3_additional_kwargs=extra_args)
sess.pandas.to_parquet(
    path="s3://..."
)

Reading from AWS Athena to Pandas

import awswrangler as wr

df = wr.pandas.read_sql_athena(
    sql="select * from table",
    database="database"
)

Reading from AWS Athena to Pandas in chunks (For memory restrictions)

import awswrangler as wr

df_iter = wr.pandas.read_sql_athena(
    sql="select * from table",
    database="database",
    max_result_size=512_000_000  # 512 MB
)

for df in df_iter:
    print(df)  # Do whatever you want

Reading from AWS Athena to Pandas with the blazing fast CTAS approach

import awswrangler as wr

sess = wr.Session(athena_ctas_approach=True)
df = sess.pandas.read_sql_athena(
    sql="select * from table",
    database="database"
)

Reading from Glue Catalog (Parquet) to Pandas

import awswrangler as wr

df = wr.pandas.read_table(database="DATABASE_NAME", table="TABLE_NAME")

Reading from S3 (Parquet) to Pandas

import awswrangler as wr

df = wr.pandas.read_parquet(path="s3://...", columns=["c1", "c3"], filters=[("c5", "=", 0)])

Reading from S3 (CSV) to Pandas

import awswrangler as wr

df = wr.pandas.read_csv(path="s3://...")

Reading from S3 (CSV) to Pandas in chunks (For memory restrictions)

import awswrangler as wr

df_iter = wr.pandas.read_csv(
    path="s3://...",
    max_result_size=512_000_000  # 512 MB
)

for df in df_iter:
    print(df)  # Do whatever you want

Reading from CloudWatch Logs Insights to Pandas

import awswrangler as wr

df = wr.pandas.read_log_query(
    log_group_names=[LOG_GROUP_NAME],
    query="fields @timestamp, @message | sort @timestamp desc | limit 5",
)

Typical Pandas ETL

import pandas
import awswrangler as wr

df = pandas.read_...  # Read from anywhere

# Typical Pandas, Numpy or Pyarrow transformation HERE!

wr.pandas.to_parquet(  # Storing the data and metadata to Data Lake
    dataframe=df,
    database="database",
    path="s3://...",
    partition_cols=["col_name"],
)

Loading Pandas Dataframe to Redshift

import awswrangler as wr

wr.pandas.to_redshift(
    dataframe=df,
    path="s3://temp_path",
    schema="...",
    table="...",
    connection=con,
    iam_role="YOUR_ROLE_ARN",
    mode="overwrite",
    preserve_index=False,
)

Extract Redshift query to Pandas DataFrame

import awswrangler as wr

df = wr.pandas.read_sql_redshift(
    sql="SELECT ...",
    iam_role="YOUR_ROLE_ARN",
    connection=con,
    temp_s3_path="s3://temp_path")

Loading Pandas Dataframe to Aurora (MySQL/PostgreSQL)

import awswrangler as wr

wr.pandas.to_aurora(
    dataframe=df,
    connection=con,
    schema="...",
    table="..."
)

Extract Aurora query to Pandas DataFrame (MySQL)

import awswrangler as wr

df = wr.pandas.read_sql_aurora(
    sql="SELECT ...",
    connection=con
)

PySpark

Loading PySpark Dataframe to Redshift

import awswrangler as wr

wr.spark.to_redshift(
    dataframe=df,
    path="s3://...",
    connection=conn,
    schema="public",
    table="table",
    iam_role="IAM_ROLE_ARN",
    mode="append",
)

Register Glue table from Dataframe stored on S3

import awswrangler as wr

dataframe.write \
        .mode("overwrite") \
        .format("parquet") \
        .partitionBy(["year", "month"]) \
        .save(compression="gzip", path="s3://...")
sess = wr.Session(spark_session=spark)
sess.spark.create_glue_table(
    dataframe=dataframe,
    file_format="parquet",
    partition_by=["year", "month"],
    path="s3://...",
    compression="gzip",
    database="my_database")

Flatten nested PySpark DataFrame

import awswrangler as wr

sess = awswrangler.Session(spark_session=spark)
dfs = sess.spark.flatten(dataframe=df_nested)
for name, df_flat in dfs.items():
    print(name)
    df_flat.show()

General

Deleting a bunch of S3 objects (parallel)

import awswrangler as wr

wr.s3.delete_objects(path="s3://...")

Get CloudWatch Logs Insights query results

import awswrangler as wr

results = wr.cloudwatchlogs.query(
    log_group_names=[LOG_GROUP_NAME],
    query="fields @timestamp, @message | sort @timestamp desc | limit 5",
)

Load partitions on Athena/Glue table (repair table)

import awswrangler as wr

wr.athena.repair_table(database="db_name", table="tbl_name")

Create EMR cluster

import awswrangler as wr

cluster_id = wr.emr.create_cluster(
    cluster_name="wrangler_cluster",
    logging_s3_path=f"s3://BUCKET_NAME/emr-logs/",
    emr_release="emr-5.27.0",
    subnet_id="SUBNET_ID",
    emr_ec2_role="EMR_EC2_DefaultRole",
    emr_role="EMR_DefaultRole",
    instance_type_master="m5.xlarge",
    instance_type_core="m5.xlarge",
    instance_type_task="m5.xlarge",
    instance_ebs_size_master=50,
    instance_ebs_size_core=50,
    instance_ebs_size_task=50,
    instance_num_on_demand_master=1,
    instance_num_on_demand_core=1,
    instance_num_on_demand_task=1,
    instance_num_spot_master=0,
    instance_num_spot_core=1,
    instance_num_spot_task=1,
    spot_bid_percentage_of_on_demand_master=100,
    spot_bid_percentage_of_on_demand_core=100,
    spot_bid_percentage_of_on_demand_task=100,
    spot_provisioning_timeout_master=5,
    spot_provisioning_timeout_core=5,
    spot_provisioning_timeout_task=5,
    spot_timeout_to_on_demand_master=True,
    spot_timeout_to_on_demand_core=True,
    spot_timeout_to_on_demand_task=True,
    python3=True,
    spark_glue_catalog=True,
    hive_glue_catalog=True,
    presto_glue_catalog=True,
    bootstraps_paths=None,
    debugging=True,
    applications=["Hadoop", "Spark", "Ganglia", "Hive"],
    visible_to_all_users=True,
    key_pair_name=None,
    spark_jars_path=[f"s3://...jar"],
    maximize_resource_allocation=True,
    keep_cluster_alive_when_no_steps=True,
    termination_protected=False,
    spark_pyarrow=True,
    tags={
        "foo": "boo"
    }
)
print(cluster_id)

Athena query to receive the result as python primitives (Iterable[Dict[str, Any])

import awswrangler as wr

for row in wr.athena.query(query="...", database="..."):
    print(row)

Load and unzip SageMaker job output

import awswrangler as wr

outputs = wr.sagemaker.get_model("JOB_NAME")

Load and unzip SageMaker job output

import awswrangler as wr

outputs = wr.sagemaker.get_job_outputs("JOB_NAME")

Diving Deep

Relational Databases (SQL) - (Oracle, PostgreSQL, MySQL, Microsoft SQL Server, etc)

Pandas and PySpark already have great interfaces to handle integrations with relational databases:

1. https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.to_sql.html
2. https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.read_sql.html
3. https://spark.apache.org/docs/latest/api/python/pyspark.sql.html?highlight=jdbc#pyspark.sql.DataFrameReader.jdbc
4. https://spark.apache.org/docs/latest/api/python/pyspark.sql.html?highlight=jdbc#pyspark.sql.DataFrameWriter.jdbc

AWS Data Wrangler does not want to reinvent the wheel. And will only implement the integrations not covered by the natives Pandas and PySpark APIs.

E.g.:

• MySQL Aurora LOAD and UNLOAD through S3
• PostgreSQL Aurora COPY through aws_s3 extension and the S3 service itself

Parallelism, Non-picklable objects and GeoPandas

AWS Data Wrangler tries to parallelize everything that is possible (I/O and CPU bound task). You can control the parallelism level using the parameters:

• procs_cpu_bound: number of processes that can be used in single node applications for CPU bound case (Default: os.cpu_count())
• procs_io_bound: number of processes that can be used in single node applications for I/O bound cases (Default: os.cpu_count() * PROCS_IO_BOUND_FACTOR)

Both can be defined on Session level or directly in the functions.

Some special cases will not work with parallelism:

• GeoPandas
• Columns with non-picklable objects

To handle that use procs_cpu_bound=1 and avoid the distribution of the dataframe.

Pandas with null object columns (UndetectedType exception)

Pandas has a too generic "data type" named object. Pandas object columns can be string, dates, etc, etc, etc. We can handle this object column fine inferring the types of theses objects inside the values, Pyarrow does that like a charm. So the real problem starts when we have a completely null object column because we don't have anything to infer.

To work with null object columns you can explicitly set the expected Athena data type for the target table doing:

import awswrangler as wr
import pandas as pd

df = pd.DataFrame({
    "col": [1, 2],
    "col_string_null": [None, None],
    "col_date_null": [None, None],
})

wr.pandas.to_parquet(
    dataframe=df,
    database="DATABASE",
    path=f"s3://...",
    cast_columns={
      "col_string_null": "string",
      "col_date_null": "date"
    })

Athena to Pandas Flow (PARALLEL/CTAS)

Pandas to Redshift Flow

Spark to Redshift Flow

Step By Step

Setting Up Lambda Layer

Go to GitHub's release section and download the layer bundle related to the desired version. Also select between Python 3.6 or 3.7.

Go to the AWS console and open the S3 panel. Upload the layer bundle to any S3 bucket in the desired AWS region.

Copy the S3 object URL.

Go to the AWS Lambda Panel, get in the layer's section (left side) and click to create one.

Fill the fields (Use the pasted URL) and create your layer.

Go to your AWS Lambda and use it!

Contributing

• AWS Data Wrangler practically only makes integrations. So we prefer to dedicate our energy / time writing integration tests instead of unit tests. We really like an end-to-end approach for all features.

• All integration tests are between a local Docker container and a remote/real AWS service.

• We have a Docker recipe to set up the local end (testing/Dockerfile).

• We have a Cloudformation to set up the AWS end (testing/template.yaml).

Steps

DISCLAIMER: Make sure to know what you are doing. This steps will charge some services on your AWS account. And requires a minimum security skills to keep your environment safe.

• Pick up a Linux or MacOS.

• Install Python 3.6+

• Install Docker and configure at least 4 cores and 8 GB of memory

• Fork the AWS Data Wrangler repository and clone that into your development environment

• Go to the project's directory create a Python's virtual environment for the project (python -m venv venv && source venv/bin/activate)

• Run ./setup-dev-env.sh

• Go to the testing directory

• Configure the parameters.json file with your AWS environment infos (Make sure that your Redshift will not be open for the World! Configure your security group to only give access for your IP.)

• Deploy the Cloudformation stack ./deploy-cloudformation.sh

• Open the docker image ./open-image.sh

• Inside the image you finally can run ./run-tests.sh