Create an AWS Lambda Layer that includes pandas and pyarrow

cdk-snippet.ts

// Creating an AWS Lambda Layer with pandas and pyarrow is harder than it might seem,
// as simply `pip install pandas pyarrow` will lead to a deployment package that is > 250 MB
// which is not allowed by AWS Lambda.
// In this snippet, that deployment package is trimmed down, to make it fit (and still work)

import * as lambda from "aws-cdk-lib/aws-lambda";

const layerInstallCommand = [
  "bash",
  "-c",
  [
    "mkdir /asset-output/python",
    "pip install -r requirements.txt -t /asset-output/python",
    "rm -rf /asset-output/python/botocore", // Lambda runtime already has boto itself
    "rm -rf /asset-output/python/boto3", // Lambda runtime already has boto itself
    "rm -rf /asset-output/python/bin", // No need to execute CLI binaries
    "find /asset-output/python -name '*.so' -type f -exec strip \"{}\" \\;", // Not sure why this is needed, copied it from AWS Data Wrangler build
    "find /asset-output/python -d -regex '.*/tests' -exec rm -r {} +", // Get rid of tests
    "find /asset-output/python -d -regex '.*/__pycache__' -exec rm -r {} +", // Get rid of pycache
    "find /asset-output/python -type f -regex '^.*\\.py[co]$' -delete", // Get rid of pycache
  ].join(" && "),
];

const pandasLayer = new lambda.LayerVersion(self, "PandasLayer", {
  code: lambda.Code.fromAsset(
    "../pandas_layer", // Point this at a directory with a requirements.txt file with pandas and pyarrow
    {
      bundling: {
        image: lambda.Runtime.PYTHON_3_9.bundlingImage,
        command: layerInstallCommand,
      },
  }),
  compatibleRuntimes: [lambda.Runtime.PYTHON_3_9],
  compatibleArchitectures: [
    lambda.Architecture.X86_64, // Match this to the architecture of the machine where you are CDK synthing on (!)
  ],
});

@varunvilva-kickdrum Unfortunately I don't have the requirements.txt file anymore, but I have a better solution for you:

Instead of building a custom layer, use the AWS SDK for pandas (formerly AWS Data Wrangler) managed layer. It includes pyArrow pre-built and optimized for Lambda, and it now supports Python 3.13!

For Python 3.13, add this layer ARN to your Lambda function:

arn:aws:lambda::336392948345:layer:AWSSDKPandas-Python313:4

Just replace with your actual region (e.g., us-east-1).

This is much simpler than building a custom pyArrow layer and should resolve your "Cannot find the pyarrow.libs" error.

Documentation: https://aws-sdk-pandas.readthedocs.io/en/stable/layers.html

Note: Make sure your Lambda function has at least 512MB of memory when using this layer.

Thanks so much, @ottokruse ! Really appreciate your help and the suggestion — it worked perfectly. 🙏

Nice!

@varunvilva-kickdrum Unfortunately I don't have the requirements.txt file anymore, but I have a better solution for you:

Instead of building a custom layer, use the AWS SDK for pandas (formerly AWS Data Wrangler) managed layer. It includes pyArrow pre-built and optimized for Lambda, and it now supports Python 3.13!

For Python 3.13, add this layer ARN to your Lambda function:

arn:aws:lambda::336392948345:layer:AWSSDKPandas-Python313:4

Just replace with your actual region (e.g., us-east-1).

This is much simpler than building a custom pyArrow layer and should resolve your "Cannot find the pyarrow.libs" error.

Documentation: https://aws-sdk-pandas.readthedocs.io/en/stable/layers.html

Note: Make sure your Lambda function has at least 512MB of memory when using this layer.

Great solution, thank you for the detailed info!