如何使用无服务器将 librosa (python lib) 添加到 aws lambda

Question

I'm trying to use https://www.serverless.com/ to manage dependencies for a simple aws lambda function.

A key dependency is the librosa python library ( https://github.com/librosa/librosa ).

I am new to serverless , so I'd appreciate guidance with the serverless.yml file.

So far I have

service: my-app-123
app: my-app-123-app
org: my-org

frameworkVersion: '2'

provider:
  name: aws
  runtime: python3.6

stage: dev
region: eu-west-2

plugins:
  - serverless-python-requirements
functions:
  hello:
    handler: handler.hello

package:
  individually: false
  exclude:
    - '**/*'
  include:
    - handler.py

and requirements.txt is:

librosa==0.8.0
pandas==1.1.5

When I tried the above I got this error:

An error occurred: HelloLambdaFunction - Unzipped size must be smaller than 262144000 bytes (Service: AWSLambdaInternal; Status Code: 400; Error Code: InvalidParameterValueException; Request ID: f6bd001e-5592-44c8-8f10-3b56180d7ec7; Proxy: null).

Answer 1

You're hitting the size limitations of your AWS Lambda deployment package, see here . It can be maximal 250 MB unzipped (that's what your error message tells you).

Have a look at the section Dealing with Lambda's size limitations in the documentation of the serverless-python-requirements plugin.

In short what serverless-python-requirements recommends/supports is:

• Compress the dependency
• Add a Lambda Layer (eg librosa alone)

---

Compression works by adding the following in your serverless.yml

custom:
  pythonRequirements:
    zip: true

and then add this to your handler before using your dependencies:

try:
  import unzip_requirements
except ImportError:
  pass

---

But some users seem to have had issues with librosa and couldn't seem to make it work, see this GitHub issue for example or this other SO question .

A suggestion in there was to split up your Lambdas and chain them, ie one Lambda which has only librosa as a dependency (maybe with compression of the dependency) and then handing off for further processing to a second Lambda, I don't know what you're trying to do if this is feasible for you.