s3zipper is a couple of Lambda functions that let you use S3 Object Lambda Access Points to download ZIP files of directories[1] in S3 on-demand. Once deployed, you don't need any special tools installed on your client - anything that can download from S3 will do.

First, deploy the template:

sam build && sam deploy \
  --stack-name myolap \
  --resolve-s3 \
  --capabilities CAPABILITY_IAM \
  --parameter-overrides \
      ParameterKey=BucketName,ParameterValue=my-target-s3-bucket \
      ParameterKey=RealContentLength,ParameterValue=false

Second, you'll need to get the alias of the access point. It would be nice if this was emitted by the stack deployment, but I think there's a bug in the CloudFormation resource definition. So you have to run this command:

aws s3control get-access-point-for-object-lambda \
  --account-id $(aws sts get-caller-identity --output text --query Account) \
  --name myolap \
  --output text \
  --query Alias.Value

That will return a name like myolap-somerandomstring--ol-s3. You can think of that as the name of the "proxy bucket" that will invoke s3zipper.

Finally, download a ZIP! If you have a bunch of objects at s3://my-target-s3-bucket/prefix/to/files/..., then you can download a ZIP like this:

aws s3 cp s3://myolap-somerandomstring--ol-s3/prefix/to/files.zip .

[1]: They're technically prefixes, S3 doesn't have directories in a file-system sense. But it's close enough.

You will see progress that looks like this:

Completed 539.2 MiB/1.0 KiB (29.3 MiB/s) with 1 file(s) remaining

This can be fixed by setting the RealContentLength stack parameter to true. The issue is that for files over 5MB, the AWS CLI will download the file in multiple chunks in parallel for improved performance - but s3zipper doesn't support the Range requests functionality for that. We work around that limitation by lying about the file size in the Content-Length header. If you enable the real Content-Length, you should force the CLI to use only one chunk with this:

aws configure set default.s3.multipart_threshold 5TB

S3 Object Lambda access points do provide functionality to ensure that the end-user is correctly attributed in CloudTrail, but it doesn't work for our use-case as we need to initiate many s3:GetObject calls (for each file in the ZIP) in response to the user's original API call.

S3 OLAP enforces a 60 second timeout on the Lambda function. That means that the largest file that can be returned is determined by how much data can be sent in that time window. The bandwidth depends on the number of objects in the ZIP and how large they are. In my limited testing, I've seen rates between 10 MB/s and 60 MB/s, which translates to a maximum ZIP size of 600MB - 3.6 GB.

Here's a sequence diagram that explains how it works. The key thing to note is that everything is streamed, so the user starts to receive the ZIP file as soon as the first object is retrieved from S3. Streaming also means that memory usage in the Lambda function never grows above 40 MB even for a ZIP file that is multiple gigabytes.

sequenceDiagram
participant U as User
participant O as S3 Object Lambda Access Point
participant F as s3zipper Lambda Function
participant B as Target Bucket
U->>O: GetObject(Bucket=alias-name, Key=my/prefix.zip)
O->>F: GetObject(Bucket=alias-name, Key=my/prefix.zip)
F->>O: start WriteGetObjectResponse stream
note left of F: This happens concurrently with the loop below
O->>U: start streaming ZIP file response 
F->>B: ListObjectsV2(Bucket=target-bucket, Prefix=my/prefix)
B->>F: <list of objects>
loop For each object
F->>B: GetObject(Bucket=target-bucket, Key=my/prefix/${object})
B->>F: <contents of object>
F->>O: append object to WriteGetObjectResponse stream
end
F->>O: Write ZIP file metadata to stream and close stream
O->>U: finish sending file to user