Question

I’m working on a machine with limited memory, and I’d like to upload a dynamically generated (not-from-disk) file in a streaming manner to S3. In other words, I don’t know the file size when I start the upload, but I’ll know it by the end. Normally a PUT request has a Content-Length header, but perhaps there is a way around this, such as using multipart or chunked content-type.

S3 can support streaming uploads. For example, see here:

http://blog.odonnell.nu/posts/streaming-uploads-s3-python-and-poster/

My question is, can I accomplish the same thing without having to specify the file length at the start of the upload?

Answer 1

You have to upload your file in 5MiB+ chunks via S3’s multipart API. Each of those chunks requires a Content-Length but you can avoid loading huge amounts of data (100MiB+) into memory.

• Initiate S3 Multipart Upload.
• Gather data into a buffer until that buffer reaches S3’s lower chunk-size limit (5MiB). Generate MD5 checksum while building up the buffer.
• Upload that buffer as a Part, store the ETag (read the docs on that one).
• Once you reach EOF of your data, upload the last chunk (which can be smaller than 5MiB).
• Finalize the Multipart Upload.

S3 allows up to 10,000 parts. So by choosing a part-size of 5MiB you will be able to upload dynamic files of up to 50GiB. Should be enough for most use-cases.

However: If you need more, you have to increase your part-size. Either by using a higher part-size (10MiB for example) or by increasing it during the upload.

First 25 parts:   5MiB (total:  125MiB)
Next 25 parts:   10MiB (total:  375MiB)
Next 25 parts:   25MiB (total:    1GiB)
Next 25 parts:   50MiB (total: 2.25GiB)
After that:     100MiB

This will allow you to upload files of up to 1TB (S3’s limit for a single file is 5TB right now) without wasting memory unnecessarily.

---

A note on your link to Sean O’Donnells blog:

His problem is different from yours – he knows and uses the Content-Length before the upload. He wants to improve on this situation: Many libraries handle uploads by loading all data from a file into memory. In pseudo-code that would be something like this:

data = File.read(file_name)
request = new S3::PutFileRequest()
request.setHeader('Content-Length', data.size)
request.setBody(data)
request.send()

His solution does it by getting the Content-Length via the filesystem-API. He then streams the data from disk into the request-stream. In pseudo-code:

upload = new S3::PutFileRequestStream()
upload.writeHeader('Content-Length', File.getSize(file_name))
upload.flushHeader()

input = File.open(file_name, File::READONLY_FLAG)

while (data = input.read())
  input.write(data)
end

upload.flush()
upload.close()