Use MultipartUpload for uploading chunks to s3

[juliusknorr]

Implements #19414

This allows chunked upload streaming the chunks directly to S3 so that the final MOVE request no longer needs to assemble the final file on the Nextcloud server.

Ref: https://docs.aws.amazon.com/AmazonS3/latest/userguide/mpuoverview.html

Current API changes to trigger the new upload mechanism (required changes on the client side)

The DAV capabilities expose a feature flag s3-multipart if available.

Additional requirements on top of the current chunking API in order to trigger it:

• New additional header is required on all chunked upload requests (MKCOL/PUT/MOVE)
  • Destination: http://nextcloud.dev.local/remote.php/dav/files/admin/myfiletarget
• Limitations
  • Chunk filename must be numberic between 1 and 10000
  • Chunks must be at least 5MB of size (except for the last one)
  • Chunks cannot be downloaded from the upload directory

Summary of how S3 handles MultipartUpload

• CreateMultipartUpload is called with the destination bucket, object key and will return an uploadId
• UploadPart gets called for each part with the bucket, objectKey, uploadId, contentLength and the content, it returns a set of an ETag and the UploadId which need to be stored for the CompleteMultipartUpload
• CompleteMultipartUpload Receives the list of parts and merges them directly on S3 to the final object

How it currently works

• MKCOL Create the UploadFolder, initiate the MultipartUpload through the IChunkedFileWrite storage and persist the returned chunkToken (uploadId) in the dav properties of the folder
  • If the destination file does not exist we create a .target file in the upload folder to have a filecache entry
  • The destination is required upfront to allow writing directly to the target file
• PUT Writes the stream data directly to the S3 object through the IChunkedFileWrite storage
  • Validates the DAV chunk filename name to be numeric between 1 and 10000
• MOVE the MultipartUpload will be completed
  • if the file exists we are done
  • if the file didn't exist, call rename to move the .target file from the upload folder to the destination
• If a storage doesn't exist the new API calls will just fallback to the previous chunking mechanism

Test script

Bash script to test the upload

#!/bin/bash
# Usage: ./upload.sh targetname1
# Uploads a file consisting of 3 chunks and downloads it a again to verify the upload

if [ -z "$1" ]
  then
    echo "Usage: ./s3-chunk filepath"
	exit 1
fi

CHUNKSIZE=10
CHUNKS=2
> /tmp/one.testbin
for i in $(seq 1 $CHUNKS)
do
	CHUNKSIZE=$(( $RANDOM % 10 + 5 ))
	dd if=/dev/random of=/tmp/$i.testbin bs=${CHUNKSIZE}M count=1 2>/dev/null
	cat /tmp/$i.testbin >> /tmp/one.testbin
done

ls -l /tmp/*.testbin

SERVER=http://nextcloud.dev.local
USER=admin
PASS=admin
FILENAME=$1
timestamp=$(date +%s)

#echo "=> Cleanup existing file"
#curl -X DELETE -u $USER:$PASS "$SERVER/remote.php/dav/files/$USER/$FILENAME"



DESTINATION="$SERVER/remote.php/dav/files/admin/$FILENAME"

echo "Uploading to $USER:$PASS $SERVER/remote.php/dav/uploads/$USER/$timestamp"
echo "=> MKCOL"
curl -X MKCOL -u $USER:$PASS $SERVER/remote.php/dav/uploads/$USER/$timestamp \
	-H "Destination: $DESTINATION"

echo "=> PUT"
for i in $(seq 1 $CHUNKS | shuf)
do
	echo "- PART $i"
	(curl -X PUT -u $USER:$PASS $SERVER/remote.php/dav/uploads/$USER/$timestamp/$i -T /tmp/$i.testbin -H "X-S3-Multipart-Destination: files/admin/$FILENAME" \
		-H "Destination: $DESTINATION"; echo " - DONE $i") &
done

wait

# exit 1;

echo "=> MOVE"
curl -X MOVE -u $USER:$PASS --header "Destination:$SERVER/remote.php/dav/files/$USER/$FILENAME" $SERVER/remote.php/dav/uploads/$USER/$timestamp/.file \
	-H "Destination: $DESTINATION"

curl -u $USER:$PASS $SERVER/remote.php/dav/files/$USER/$FILENAME > /tmp/compare.testbin

md5sum /tmp/one.testbin
md5sum /tmp/compare.testbin

ls -l /tmp/*.testbin

rm /tmp/*.testbin

[juliusknorr]

@nextcloud/server-backend I didn't get that far in finishing this as expected, but some intermediate feedback on the approach with the storage abstraction and the chunking plugin would already be much appreciated.

[blizzz]

/backport to stable26

[pierreozoux]

🚀

This is amazing :) Thanks!

(hope it is ok to "pollute" github PR with praising :) it is easy to complain about bugs, but we should also learn to celebrate :) )

[backportbot-nextcloud]

The backport to stable-23 failed. Please do this backport manually.

[backportbot-nextcloud]

The backport to stable26 failed. Please do this backport manually.

[ir0nhide]

@juliushaertl great work, thank you very much!
Looking at the newly added code in https://github.com/nextcloud/server/blob/master/lib/private/Files/ObjectStore/S3.php, i am wondering if it is also required to add SSE-C headers here like done in e.g. https://github.com/nextcloud/server/blob/master/lib/private/Files/ObjectStore/S3ObjectTrait.php#L62 to support #32798 with multipart upload?

[backportbot-nextcloud]

The backport to stable-22 failed. Please do this backport manually.

[backportbot-nextcloud]

The backport to stable22 failed. Please do this backport manually.

[backportbot-nextcloud]

The backport to stable23 failed. Please do this backport manually.

[backportbot-nextcloud]

The backport to stable24 failed. Please do this backport manually.

[juliusknorr]

@juliushaertl great work, thank you very much!
Looking at the newly added code in master/lib/private/Files/ObjectStore/S3.php, i am wondering if it is also required to add SSE-C headers here like done in e.g. master/lib/private/Files/ObjectStore/S3ObjectTrait.php#L62 to support #32798 with multipart upload?

Thanks, good catch indeed. I'll take care of that in the follow up.

[unteem]

Awesome feature :) ! Thanks a lot

Are there any reasons for it not being backported to v25 ?

[AndyScherzinger]

We usually only backport security fixes or critical bugfixes. Any kind of performance improvement as well as features gets pinned to the upcoming major release.