As reported by a customer, when performing frequent lazy rendition exports of folderish documents that have many descendants, many of which have small content, many small blobs are downloaded from S3 to the LocalBlobStore before each can be added to the export archive. This takes a long time.
In an NFS-based environment, each file is directly copied from the NFS-based FileSystemBlobProvider to the export archive, which is fast. Though the NFS-based solution will significantly outperform any S3-based solution in almost all scenarios, we need to optimize the S3-based solution to the maximum extent. At a minimum, Nuxeo should be adding significant load to the network and/or the S3 store which is not currently the case.