[mahyarm]

The typical pattern for photo users will be a large amount of upload at the start as they upload their collection, then occasional browsing of photos on the amazon website, which will be mostly smaller thumbnails. Rarely once in awhile users will request large originals. You will have an even smaller amount of people download their entire collection of originals.

With these patterns, you can make a system that puts everything on S3 at first, makes thumbnails and starts moving most things to glacier based on access patterns.

[Veratyr]

Such a system would be wonderful, particularly if it used Backblaze B2 or Google Nearline instead of Glacier (to get around the hours of delay).

Unfortunately, no such system yet exists and the complexities of dealing with many people's libraries (particularly regarding RAW files) make it more difficult than it seems (I've looked at doing this myself in the past).