Is it faster to upload and download images into a MongoDB than on disk
Lightweight web servers (lighttpd, nginx) do a pretty good job of serving content from the filesystem. Since the OS acts as a caching layer they typically serve content from memory which is very fast.
If you want to serve images from mongodb the web server has to run some sort of script (python, php, ruby... of course FCGI, you can't start a new process for each image), which has to fetch data from mongodb each time the image is requested. So it's going to be slow? The benefits are automatic replication and failover if you use replica sets. If you need this and clever enough to know to achieve it with FS then go with that option... If you need a very quick implementation that's reliable then mongodb might be a faster way to do that. But if your site is going to be popular sooner or later you have to switch to the FS implementation.
BTW: you can mix these two approaches, store the image in mongodb to get instant reliability and then replicate it to the FS of a couple of servers to gain speed.
Oh one more thing.. coupling the metadata with the image seems to be nice until you realize the generated HTML and the image download is going to be two separate HTTP requests, so you have to query mongo twice, once for the metadata and once for the image.
When to use GridFS for storing files with MongoDB - the document suggests you should. It also sounds fast and reliable, and is great for backups and replication. Hope that helps.
Several benchmarks have shown MongoDB is approximately 6 times slower for file storage (via GridFS) versus using the regular old filesystem. (One compared apache, nginx, and mongo)
However, there are strong reasons to use MongoDB for file storage despite it being slower -- #1 free backup from Mongo's built-in sharding/replication. This is a HUGE time saver. #2 ease of admin, storing metadata, not having to worry about directories, permissions, etc. Also a HUGE time saver.
Our photo back-end was realized years ago in a huge gob of spaghetti code that did all kinds of stuff (check or create user dir, check or create date dirs, check for name collision, set perms), and a whole other mess did backups.
We've recently changed everything over to Mongo. In our experience, Mongo is a bit slower (it may be 6 times slower but it doesn't feel like 6 times slower), and anyway- so what? All that spaghetti is out the window, and the new Mongo+photo code is much smaller, tighter and logic simpler. Never going back to file system.