Boxed Ice is probably one of the larger MongoDB production deployments, and no doubt letting the DB handle the data growth will be a good thing for them. If anyone is going to run into issues deploying this, they will.
I have great hopes for Mongo 1.6: the automatic partioning and failover make it a very useful addition to startups' toolkits who need this kind of data storage. It's like having your very own Amazon SimpleDB that's cheaper to build and you control it 100%.
If sharding and repliaction is a major concern, I'd also take a look at riak or cassandra. Especially as a simple key-value store, they offer more interesting balancing and error recovery features than MongoDB.
MongoDB does however have a rich data-schema and allows queries (although the necessary indexes bring up ram consumption a fair bit and might become a problem)
Yeah I looked at them a little bit. The reason I went with Mongo for my purposes was because it works the way I thought of the data and also it has a very helpful community. It's also very easy to set up and get going with PHP.
Something that surprises me about MongoDB is that they always talk about how this is for humungous data sets.
But then MongoDB does not seem to support online/hot backups without setting the whole system to read only. Incremental backups are not supported at all, AFAICS. If a server crashes, you're up to some really lengthy fsck action before your server is back up.
I don't understand how people actually make this work. If you have a huge databases, a complete backup is going to take too long to make setting your system to read-only feasible. And if a server crashes, even if you have a hot standby replica, if you have to do several hours of fsck before you can restore regular operations, that doesn't sound good either.
It's nice to see movement in the database market, and MongoDB looks really good from a front-end side (JS, indexes, queries, ...); but I think it really is lacking some crucial features.
You don't have to set your whole system to read only for the entirety of the backup. There are three ways to avoid it:
1) Use mongodump which queries for every object in the database. It will produce a compacted file, but does not do a point-in-time snapshot.
2) Take a backup from a slave. You can either set it to read-only mode, or just shut it down and copy the files.
3) Run MongoDB on top of LVM. LVM supports almost instant snapshotting because it uses copy on write. Once the snapshot has been made (usually less than a few seconds) you can resume the server for read-write operations.
Note that all three options can be combined. You could make an LVM snapshot of a read slave then use mongodump's --dbpath option to produce compacted backup files.
I second the snapshotting using the disk system. It would be nice to be able to do this without locking the database, but LVM or ZFS snapshots take fractions of a second to create. In the mean time, pending transactions will just wait.
They intend for you to handle this by having read slaves (or using replica sets in 1.6). If your master fails, you can manually fail over to a read slave. Replica sets handle this for you automatically.
You can also pull a read slave out and do whatever you need to with it, then reconnect and have it catch up.
MongoDB is really designed to be run on at least 2 servers.
I have great hopes for Mongo 1.6: the automatic partioning and failover make it a very useful addition to startups' toolkits who need this kind of data storage. It's like having your very own Amazon SimpleDB that's cheaper to build and you control it 100%.
BTW, David gave a good talk about how Boxed Ice uses Mongo recently in London: http://skillsmatter.com/event/cloud-grid/mongouk