And, yes, the speed improvements to the ruby driver are very much appreciated :)
* Does it matter if you lose the last 5sec. worth of updates? The last 5 minutes? The last day?
If you can lose 5sec. worth of updates, a MongoDB replication pair is just fine. If you can lose a day's worth of updates (or can easily reconstruct the database contents from other sources), you can try out pretty much anything without bad repercussions. If you can't lose anything, you're pretty much limited to the most conservative databases (the SQL bunch).
* What's the most obvious unit of data that you're working with?
If you always update single values (or add things to lists/sets), Redis is an excellent choice. If you have fixed-size records, SQL or one of the table-based options (Cassandra, Hbase) may be for you. If you have documents with substantial internal structure, a document store (MongoDB, CouchDB, or Lotus Notes if you want something expensive and commercial) would be a good option.
* How much data do you have?
If all of your data fits into memory (and for the price of another server, you may well get enough memory to fit all of your data), you can go pretty far with a single server. If it fits on a single set of hard disks, you'd want replication, not sharding, so that the risk of losing data is minimal. If your data is much larger than that, your only hope is a sharding setup - either with SQL+spit+glue, or Cassandra/HBase, or some version of MongoDB where sharding is stable enough for production use (I do remember seeing warnings - so the current version may or may not fit that description).
It is absolutely OK to lose analytics data occasionally, and indeed with the variety of ways to bork that (js is off, user agent prefetches undisplayed page, bot action, etc) if your stats aren't robust against it you are screwed anyhow.
For instance, I routinely delete web server logs older than 30 days, on the assumption that if I didn't need it in the last 30 days I'll probably never need it. Every now and then this bites me and I need more than 30 days data to test some assumption, then I will just have to wait for a bit. (for events that occur infrequently enough).
Website visit counters are a great example. I don't think many people care if a few visits get lost once in a while.
We use MongoDB for storing tons and tons of analytics data for which we don't care if some stuff occasionally gets lost in a server crash. The data really fits MongoDB well and it would have been a nightmare if we were to use an SQL database for this. But for bank transactions we wouldn't even consider MongoDB.
The write lock might be a problem for some people. On the other hand MongoDB supports easy sharding, much easier than with SQL. Sharding allows us to scale horizontally which is a huge plus for our data.