"My stuff wouldn't work on PostgreSQL at all, I have so much data that I must use sharding."
We had to implement sharding as well and chose to do it manually in PostgreSQL. Luckily our schema made that relatively easy and natural. YMMV
It's setup for multi box (each schema is mapped to a hostname in code) but I simply haven't had a reason to move to more boxes yet. The schema feature is a nice, convenient way to pre-shard like this so that growing to more boxes doesn't require rehashing for a very long time if ever (depending on how much sharding you do up front). You just move schemas/shards as needed using the standard dump and restore tools and update the schema->hostname mapping in the code.
Thankfully most of the joins happen within a shard (hashing and sharding on something like a user_id) with the exception being various analysis and aggregation queries.
Using PostgreSQL's schemas is admittedly not too different from just using many DBs in MySQL or something else but in practice I've found that extra layer of organization helps keep things neater. I can backup, move, or delete a specific schema/shard or I can backup, move, etc all shards on a machine by operating on the containing database.
You're absolutely right about the working set being the important bit to keep in memory (and all the indexes you use). I clarified that in person thanks to a great question from the audience.
My entire point wrt it slowing-down-outside-of-RAM is that it happens with any database. My point was that any database can be fast if it's in RAM and non-durable; yet Mongo trumpets performance under these unsafe scenarios while other databases use safe defaults.
To make matters worse, MongoDB data can suffer from poor data locality over time. Since caching is completely left up to the kernel, 1 hot document, even if it's only 200 bytes, will cause an entire page to remain in memory. The other bytes in the page may be cold documents or null data.
This is why flip/flopping is useful in current versions and in-place compaction so important for future versions. For prior research see defrag.exe.
I wish you luck in your attempts to tune performance by tweaking _id. I could only take doing hacks like that for so long before I looked for greener pastures.