They seems like to talk only about performance, which I believe it's not the case compared to other hybrid RAID types.
What are you asking -- "why does MongoDB recommend EBS over local disks" or "why does MongoDB recommend RAID-10 instead of a single EBS?" I think both questions are fairly well answered in the link you gave, but to recap:
- EBS vs local disk: This is a no-brainer if you want your MongoDB data to survive an instance crash or reboot. Local disk goes away when the instance does.
- RAID-10 of EBS: Put simply, EBS performance is sh...ockingly poor (and, worse, variable). A RAID-10 at least spreads the load across a lot of volumes.
Edit: So, there's a third question then. The actual benchmarking data linked to from the link in your comment is gone, so we only have the commentary, but I'd say it's a bad benchmark because the logic involved is wrong. The per-block I/O needed for different RAID levels are:
- RAID1: N writes across N disks
- RAID5: N-2 reads (collect existing, unchanged data in the stripe -- possibly cached, if you're lucky), 2 writes across N disks
- RAID10: 2 writes across N disks
I can't explain why the benchmark got the results it did, because the data is gone, but my guess is that not enough test runs were done to smooth out the inherent variability in testing a shared resource. I did a lot of this sort of work for Heroku way back when, and the variability in runs over time (and over volumes) was sh...ocking.
As for the wikipedia article you cite, it's irrelevant in this case because it's all about RAID-5 in a high-end SAN appliance; you're doing RAID-5 inside a VM, which lacks the ability to make "the task of parity recalculation and redistribution [...] performed as a back-end process transparent to the host" -- amongst a variety of other differences between "big-ass SAN" and "software RAID".
10gen did a webinar not too long ago specifically about using MongoDB on AWS. It deals with two issues: (1) Performance and (2) Fault tolerance in the wake of the recent AWS multi-day outage in a single availability zone. Their solution is RAID 10 across many nodes.
I'd suggest you go through the slides, as they can explain much better than I can in a SF post.