[HCoop-Discuss] VPS hosting for HCoop

[Adam Chlipala]

This discussion seems to have sputtered out.  Personally, I'm thinking 
of 100% VPS hosting (probably spread among multiple providers to some 
extent) as the best option for HCoop going forward.  Rob Gubler 
mentioned some file system/IO failures he's experienced with 
virtualization.  Shaun Empie said he doesn't think moving to 100% VPS is 
a good idea, but he didn't give any reasons, and he hasn't responded yet 
to a query about what they are.  Nathan Kennedy has been saying for a 
while that we should own our own servers.  Besides that, I think we've 
only seen positive opinions about the VPS option.

Here's my summary of the relevant issues, in decreasing order of importance:

VPS Pros:
- Virtual servers can be accessed from anywhere, as long as they're in 
good enough shape to be accessible via console.  We can always at least 
reboot them remotely for free.  This means that we lose _all_ geographic 
constraints on where admins need to be to handle any possible failure.  
Thus, we can have 100% time coverage for emergency response, without 
needing to find anyone willing to work the admin graveyard shift, since 
we can spread admins across time zones.  Also, even if our servers live 
in a relatively high cost-of-living area which has good 'net 
connectivity, we can still hire admins anywhere in the world.
- We don't have to purchase, diagnose, or replace hardware.  The VPS 
provider will do that for us.  We just need to be ready to move 
functionality from broken nodes reasonably quickly, probably using the 
provider's control panel interface.
- We can add new machines with a few control panel clicks, in response 
to load increases or anything else.
- For underutilized virtual machines, we probably save at least a little 
bit of money compared to dedicated servers or colocation of HCoop-owned 
servers.

VPS Cons:
- Crazy stuff might happen with simulated hardware.  I don't know how 
likely it is that we'd experience anything like this with one of the 
major providers, though, since they spend a lot of time tweaking kernels 
to avoid such trouble.
- We wouldn't own the machines.  (Is this really a problem?  Maybe it's 
just old-fashioned to think of owning machines instead of data.)
- We would be sharing machines with other customers, which could lead to 
not enough computational resources being available for us, in periods of 
high load.
- We would be sharing machines, so new potential security problems 
emerge.  (Interesting note on this and the last point: Linode apparently 
has an unlisted VPS kind big enough that it actually takes up its own 
8-core machine, for $800/mo..)

Did I miss anything?