==== Zooko says: ====

Different users of Tahoe-LAFS have different desires for "Which servers should I upload which shares to?".

 * allmydata.com wants to upload to a random selection, evenly distributed among servers which are not full; This is, unsurprisingly, what Tahoe v1.4 currently does.
 * Brian has mentioned that an allmydata.com-style deployment might prefer to have the servers with more remaining capacity receiving more shares, thus "filling up faster" than the servers with less remaining capacity.
 * Kevin Reid wants, at least for one of his use cases, to specify several servers each of which is guaranteed to get at least K shares of each file, in addition to potentially other servers also getting shares.
 * Shawn Willden wants, likewise, to specify a server (e.g. his mom's PC) which is guaranteed to get at least K shares of certain files (the family pictures and movies files).
 * Some people -- I'm sorry I forget who -- have said they want to upload at least K shares to the K fastest servers.
 * Jake Appelbaum and Harold Gonzales want to specify a set of servers which collectively are guaranteed to have at least K shares -- they intend to use this to specify the ones that are running as Tor hidden services and thus are attack-resistant (but also extra slow-and-expensive to reach).  Interestingly the server selection policy on ''download'' should be that the K servers which are Tor hidden services should be downloaded from as a last resort. 
 * Several people -- again I'm sorry I've forgotten specific attribution -- want to identify which servers live in which cluster or co-lo or geographical area, and then to distribute shares evenly across clusters/colos/geographical-areas instead of evenly across servers.
  * Here's an example of this desire, Nathan Eisenberg asked on the mailing list for "Proximity Aware Decoding": http://allmydata.org/pipermail/tahoe-dev/2009-December/003286.html

As I have emphasized a few times, we really should not try to write a super-clever algorithm into Tahoe which satisfies all of these people, plus all the other crazy people that will be using Tahoe-LAFS for other things in the future.  Instead, we need some sort of configuration language or plugin system so that each crazy person can customize their own crazy server selection policy.  I don't know the best way to implement this yet -- a domain specific language?  Implement the above-mentioned list of seven policies into Tahoe-LAFS and have an option to choose which of the seven you want for this upload?  My current favorite approach is: you give me a Python function.  When the time comes to upload a file, I'll call that function and then use whichever servers it said to use.

==== Brian says: ====

Having a function or class to control server-selection is a great idea. The
current code already separates out responsibility for server-selection into a
distinct class, at least for immutable files
(source:src/allmydata/immutable/upload.py#L131 {{{Tahoe2PeerSelector}}}). It
would be pretty easy to make the uploader use different classes according to
a {{{tahoe.cfg}}} option.

However, there are some additional properties that need to be satified by the
server-selection algorithm for it to work at all. The basic Tahoe model is
that the filecap is both necessary and sufficient (given some sort of grid
membership) to recover the file. This means that the eventual
'''downloader''' needs to be able to find the same servers, or at least have
a sufficiently high probability of finding "enough" servers within a
reasonable amount of time, using only information which is found in the
filecap.

If the downloader is allowed to ask every server in the grid for shares, then
anything will work. If you want to keep the download setup time low, and/or
if you expect to have more than a few dozen servers, then the algorithm needs
to be able to do something better. Note that this is even more of an issue
for mutable shares, where it is important that publish-new-version is able to
track down and update all of the old shares: the chance of accidental
rollback increases when it cannot reliably/cheaply find them all.

Another potential goal is for the download process to be tolerant of new
servers, removed servers, and shares that have been moved (possibly as the
result of repair or "rebalancing"). Some use cases will care about this,
while others may never change the set of active servers and won't care.

It's worth pointing out the properties we were trying to get when we came up
with the current "TahoeTwo" algorithm:

 * for mostly static grids, download uses minimal do-you-have-share queries
 * adding one server should only increase download search time by 1/numservers
 * repair/rebalancing/migration may move shares to new places, including
   servers which weren't present at upload time, and download should be able
   to find and use these shares, even though the filecap doesn't change
 * traffic load-balancing: all non-full servers get new shares at the same
   bytes-per-second, even if serverids are not uniformly distributed

We picked the pseudo-random permuted serverlist [dependent on the StorageIndex]
to get these properties. I'd love to be able to get stronger diversity among
hosts, racks, or data centers, but I don't yet know how to get that '''and'''
get the properties listed above, while keeping the filecaps small.