Allmydata "Tahoe" Architecture

OVERVIEW

At a high-level this system consists of three layers: the grid, the
filesystem, and the application.

The lowest layer is the "grid", a mapping from capabilities to
data. The capabilities are relatively short ascii strings, each used
as a reference to an arbitrary-length sequence of data bytes. This
data is encrypted and distributed across a number of nodes, such that
it will survive the loss of most of the nodes.

The middle layer is the decentralized filesystem: a directed graph in
which the intermediate nodes are directories and the leaf nodes are
files. The leaf nodes contain only the file data -- they contain no
metadata about the file other than the length.  The edges leading to
leaf nodes have metadata attached to them about the file they point
to.  Therefore, the same file may be associated with different
metadata if it is dereferenced through different edges.

The top layer consists of the applications using the filesystem.
Allmydata.com uses it for a backup service: the application
periodically copies files from the local disk onto the decentralized
filesystem We later provide read-only access to those files, allowing
users to recover them.  The filesystem can be used by other
applications, too.


THE GRID OF STORAGE SERVERS

The grid is composed of peer nodes -- processes running on
computers.  They establish TCP connections to each other using Foolscap, a
secure remote message passing library.

Each peer offers certain services to the others. The primary service
is that of the storage server, which holds data in the form of
"shares".  Shares are encoded pieces of files.  There are a
configurable number of shares for each file, 10 by default.  Normally,
each share is stored on a separate server, but a single server can
hold multiple shares for a single file.

Peers learn about each other through an "introducer". Each peer
connects to a central introducer at startup, and receives a list of
all other peers from it. Each peer then connects to all other peers,
creating a fully-connected topology.  In the current release, nodes
behind NAT boxes will connect to all nodes that they can open
connections to, but they cannot open connections to other nodes behind
NAT boxes.  Therefore, the more nodes behind NAT boxes, the less the
topology resembles the intended fully-connected topology.


FILE ENCODING

When a peer stores a file on the grid, it first encrypts the file,
using a key that is optionally derived from the hash of the file
itself.  It then segments the encrypted file into small pieces, in
order to reduce the memory footprint, and to decrease the lag between
initiating a download and receiving the first part of the file; for
example the lag between hitting "play" and a movie actually starting.

The peer then erasure-codes each segment, producing blocks such that
only a subset of them are needed to reconstruct the segment (by
default 3 out of 10 of the blocks).  It sends one block from each
segment to a given server.  The set of blocks on a given server
constitutes a "share".  Only a subset of the shares (3 out of 10) are
needed to reconstruct the file.

A tagged hash of the encryption key is used to form the "storage
index", which is used for both server selection (described below) and
to index shares within the StorageServers on the selected peers.

A variety of hashes are computed while the shares are being produced,
to validate the plaintext, the ciphertext, and the shares
themselves. Merkle hash trees are also produced to enable validation
of individual segments of plaintext or ciphertext without requiring
the download/decoding of the whole file. These hashes go into the
"Capability Extension Block", which will be stored with each share.

The capability contains the encryption key, the hash of the Capability
Extension Block, and any encoding parameters necessary to perform the
eventual decoding process.  For convenience, it also contains the size
of the file being stored.


On the download side, the node that wishes to turn a capability into a
sequence of bytes will obtain the necessary shares from remote nodes, break
them into blocks, use erasure-decoding to turn them into segments of
ciphertext, use the decryption key to convert that into plaintext, then emit
the plaintext bytes to the output target (which could be a file on disk, or
it could be streamed directly to a web browser or media player).

All hashes use SHA-256, and a different tag is used for each purpose.
Netstrings are used where necessary to insure these tags cannot be confused
with the data to be hashed. All encryption uses AES in CTR mode. The erasure
coding is performed with zfec.

A Merkle Hash Tree is used to validate the encoded blocks before they are fed
into the decode process, and a transverse tree is used to validate the shares
as they are retrieved. A third merkle tree is constructed over the plaintext
segments, and a fourth is constructed over the ciphertext segments.  All
necessary hashes are stored with the shares, and the hash tree roots are put
in the Capability Extension Block. The final hash of the extension block goes
into the capability itself.

Note that the number of shares created is fixed at the time the file is
uploaded: it is not possible to create additional shares later. The use of a
top-level hash tree also requires that nodes create all shares at once, even
if they don't intend to upload some of them, otherwise the hashroot cannot be
calculated correctly.


Capabilities

Capabilities to immutable files represent a specific set of bytes.  Think of
it like a hash function: you feed in a bunch of bytes, and you get out a
capability, which is deterministically derived from the input data: changing
even one bit of the input data will result in a completely different
capability.

Read-only capabilities to mutable files represent the ability to get a set of
bytes representing a version of the file.  Each read-only capability is
unique.  In fact, each mutable file has a unique public/private key pair
created when the mutable file is created, and the read-only capability to
that file includes a secure hash of the public key.

Read-write capabilities to mutable files represent the ability to read the
file (just like a read-only capability) and also to write a new version of
the file, overwriting any extant version.  Read-write capabilities are unique
-- each one includes the secure hash of the private key associated with that
mutable file.

The capability provides both "location" and "identification": you can use it
to retrieve a set of bytes, and then you can use it to validate ("identify")
that these potential bytes are indeed the ones that you were looking for.

The "grid" layer is insufficient to provide a virtual drive: an actual
filesystem requires human-meaningful names.  Capabilities sit on the
"global+secure" edge of Zooko's Triangle[1]. They are self-authenticating,
meaning that nobody can trick you into using a file that doesn't match the
capability you used to refer to that file.


SERVER SELECTION

When a file is uploaded, the encoded shares are sent to other peers. But to
which ones? The "server selection" algorithm is used to make this choice.

In the current version, the verifierid is used to consistently-permute the
set of all peers (by sorting the peers by HASH(verifierid+peerid)). Each file
gets a different permutation, which (on average) will evenly distribute
shares among the grid and avoid hotspots.

We use this permuted list of peers to ask each peer, in turn, if it will hold
on to share for us, by sending an 'allocate_buckets() query' to each one.
Some will say yes, others (those who are full) will say no: when a peer
refuses our request, we just take that share to the next peer on the list. We
keep going until we run out of shares to place. At the end of the process,
we'll have a table that maps each share number to a peer, and then we can
begin the encode+push phase, using the table to decide where each share
should be sent.

Most of the time, this will result in one share per peer, which gives us
maximum reliability (since it disperses the failures as widely as possible).
If there are fewer useable peers than there are shares, we'll be forced to
loop around, eventually giving multiple shares to a single peer. This reduces
reliability, so it isn't the sort of thing we want to happen all the time,
and either indicates that the default encoding parameters are set incorrectly
(creating more shares than you have peers), or that the grid does not have
enough space (many peers are full). But apart from that, it doesn't hurt. If
we have to loop through the peer list a second time, we accelerate the query
process, by asking each peer to hold multiple shares on the second pass. In
most cases, this means we'll never send more than two queries to any given
peer.

If a peer is unreachable, or has an error, or refuses to accept any of our
shares, we remove them from the permuted list, so we won't query them a
second time for this file. If a peer already has shares for the file we're
uploading (or if someone else is currently sending them shares), we add that
information to the share-to-peer table. This lets us do less work for files
which have been uploaded once before, while making sure we still wind up with
as many shares as we desire.

If we are unable to place every share that we want, but we still managed to
place a quantity known as "shares of happiness", we'll do the upload anyways.
If we cannot place at least this many, the upload is declared a failure.

The current defaults use k=3, shares_of_happiness=7, and N=10, meaning that
we'll try to place 10 shares, we'll be happy if we can place 7, and we need
to get back any 3 to recover the file. This results in a 3.3x expansion
factor. In general, you should set N about equal to the number of peers in
your grid, then set N/k to achieve your desired availability goals.

When downloading a file, the current release just asks all known peers for
any shares they might have, chooses the minimal necessary subset, then starts
downloading and processing those shares. A later release will use the full
algorithm to reduce the number of queries that must be sent out. This
algorithm uses the same consistent-hashing permutation as on upload, but
stops after it has located k shares (instead of all N). This reduces the
number of queries that must be sent before downloading can begin.

The actual number of queries is directly related to the availability of the
peers and the degree of overlap between the peerlist used at upload and at
download. For stable grids, this overlap is very high, and usually the first
k queries will result in shares. The number of queries grows as the stability
decreases. Some limits may be imposed in large grids to avoid querying a
million peers; this provides a tradeoff between the work spent to discover
that a file is unrecoverable and the probability that a retrieval will fail
when it could have succeeded if we had just tried a little bit harder. The
appropriate value of this tradeoff will depend upon the size of the grid, and
will change over time.

Other peer selection algorithms are possible. One earlier version (known as
"tahoe 3") used the permutation to place the peers around a large ring,
distributed shares evenly around the same ring, then walks clockwise from 0
with a basket: each time we encounter a share, put it in the basket, each
time we encounter a peer, give them as many shares from our basket as they'll
accept. This reduced the number of queries (usually to 1) for small grids
(where N is larger than the number of peers), but resulted in extremely
non-uniform share distribution, which significantly hurt reliability
(sometimes the permutation resulted in most of the shares being dumped on a
single peer).

Another algorithm (known as "denver airport"[2]) uses the permuted hash to
decide on an approximate target for each share, then sends lease requests via
Chord routing. The request includes the contact information of the uploading
node, and asks that the node which eventually accepts the lease should
contact the uploader directly. The shares are then transferred over direct
connections rather than through multiple Chord hops. Download uses the same
approach. This allows nodes to avoid maintaining a large number of long-term
connections, at the expense of complexity, latency, and reliability.


SWARMING DOWNLOAD, TRICKLING UPLOAD

Because the shares being downloaded are distributed across a large number of
peers, the download process will pull from many of them at the same time. The
current encoding parameters require 3 shares to be retrieved for each
segment, which means that up to 3 peers will be used simultaneously. For
larger networks, 8-of-22 encoding could be used, meaning 8 peers can be used
simultaneously. This allows the download process to use the sum of the
available peers' upload bandwidths, resulting in downloads that take full
advantage of the common 8x disparity between download and upload bandwith on
modern ADSL lines.

On the other hand, uploads are hampered by the need to upload encoded shares
that are larger than the original data (3.3x larger with the current default
encoding parameters), through the slow end of the asymmetric connection. This
means that on a typical 8x ADSL line, uploading a file will take about 32
times longer than downloading it again later.

Smaller expansion ratios can reduce this upload penalty, at the expense of
reliability. See RELIABILITY, below. A project known as "offloaded uploading"
can eliminate the penalty, if there is a node somewhere else in the network
that is willing to do the work of encoding and upload for you.


VDRIVE and DIRNODES: THE VIRTUAL DRIVE LAYER

The "virtual drive" layer is responsible for mapping human-meaningful
pathnames (directories and filenames) to pieces of data. The actual bytes
inside these files are referenced by capability, but the "vdrive" is where
the directory names, file names, and metadata are kept.

In the current release, the virtual drive is a graph of "dirnodes". Each
dirnode represents a single directory, and thus contains a table of named
children. These children are either other dirnodes or actual files. All
children are referenced by their capability. Each client creates a "private
vdrive" dirnode at startup. The clients also receive access to a "global
vdrive" dirnode from the central introducer/vdrive server, which is shared
between all clients and serves as an easy demonstration of having multiple
writers for a single dirnode.

The dirnode itself has two forms of capability: one is read-write and the
other is read-only. The table of children inside the dirnode has a read-write
and read-only capability for each child. If you have a read-only capability
for a given dirnode, you will not be able to access the read-write capability
of the children.  This results in "transitively read-only" dirnode access.

By having two different capabilities, you can choose which you want to share
with someone else. If you create a new directory and share the read-write
capability for it with a friend, then you will both be able to modify its
contents. If instead you give them the read-only capability, then they will
*not* be able to modify the contents. Any capability that you receive can be
attached to any dirnode that you can modify, so very powerful
shared+published directory structures can be built from these components.

This structure enable individual users to have their own personal space, with
links to spaces that are shared with specific other users, and other spaces
that are globally visible. Eventually the application layer will present
these pieces in a way that allows the sharing of a specific file or the
creation of a "virtual CD" as easily as dragging a folder onto a user icon.

In the current release, these dirnodes are *not* distributed. Instead, each
dirnode lives on a single host, in a file on it's local (physical) disk. In
addition, all dirnodes are on the same host, known as the "Introducer And
VDrive Node". This simplifies implementation and consistency, but obviously
has a drastic effect on reliability: the file data can survive multiple host
failures, but the vdrive that points to that data cannot. Fixing this
situation is a high priority task.


LEASES, REFRESHING, GARBAGE COLLECTION

Shares are uploaded to a storage server, but they do not necessarily stay
there forever. We are anticipating three main share-lifetime management modes
for Tahoe: 1) per-share leases which expire, 2) per-account timers which
expire and cancel all leases for the account, and 3) centralized account
management without expiration timers.

Multiple clients may be interested in a given share, for example if two
clients uploaded the same file, or if two clients are sharing a directory and
both want to make sure the files therein remain available. Consequently, each
share (technically each "bucket", which may contain multiple shares for a
single storage index) has a set of leases, one per client. One way to
visualize this is with a large table, with shares (i.e. buckets, or storage
indices, or files) as the rows, and accounts as columns. Each square of this
table might hold a lease.

Using limited-duration leases reduces the storage consumed by clients who
have (for whatever reason) forgotten about the share they once cared about.
Clients are supposed to explicitly cancel leases for every file that they
remove from their vdrive, and when the last lease is removed on a share, the
storage server deletes that share. However, the storage server might be
offline when the client deletes the file, or the client might experience a
bug or a race condition that results in forgetting about the file. Using
leases that expire unless otherwise renewed ensures that these lost files
will not consume storage space forever. On the other hand, they require
periodic maintenance, which can become prohibitively expensive for large
grids. In addition, clients who go offline for a while are then obligated to
get someone else to keep their files alive for them.


In the first mode, each client holds a limited-duration lease on each share
(typically one month), and clients are obligated to periodically renew these
leases to keep them from expiring (typically once a week). In this mode, the
storage server does not know anything about which client is which: it only
knows about leases.

In the second mode, each server maintains a list of clients and which leases
they hold. This is called the "account list", and each time a client wants to
upload a share or establish a lease, it provides credentials to allow the
server to know which Account it will be using. Rather than putting individual
timers on each lease, the server puts a timer on the Account. When the
account expires, all of the associated leases are cancelled.

In this mode, clients are obligated to renew the Account periodically, but
not the (thousands of) individual share leases. Clients which forget about
files are still incurring a storage cost for those files. An occasional
reconcilliation process (in which the client presents the storage server with
a list of all the files it cares about, and the server removes leases for
anything that isn't on the list) can be used to free this storage, but the
effort involved is large, so reconcilliation must be done very infrequently.

Our plan is to have the clients create their own Accounts, based upon the
possession of a private key. Clients can create as many accounts as they
wish, but they are responsible for their own maintenance. Servers can add up
all the leases for each account and present a report of usage, in bytes per
account. This is intended for friendnet scenarios where it would be nice to
know how much space your friends are consuming on your disk.

In the third mode, the Account objects are centrally managed, and are not
expired by the storage servers. In this mode, the client presents credentials
that are issued by a central authority, such as a signed message which the
storage server can verify. The storage used by this account is not freed
unless and until the central account manager says so.

This mode is more appropriate for a commercial offering, in which use of the
storage servers is contingent upon a monthly fee, or other membership
criteria. Being able to ask the storage usage for each account (or establish
limits on it) helps to enforce whatever kind of membership policy is desired.


Each lease is created with a pair of secrets: the "renew secret" and the
"cancel secret". These are just random-looking strings, derived by hashing
other higher-level secrets, starting with a per-client master secret. Anyone
who knows the secret is allowed to restart the expiration timer, or cancel
the lease altogether. Having these be individual values allows the original
uploading node to delegate these capabilities to others.

In the current release, clients provide lease secrets to the storage server,
and each lease contains an expiration time, but there is no facility to
actually expire leases, nor are there explicit owners (the "ownerid" field of
each lease is always set to zero). In addition, many features have not been
implemented yet: the client should claim leases on files which are added to
the vdrive by linking (as opposed to uploading), and the client should cancel
leases on files which are removed from the vdrive, but neither has been
written yet. This means that shares are not ever deleted in this release.


FILE REPAIRER

Shares may go away because the storage server hosting them has suffered a
failure: either temporary downtime (affecting availability of the file), or a
permanent data loss (affecting the reliability of the file). Hard drives
crash, power supplies explode, coffee spills, and asteroids strike. The goal
of a robust distributed filesystem is to survive these setbacks.

To work against this slow, continually loss of shares, a File Checker is used
to periodically count the number of shares still available for any given
file. A more extensive form of checking known as the File Verifier can
download the ciphertext of the target file and perform integrity checks (using
strong hashes) to make sure the data is stil intact. When the file is found
to have decayed below some threshold, the File Repairer can be used to
regenerate and re-upload the missing shares. These processes are conceptually
distinct (the repairer is only run if the checker/verifier decides it is
necessary), but in practice they will be closely related, and may run in the
same process.

The repairer process does not get the full capability of the file to be
maintained: it merely gets the "repairer capability" subset, which does not
include the decryption key. The File Verifier uses that data to find out
which peers ought to hold shares for this file, and to see if those peers are
still around and willing to provide the data. If the file is not healthy
enough, the File Repairer is invoked to download the ciphertext, regenerate
any missing shares, and upload them to new peers. The goal of the File
Repairer is to finish up with a full set of 100 shares.

There are a number of engineering issues to be resolved here. The bandwidth,
disk IO, and CPU time consumed by the verification/repair process must be
balanced against the robustness that it provides to the grid. The nodes
involved in repair will have very different access patterns than normal
nodes, such that these processes may need to be run on hosts with more memory
or network connectivity than usual. The frequency of repair runs directly
affects the resources consumed. In some cases, verification of multiple files
can be performed at the same time, and repair of files can be delegated off
to other nodes.

The security model we are currently using assumes that peers who claim to
hold a share will actually provide it when asked. (We validate the data they
provide before using it in any way, but if enough peers claim to hold the
data and are wrong, the file will not be repaired, and may decay beyond
recoverability). There are several interesting approaches to mitigate this
threat, ranging from challenges to provide a keyed hash of the allegedly-held
data (using "buddy nodes", in which two peers hold the same block, and check
up on each other), to reputation systems, or even the original Mojo Nation
economic model.


SECURITY

The design goal for this project is that an attacker may be able to deny
service (i.e. prevent you from recovering a file that was uploaded earlier)
but can accomplish none of the following three attacks:

 1) violate confidentiality: the attacker gets to view data to which you have
    not granted them access
 2) violate consistency: the attacker convinces you that the wrong data is
    actually the data you were intending to retrieve
 3) violate mutability: the attacker gets to modify a dirnode (either the
    pathnames or the file contents) to which you have not given them
    mutability rights

Data validity and consistency (the promise that the downloaded data will
match the originally uploaded data) is provided by the hashes embedded in the
capability. Data confidentiality (the promise that the data is only readable
by people with the capability) is provided by the encryption key embedded in
the capability.  Data availability (the hope that data which has been
uploaded in the past will be downloadable in the future) is provided by the
grid, which distributes failures in a way that reduces the correlation
between individual node failure and overall file recovery failure.

Many of these security properties depend upon the usual cryptographic
assumptions: the resistance of AES and RSA to attack, the resistance of
SHA256 to pre-image attacks, and upon the proximity of 2^-128 and 2^-256 to
zero. A break in AES would allow a confidentiality violation, a pre-image
break in SHA256 would allow a consistency violation, and a break in RSA would
allow a mutability violation. The discovery of a collision in SHA256 is
unlikely to allow much, but could conceivably allow a consistency violation
in data that was uploaded by the attacker. If SHA256 is threatened, further
analysis will be warranted.

There is no attempt made to provide anonymity, neither of the origin of a
piece of data nor the identity of the subsequent downloaders. In general,
anyone who already knows the contents of a file will be in a strong position
to determine who else is uploading or downloading it. Also, it is quite easy
for a coalition of more than 1% of the nodes to correlate the set of peers
who are all uploading or downloading the same file, even if the attacker does
not know the contents of the file in question.

Also note that the file size and (when convergence is being used) a keyed
hash of the plaintext are not protected. Many people can determine the size
of the file you are accessing, and if they already know the contents of a
given file, they will be able to determine that you are uploading or
downloading the same one.

A likely enhancement is the ability to use distinct encryption keys for each
file, avoiding the file-correlation attacks at the expense of increased
storage consumption. This is known as "non-convergent" encoding.

The capability-based security model is used throughout this project. dirnode
operations are expressed in terms of distinct read and write capabilities.
Knowing the read-capability of a file is equivalent to the ability to read
the corresponding data. The capability to validate the correctness of a file
is strictly weaker than the read-capability (possession of read-capability
automatically grants you possession of validate-capability, but not vice
versa). These capabilities may be expressly delegated (irrevocably) by simply
transferring the relevant secrets. 

The application layer can provide whatever security/access model is desired,
but we expect the first few to also follow capability discipline: rather than
user accounts with passwords, each user will get a FURL to their private
dirnode, and the presentation layer will give them the ability to break off
pieces of this vdrive for delegation or sharing with others on demand.


RELIABILITY

File encoding and peer selection parameters can be adjusted to achieve
different goals. Each choice results in a number of properties; there are
many tradeoffs.

First, some terms: the erasure-coding algorithm is described as K-out-of-N
(for this release, the default values are K=3 and N=10). Each grid will have
some number of peers; this number will rise and fall over time as peers join,
drop out, come back, and leave forever. Files are of various sizes, some are
popular, others are rare. Peers have various capacities, variable
upload/download bandwidths, and network latency. Most of the mathematical
models that look at peer failure assume some average (and independent)
probability 'P' of a given peer being available: this can be high (servers
tend to be online and available >90% of the time) or low (laptops tend to be
turned on for an hour then disappear for several days). Files are encoded in
segments of a given maximum size, which affects memory usage.

The ratio of N/K is the "expansion factor". Higher expansion factors improve
reliability very quickly (the binomial distribution curve is very sharp), but
consumes much more grid capacity. The absolute value of K affects the
granularity of the binomial curve (1-out-of-2 is much worse than
50-out-of-100), but high values asymptotically approach a constant that
depends upon 'P' (i.e. 500-of-1000 is not much better than 50-of-100).

Likewise, the total number of peers in the network affects the same
granularity: having only one peer means a single point of failure, no matter
how many copies of the file you make. Independent peers (with uncorrelated
failures) are necessary to hit the mathematical ideals: if you have 100 nodes
but they are all in the same office building, then a single power failure
will take out all of them at once. The "Sybil Attack" is where a single
attacker convinces you that they are actually multiple servers, so that you
think you are using a large number of independent peers, but in fact you have
a single point of failure (where the attacker turns off all their machines at
once). Large grids, with lots of truly-independent peers, will enable the
use of lower expansion factors to achieve the same reliability, but increase
overhead because each peer needs to know something about every other, and the
rate at which peers come and go will be higher (requiring network maintenance
traffic). Also, the File Repairer work will increase with larger grids,
although then the job can be distributed out to more peers.

Higher values of N increase overhead: more shares means more Merkle hashes
that must be included with the data, and more peers to contact to retrieve
the shares. Smaller segment sizes reduce memory usage (since each segment
must be held in memory while erasure coding runs) and increases "alacrity"
(since downloading can validate a smaller piece of data faster, delivering it
to the target sooner), but also increase overhead (because more blocks means
more Merkle hashes to validate them).

In general, small private grids should work well, but the participants will
have to decide between storage overhead and reliability. Large stable grids
will be able to reduce the expansion factor down to a bare minimum while
still retaining high reliability, but large unstable grids (where nodes are
coming and going very quickly) may require more repair/verification bandwidth
than actual upload/download traffic.

Tahoe nodes that run a webserver have a page dedicated to provisioning
decisions: this tool may help you evaluate different expansion factors and
view the disk consumption of each. It is also acquiring some sections with
availability/reliability numbers, as well as preliminary cost analysis data.
This tool will continue to evolve as our analysis improves.

------------------------------

[1]: http://en.wikipedia.org/wiki/Zooko%27s_triangle

[2]: all of these names are derived from the location where they were
     concocted, in this case in a car ride from Boulder to DEN. To be
     precise, "tahoe 1" was an unworkable scheme in which everyone who holds
     shares for a given file would form a sort of cabal which kept track of
     all the others, "tahoe 2" is the first-100-peers in the permuted hash,
     and this document descibes "tahoe 3", or perhaps "potrero hill 1".