Nicknym

Automatic discovery and validation of public keys.

Introduction

What is Nicknym?

Nicknym is a protocol to map user nicknames to public keys. With Nicknym, the user is able to think solely in terms of nickname, while still being able to communicate with a high degree of security (confidentiality, integrity, and authenticity). Essentially, Nicknym is a system for binding human-memorable nicknames to a cryptographic key via automatic discovery and automatic validation.

Nicknym is a federated protocol: a Nicknym address is in the form username@domain just alike an email address and Nicknym includes both a client and a server component. Although the client can fall back to legacy methods of key discovery when needed, domains that run the Nicknym server component enjoy much stronger identity guarentees.

Nicknym is key agnostic, and supports whatever public key information is available for an address (OpenPGP, OTR, X.509, RSA, etc). However, Nicknym enforces a strict one-to-one mapping of address to public key.

Why is Nicknym needed?

Existing forms of secure identity are deeply flawed. These systems rely on either a single trusted entity (e.g. Skype), a vulnerable Certificate Authority system (e.g. S/MIME), or key identifiers that are not human memorable (e.g. fingerprints used in OpenPGP, OTR, etc). When an identity system is hard to use, it is effectively compromised because too few people take the time to use it properly.

The broken nature of existing identities systems (either in security or in usability) is especially troubling because identity remains a bedrock precondition for any message security: you cannot ensure confidentiality or integrity without confirming the authenticity of the other party. Nicknym is a protocol to solve this problem in a way that is backward compatible, easy for the user, and includes very strong authenticity.

Goals

High level goals

  • Pseudo-anonymous and human friendly addresses in the form username@domain.
  • Automatic discovery and validation of public keys associated with an address.
  • The user should be able to use Nicknym without understanding anything about public/private keys or signatures.

Technical goals

  • Wide utility: nicknym should be a general purpose protocol that can be used in wide variety of contexts.
  • No revocation: instead of key revocation, support short lived keys that frequently and automatically refresh.
  • Prevent dangerous actions: Nicknym should fail hard when there is a possibility of an attack.
  • Minimize false positives: because Nicknym fails hard, we should minimize false positives where it fails incorrectly.
  • Resistant to malicious actors: Nicknym should be externally auditable in order to assure service providers are not compromised or advertising bogus keys.
  • Resistant to association analysis: Nicknym should not reveal to any actor or network observer a map of a user’s associations.

Non-goals

  • Nicknym does not try to create a decentralized peer-to-peer identity system. Nicknym is federated, akin to the way email is federated.

The binding problem

Nicknym attempts to solve the problem of binding a human memorable identifier to a cryptographic key. If you have the identifier, you should be able to get the key with a high level of confidence, and vice versa. The goal is to have federated, human memorable, globally unique public keys.

There are a number of established methods for binding identifier to key:

The methods differ widely, but they all try to solve the same general problem of proving that a person or organization is in control of a particular key.

Nicknym overview

Nicknym solves the binding problem by using a combination of methods, utilizing TOFU, X.509, Network Perspective, and additional methods we call “Provider Keys” and “Federated Web of Trust” (FWOT).

  1. Nicknym starts with TOFU of user keys, because it is easy to do and backward compatible with legacy providers. In TOFU, your client naively accept the key of another user when it first encounters it. When you accept a key via TOFU, you are making a bet that possible attackers against you did not have the foresight to specifically target you with a false key during discovery.
  2. Next, we add X.509. For those providers that publish the public keys of their users, we require that these keys be fetched over validated TLS. This makes third party attacks against TOFU more difficult, but also places a lot of trust in the providers (and the Certificate Authorities).
  3. Next, we add a simple form of Network Perspective where the client can ask one provider what key another provider is distributing. This allows a user’s client to be able to audit their provider and keep them honest in an automated manner. If a service provider distributes bogus keys, their users and other providers will be quickly alerted to the problem.
  4. Next, we add Provider Keys. If a service provider supports nicknym, the public keys of its users are additionally signed by a “provider key”. If your client has the correct provider key, you no longer need to TOFU the keys of the provider’s users. This has the benefit making it possible for a user to issue new keys, and to add support for very short-lived keys rather than trying to use key revocation. A service provider is much less likely to lose their private key or have it compromised, a significant problem with TOFU of user keys.
  5. Finally, we add a Federated Web of Trust. The system works like this: each service provider is responsible for the due diligence of properly signing the keys of a few other providers, akin to the distributed web of trust model of OpenPGP, but with all the hard work of proper signature validation placed upon the service provider. When a user communicates with another party who happens to use a service provider that participates in the FWOT, the user’s software will automatically trace a chain of signature from the other party’s key, to their service provider, to the user’s own service provider (with some possible intermediary signatures). This allows for identity that is verified through an end-to-end trust path from any user to any other user in a way that can be automated and is human memorable. Support for a FWOT allows us to bypass entirely X.509 Certificate Authorities, to gracefully handle short lived provider keys, and to handle emergency re-key events if a provider’s key is lost.

As we move down this list, each measure taken gets more complicated, requires more provider cooperation, and provides less additional benefit than the one before it. Nevertheless, each measure contributes some important benefit toward the goal of automatic binding of user identity to public key.

Questions

Why not use WoT? Most users are empirically unable to properly maintain a web of trust. The concepts are hard, it is easy to mess up the signing practice, most people default to TOFU anyway, and very few users use revocation properly. Most importantly, the WOT exposes a user’s social network, potentially highly sensitive information in its own right. When first proposed, WOT was a clever innovation, but contemporary threats have greatly reduced its usefulness.

Why not use DANE/DNSSEC? DANE is great for discovery and validation of server keys, but there are many reasons why it is not so good for user keys: DNS records are slow to update; DNS queries are observable, unlike HTTP over TLS; it is difficult for a provider to publish thousands of keys in DNS; it is much easier for a client to do a simple HTTP fetch (and more possible for HTML5 clients). Also, RSA Public keys will soon be too big for UDP packets (though this is not true of ECC), so putting keys in DNS will mean putting a URL to a key in DNS, so you might as well just use HTTP anyway.

Why not use Shared Secret? Shared secrets, like with the Socialist Millionaire protocol, are cool in theory but prone to user error and frustration in practice. A typical user is not in a position to have established a prior secret with most of the people they need to make first contact with. Shared secrets also cannot be scaled to a group setting. Finally, shared secrets are often typed incorrectly (e.g. was the secret “Invisible Zebra” or “invisibleZebra”? This could be fixed with rules for secret normalization, but this is tricky and language specific). For the special case of advanced users with special security needs, however, a shared secret provides a much stronger validation than other methods of key binding (so long as the validation window is small).

Why not use Mail-back Verification? If the provider distributes user keys, there is not any benefit to mail-back verification. The nicknym protocol could potentially benefit from a future enhancement to support mail-back for users on a non-cooperating legacy provider. However, at its best, mail-back is a very weak form of key validation.

Why not use Global Append-only Log? Maybe we should, they are neat. However, current implementations are slow, resource intensive, and experimental (e.g. namecoin).

Why not use Nonverbal Feedback? ZRTP can use non-verbal clues to establish secure identity because of the nature of a live phone call. This doesn’t work for text only messaging.

Why not use the key fingerprint as the unique identifier? This is the strategy taken by all systems for peer-to-peer messaging (e.g. retroshare, bitmessage, etc). Depending on the length of the fingerprint, this method is very secure. Essentially, this approach neatly solves the binding problem by collapsing the key and the identifier together as one. The problem, of course, is that this is not very user friendly. Users must either have pre-arranged some way to exchange fingerprints, or they must fall back to one of the other methods for verification (shared secret, WoT, etc). The friction associated with pre-arranged sharing of fingerprints can be reduced with technology, using QR-codes and hand held devices, for example. In the best case scenario, however, fingerprints as identifiers will always be much less user friendly than simple username@domain.org addresses. The motivating premise behind Nicknym is that when an identity system is hard to use, it is effectively compromised because too few people take the time to use it properly.

Related work

WebID and Mozilla Persona

What about WebID or Mozilla Persona? These are both interesting standards for cryptographically proving identify, so why do we need something new?

These protocols, and the poorly conceived OpenID Connect, are designed to address a fundamentally different problem: authenticating a user to a website. The problem of authenticating users to one another requires a different architecture entirely. There are some similarities, however, and in the long run a Nicknym provider could also be a WebID and Mozilla Persona provider.

STEED

STEED is a proposal with very similar goals to Nicknym. In a nutshell, Nicknym basically looks very similar to STEED when the domain owner does not support Nicknym. STEED includes four main ideas:

  • trust upon first contact: Nicknym uses this as well, although this is the fallback mechanism when others fail.
  • automatic key distribution and retrieval: Nicknym uses this as well, although we used HTTP for this instead of DNS.
  • automatic key generation: Nicknym is designed specifically to support automatic key generation, but this is outside the scope of the Nicknym protocol and it is not required.
  • opportunistic encryption: Again, Nicknym is designed to support opportunistic encryption, but does not require it.

Additional differences include:

  • Nicknym is key agnostic: Nicknym does not make an assumption about what types of public keys a user wants to associate with their address.
  • Nicknym is protocol agnostic: Nicknym can be used with SMTP, XMPP, SIP, etc.
  • Nicknym relies on service provider adoption: With Nicknym, the strength of verification of public keys rests the degree to which a service provider adopts Nicknym. If a service provider does not support Nicknym, then effectively Nicknym opperates like STEED for that domain.

DANE

DANE, and the specific proposal for OpenPGP user keys using DANE, offer a standardized method for securely publishing and locating OpenPGP public keys in DNS.

As noted above, DANE will be very cool if ever adopted widely, but user keys are probably not a good fit for DNSSEC, because of issues of observability of DNS queries and complexity on the server and client end.

By relying on the central authority of the root DNS zone, and the authority of TLDs (many of which are of doubtful trustworthiness), DANE potentially suffers from problems of compromised or nefarious authorities. Because DNS queries are not secure, a single user is particularly vulnerable to MiTM attacks that rewrite all their DNS queries. Adopting an alternate DNS query system, like DNSCurve, DNSCrypt, an alternate HTTPS based API, or restricting DNS queries to a VPN, would go a long way to fix this problem, and would effectively turn any supporting DNS server into a network perspectives notary. Regardless, the other problems with using DANE for user keys remain.

Nicknym protocol

Definitions

  • address: A globally unique handle in the form user@domain (i.e. an email, SIP, or XMPP address).
  • provider: A service provider that offers end-user services on a particular domain.
  • user key: A public/private key pair associated with a user address. If not specified, “user key” refers to the public key.
  • provider key: A public/private key pair owned by the provider. The address associated with this key is just the domain of the service provider.
  • validated key: A key is “validated” if the nickagent has bound the user address to a public key.
  • nickagent: Client side program that manages a user’s contact list, the public keys they have encountered and validated, and the user’s own key pairs. The nickagent may also expose an API for other local applications to query for a public key.
  • nickserver: Server side daemon run by providers who support Nicknym. A nickserver is responsible for answering the question “what public key do you see for this address”?

Nickserver requests

A nickagent will attempt to discover the public key for a particular user address by contacting a nickserver. The nickserver returns JSON encoded key information in response to a simple HTTP request with a user’s address. For example:

curl -X POST -d address=alice@domain.org https://nicknym.domain.org:6425
  • The port is always 6425.
  • The HTTP verb may be POST or GET.
  • The request must use TLS (see Query security).
  • The query data should have a single field ‘address’.
  • For POST requests to nicknym.domain.org, the query data may be encrypted to the the public OpenPGP key nicknym@domain.org (see Query security).
  • The request may include an “If-Modified-Since” header. In this case, the response might be “304 Not Modified”.

Requests may be local or foreign, and for user keys or for provider keys.

  • local requests are for information that the nickserver is authoritative. In other words, when the requested address is for the same domain that the nickserver is running on.
  • foreign request are for information about other domains.
  • user key requests are for addresses in the form “username@domain”.
  • provider key requests are for addresses in the form “domain”.

Local, Provider Key request

For example:

https://nicknym.domain.org:6425/?address=domain.org

The response is the authoritative provider key for that domain.

Local, User Key request

For example:

https://nicknym.domain.org:6425/?address=alice@domain.org

The nickserver returns authoritative key information from the provider’s own user database. Every public key returned for local requests must be signed by the provider’s key.

Foreign, Provider Key request

For example:

https://nicknym.domain.org:6425/?address=otherdomain.org
  1. First, check the nickserver’s cache database of discovered keys. If the cache is not old, return this key. This step is skipped if the request is encrypted to the foreign provider’s key.
  2. Otherwise, fetch provider key from the provider’s nickserver, cache the result, and return it.

Foreign, User Key request

For example:

https://nicknym.domain.org:6425/?address=bob@otherdomain.org
  • First, check the nickserver’s database cache of nicknyms. If the cache is not old, return the key information found in the cache. This step is skipped if the request is encrypted to a foreign provider key.
  • Otherwise, attempt to contact a nickserver run by the provider of the requested address. If the nickserver exists, query that nickserver, cache the result, and return it in the response.
  • Otherwise, fall back to querying existing SKS keyservers, cache the result and return it.
  • Otherwise, return a 404 error.

If the key returned for a foreign request contains multiple user addresses, they are all ignored by nicknym except for the user address specified in the request.

Nickserver response

The nickserver will respond with one of the following status codes:

  • “200 Success”: found keys for this address, and the result is in the body of the response encoded as JSON.
  • “304 Not Modified”: if the request included an “If-Modified-Since” header and the keys in question have not been modified, the response will have status code 304.
  • “404 Not Found”: no keys were found for this address.
  • “500 Error”: An unknown error occurred. Details may be included in the body.

Responses with status code 200 include a body text that is a JSON encoded map with a field “address” plus one or more of the following fields: “openpgp”, “otr”, “rsa”, “ecc”, “x509-client”, “x509-server”, “x509-ca”. For example:

{
  "address": "alice@example.org",
  "openpgp": "6VtcDgEKaHF64uk1c/crFhRHuFW9kTvgxAWAK01rXXjrxEa/aMOyXnVQuQINBEof...."
}

Responses with status codes other than 200 include a body text that is a JSON encoded map with the following fields: “address”, “status”, and “message”. For example:

{
  "address": "bob@otherdomain.org",
  "status": 404,
  "message": "Not Found"
}

A nickserver response is always signed with a OpenPGP public signing key associated with the provider. Both successful AND unsuccessful responses are signed. Responses to successful local requests must be signed by the key associated with the address “nicknym@domain.org”. Foreign requests and non-200 responses may alternately be signed with a key associated with the address nickserver@domain.org. This allows for user keys to be signed off-site and in advance, if they so choose. The signature is ASCII armored and appended to the JSON.

{
  "address": "alice@example.org",
  "openpgp": "6VtcDgEKaHF64uk1c/crFhRHuFW9kTvgxAWAK01rXXjrxEa/aMOyXnVQuQINBEof...."
}
-----BEGIN PGP SIGNATURE-----
iQIcBAEBCgAGBQJRhWO+AAoJEIaItIgARAAl2IwP/24z9CjKjD0fd27pQs+r+e3h
p8KAYDbVac3+c3vm30DjHO/RKF4Zq6+sTAIkrFvXOwYJl9KgjMpQVV/voInjxATz
-----END PGP SIGNATURE-----

If the data in the request was encrypted to the public key nicknym@domain.org, then the JSON response and signature are additionally encrypted to the symmetric key found in the request and returned base64 encoded.

TBD: maybe we should just switch to a raw RSA or ECC signature.

Query balancing

A nickagent must choose what IP address to query by selecting randomly from among hosts that resolve from nicknym.domain.org (where domain.org is the domain name of the provider).

If a host does not response, a nickagent must skip over it and attempt to contact another host in the pool.

Query security

TLS is required for all nickserver queries.

When querying https://nicknym.domain.org, nickagent must validate the TLS connection in one of four possible ways:

  1. Using a commercial CA certificate distributed with the host operating system.
  2. Using DANE TLSA record to discover and validate the server certificate.
  3. Using a seeded CA certificate (see Discovering nickservers).
  4. Using a custom self-signed CA certificate discovered for the domain, so long as the CA certificate was discovered via #1 or #2 or #3. Custom CA certificates may be discovered for a domain by making a provider key request of a nickserver (e.g. https://nicknym.known-domain.org/?address=new-domain.org).

Optionally, a nickagent may make an encrypted query like so:

  1. Suppose the nickagent wants to make an encrypted query regarding the address alice@domain-x.org.
  2. Nickagent discovers the public key for nicknym@domain-y.org.
  3. The nickagent makes a POST request to a nickserver with two fields: address and ciphertext.
  4. The address only contains the domain part of the address (unlike an unencrypted request).
  5. The ciphertext field is encrypted to the public key for nicknym@domain-y.org. The corresponding cleartext contains the full address on the first line followed by randomly generated symmetric key on the second line.
  6. If the request was local, the nickserver handles the request. If the request for foreign, the nickserver proxies the request to the domain specified in the address field.
  7. When the request gets to the right nickserver, the body of the nickserver response is encrypted using using the symmetric key. The first line of the response specifies the cipher and mode used (allowed ciphers TBD).

Comment: although it may seem excessive to encrypt both the request via TLS and the request body via OpenPGP, the reason for this is that many requests will not use OpenPGP.

Automatic key validation

A key is “validated” if the nickagent has bound the user address to a public key.

Nicknym supports three different levels of key validation:

  • Level 3 - path trusted: A path of cryptographic signatures can be traced from a trusted key to the key under evaluation. By default, only the provider key from the user’s provider is a “trusted key”.
  • Level 2 - provider signed: The key has been signed by a provider key for the same domain, but the provider key is not validated using a trust path (i.e. it is only registered)
  • Level 1 - registered: The key has been encountered and saved, it has no signatures (that are meaningful to the nickagent).

A nickagent will try to validate using the highest level possible.

Automatic renewal

A validated public key is replaced with a new key when:

  • The new key is path trusted
  • The new key is provider signed, but the old key is only registered.
  • The new key has a later expiration, and the old key is only registered and will expire “soon” (exact time TBD).
  • The agent discovers a new subkey, but the master signing key is unchanged.

In all other cases, the new key is rejected.

The nickagent will attempt to refresh a key by making request to a nickserver of its choice when a key is past ¾ of its lifespan and again when it is about to expire.

Nicknym encourages, but does not require, the use of short lived public keys, in the range of X to Y days. It is recommended that short lived keys are not uploaded to OpenPGP keyservers.

Automatic invalidation

A key is invalidated if:

  • The old key has expired, and no new key can be discovered with equal or greater validation level.

This means validation is a one way street: once a certain level of validation is established for a user address, no client should accept any future keys for that address with a lower level of validation.

Discovering nickservers

It is entirely up to the nickagent to decide what nickservers to query. If it wanted to, a nickagent could send all its requests to a single nickserver.

However, nickagents should discover new nickservers and balance their queries to these nickservers for the purposes of availability, load balancing, network perspective, and hiding the user’s association map.

Whenever the nickagent is asked by a locally running application for a public key corresponding to an address on the domain domain.org, it may check to see if the host nicknym.domain.org exists. If the domain resolves, then the nickagent may add it to the pool of known nickservers. A nickagent should only perform this DNS check if it is able to do so over an encrypted tunnel.

Additionally, a nickagent may be distributed with an initial list of “seed” nickservers. In this case, the nickagent is distributed with a copy of the CA certificate used to validate the TLS connection with each respective seed nickserver.

Cross-provider signatures

Nicknym does not support user signatures on user keys. There is no trust path from user to user. However, a service provider may sign the provider key of another provider.

To be written.

Auditing

In order to keep the user’s provider from handing out bogus public keys, a nickagent should occasionally make foreign queries of the user’s own address against nickservers run by third parties. The recommended frequency of these queries is once per day, at a random time during local waking hours.

In order to prevent a nickserver from handing out bogus provider keys, a nickagent should query multiple nickservers before a provider key is registered or path trusted.

Possible attacks:

Attack 1 - Intercept Outgoing:

  • Attack: provider A signs an impostor key for provider B and distributes it to users of A (in order to intercept outgoing messages sent to B).
  • Countermeasure: By querying multiple nickservers for the provider key of B, the nickagent can detect if provider A is attempting to distribute impostor keys.

Attack 2 - Intercept Incoming:

  • Attack: provider A signs an impostor key for one of its own users, and distributes to users of provider B (in order to intercept incoming messages).
  • Countermeasure: By querying for its own keys, a nickagent can detect if a provider is given out bogus keys for their addresses.

Attack 3 - Association Mapping:

  • Attack: A provider tracks all the requests for key discovery in order to build a map of association.
  • Countermeasure: By performing foreign key queries via third party nickservers, an agent can prevent any particular entity from tracking their queries.

Known vulnerabilities

The nicknym protocol does not yet have a good solution for dealing with the following problems:

  • Enumeration attack: an attacker can enumerate the list of all users for a provider by simply querying every possible username combination. We have no defense against this, although it would surely take a while.
  • DDoS attack: by their very nature, nickservers perform a bit of work for every request. Because of this, they are vulnerable to be overloaded by a a flood of bogus requests.
  • Besmirch attack: a MitM attacker can sully the reputation of a provider by generating many bad responses (responses signed with the wrong key), thus leading other nickservers and nicknym agents to consider the provider compromised.

Future enhancements

Additional discovery mechanisms

In addition to nickservers and SKS keyservers, there are two other potential methods for discovering public keys:

  • Webfinger includes a standard mechanism for distributing a user’s public key via a simple HTTP request. This is very easy to implement on the server, and very easy to consume on the client side, but there are not many webfinger servers with public keys in the wild.
  • DNS is used by multiple competing standards for key discovery. When and if one of these emerges predominate, Nicknym should attempt to use this method when available.

Reference nickagent implementation

https://github.com/leapcode/keymanager

There is a reference nickagent implementation called “key manager” written in Python and integrated into the LEAP client. It uses Soledad to store its data.

Public API

refresh_keys()

updates the keys with fresh ones, as needed.

get_key(address, type)

returns a single public key for address. type is one of ‘openpgp’, ‘otr’, ‘x509’, or ‘rsa’.

send_key(address, public_key, type)

authenticates with the appropriate provider and saves the public_key in the user database.

Storage

Key manager uses Soledad for storage. GPGME, however, requires keys to be stored in keyrings, which are read from disk.

For now, Key Manager deals with this by storing each key in its own keyring. In other words, every key is in a keyring with exactly 1 key, and this keyring is stored in a Soledad document. To keep from confusing this keyring from a normal keyring, I will call it a ‘unitary keyring’.

Suppose Alice needs to communicate with Bob:

  1. Alice’s Key Manager copies to disk her private key and bob’s public key. The key manager gets these from Soledad, in the form of unitary Keyrings.
  2. Client code uses GPGME, feeding it these temporary keyring files.
  3. The keyrings are destroyed.

TBD: how best to ensure destruction of the keyring files.

An example Soledad document for an address:

{
  "address":"alice@example.org",
  "keys": [
    {
      "type": "opengpg"
      "key": "binary blob",
      "keyring": "binary blob",
      "expires_on": "2014-01-01",
      "validation": "provider_signed",
      "first_seen_at": "2013-04-01 00:11:00",
      "last_audited_at": "2013-04-02 12:00:00",
    },
    {
      "type": "otr"
      "key": "binary blob",
      "expires_on": "2014-01-01",
      "validation": "registered",
      "first_seen_at": "2013-04-01 00:11:00",
      "last_audited_at": "2013-04-02 12:00:00",
    }
  ]
}

Pseudocode

get_key

#
# return a key for an address
#
function get_key(address, type)
  if key for address exists in soledad database?
    return key
  else
    fetch key from nickserver
    save it in soledad
    return key
  end
end

send_key

#
# send the user's provider the user's key. this key will get signed by the provider, and replace any prior keys
#
function send_key(type)
  if not authenticated:
    error!
  end
  get (self.address, type)
  send (key_data, type) to the provider
end

refresh_keys

#
# update the user's db of validated keys to see if there are changes.
#
function refresh_keys()
  for each key in the soledad database (that should be checked?):
      newkey = fetch_key_from_nickserver()
      if key is about to expire and newkey complies with the renewal paramters:
          replace key with newkey
      else if fingerprint(key) != fingerprint(newkey):
          freak out, something wrong is happening? :)
          may be handle revokation, or try to get some voting for a given key and save that one (retrieve it through tor/vpn/etc and see what's the most found key or something like that.
      else:
          everything's cool for this key, continue
      end
  end
end

private fetch_key_from_nickserver

function fetch_key_from_nickserver(key)
  randomly pick a subset of the available nickservers we know about
  send a tcp request to each in this subset in parallel
  first one that opens a successful socket is used, all the others are terminated immediately
  make http request
  parse json for the keys
  return keys
end

Reference nickserver implementation

https://github.com/leapcode/nickserver

The reference nickserver is written in Ruby 1.9 and licensed GPLv3. It is lightweight and scalable (supporting high concurrency, and reasonable latency), and uses EventMachine for asynchronous network IO. Data is stored in CouchDB.