Identity/AttachedServices/StorageServerProtocol

Summary

This is a working proposal for the PiCL Storage API, to implement the concepts described in Identity/CryptoIdeas/04-Delta-Sync. It's a work in progress that will eventually obsolete Identity/AttachedServices/StorageProtocolZero.

Delta-Sync Data Model

The storage server hosts a number of independent named collections for each user. Each collection is a key-value store whose contents can be atomically modified by the client. Each modification of a collection creates a new version with corresponding version identifier, which is a signed hash of the contents of the collection at that version.

More details at Identity/CryptoIdeas/04-Delta-Sync.

Authentication

To access the storage service, a client device must authenticate by providing a BrowserID assertion and a Device ID. It will receive in exchange:

• the current version number of each collection
• a short-lived id/key pair that can be used to authenticate subsequent requests with Hawk
• a URL to which further requests should be directed

You can think of this as establishing a "login session" with the server, although we're also tunneling some basic metadata in order to reduce the number of round-trips.

Example:

   >  POST <server-url>
   >  {
   >   "assertion": <browserid assertion>,
   >   "device": <device UUID>
   >  }
   .
   <  200 OK
   <  Content-Type: application/json
   <  {
   <   "base_url": <user-specific access url>,
   <   "id": <hawk auth id>,
   <   "key": <hawk auth secret key>,
   <   "collections": {
   <     "bookmarks": <version id for bookmarks collection>,
   <     "passwords": <version id for passwords collection>,
   <     <...etc...>
   <   }
   <  }

The user and device identity information is encoded in the hawk auth id, to avoid re-sending it on each request. The server may also include additional state in this value, depending on the implementation. It's opaque to the client.

The base_url may include a unique identifier for the user, in order to improve RESTful-icity of the API. Or it might point the client to a specific data-center which houses their write master. It's opaque to the client.

Data Access

The client now makes Hawk-authenticated requests to the storage API under its assigned base_url. The following operations are available.

GET <base-url>

Get the current version id for all collections. This is the same data as returned in the session-establishment call above, but it may be useful if the client wants to refresh its view. Example:

   >  GET <base-url>
   >  Authorization:  <hawk auth parameters>
   .
   <  200 OK
   <  Content-Type: application/json
   <  {
   <   "collections": {
   <     "bookmarks": <version id for bookmarks collection>,
   <     "passwords": <version id for passwords collection>,
   <     <...etc...>
   <   }
   <  }

GET <base-url>/<collection>

Get the current version id for a specific collection. Example:

   >  GET <base-url>/<collection>
   >  Authorization:  <hawk auth parameters>
   .
   <  200 OK
   <  Content-Type: application/json
   <  {
   <   "version": <version id for this collection>
   <  }

GET <base-url>/<collection>/<version>

Get the contents of a specific version of a specific collection. In the simplest case, we GET the full contents like so:

   >  GET <base-url>/<collection>/<version>
   >  Authorization:  <hawk auth parameters>
   .
   <  200 OK
   <  Content-Type: application/json
   <  {
   <   "items": {
   <    "key1": "value1",
   <    "key2": "value2",
   <    <..etc..>
   <   }
   <  }

However, clients will usually want to request a delta from a previous version. They can do this by specifying the "from" parameter. New or updated keys are represented with their value, while deleted keys are represented with a value of null. Like so:

   >  GET <base-url>/<collection>/<version>?from=<previous version>
   >  Authorization:  <hawk auth parameters>
   .
   <  200 OK
   <  Content-Type: application/json
   <  {
   <   "items": {
   <     "key1": "value1",  // a key that was updated
   <     "key2": null       // a key that was deleted
   <   }
   <  } 

To allow reliable transfer of a large number of items, both client and server may choose to paginate responses to this query.

The client may specify "first" as the key at which to (lexicographically) start the listing, and "upto" as the key at which to stop the listing. It may also specify an integer "limit" to restrict the total number of keys sent at once. The server may enforce a default value and/or upper-bound on "limit".

If the set of items is truncated, the server will send the response argument "next" to give the next available key in iteration order. The client should make another request setting "first" equal to the provided value of "next" in order to fetch additional items.

As an example, suppose that the client requests at most two items per response, and the collection contains items "key1", "key2" and "key3". It would would need to fetch them in two batches like so:

   >  GET <base-url>/<collection>/<version>?limit=2
   >  Authorization:  <hawk auth parameters>
   .
   <  200 OK
   <  Content-Type: application/json
   <  {
   <   "next": "key3",
   <   "items": {
   <     "key1": "value1",
   <     "key2": "value2"
   <   }
   <  }
   .
   .
   >  GET <base-url>/<collection>/<version>?first=key3&limit=2
   >  Authorization:  <hawk auth parameters>
   .
   <  200 OK
   <  Content-Type: application/json
   <  {
   <   "items": {
   <     "key3": "value3"
   <   }
   <  }

XXX TODO: There are several error cases that need to be distinguished, possibly by HTTP status code or possibly by some information in the error response body:

• The requested version is not known or no longer present on the server
• We can't generate a delta from the specified "from" version to the request version
• The specified "from" version is invalid (e.g. due to lost writes during a rollback)

POST <base-url>/<collection>/<version>

Creates a new version of a specific collection. In the simplest case, we POST up the full contents of the new version like so:

   >  POST <base-url>/<collection>/<version>
   >  Authorization:  <hawk auth parameters>
   >  {
   >   "items": {
   >     "key1": "value1",
   >     "key2": "value2",
   >     <..etc..>
   >   }
   >  } 
   .
   <  201 Created

However, clients will usually want to send just the changes from a previous version. They can do this by specifying the "from" parameter. New or updated keys are represented with their value, while deleted keys are represented with a value of null. Like so:

   >  POST <base-url>/<collection>/<version>?from=<previous version>
   >  Authorization:  <hawk auth parameters>
   >  {
   >   "items": {
   >     "key1": "value1",  // a key to be updated
   >     "key2": null       // a key to be deleted
   >   }
   >  } 
   .
   <  201 Created

To guard against intermittent or unreliable connections, the client can also send data in batches. It can specify the argument "first" to indicate a key offset at which this batch begins, and the argument "upto" to specify a key offset at which this batch ends. The server will spool all the incoming items until it sees a batch with no "upto" argument, then create the new version as an atomic unit.

As an example, here is how the client might create a new version by sending items one at a time:

   >  POST <base-url>/<collection>/<version>?upto=key2
   >  Authorization:  <hawk auth parameters>
   >  {
   >   "items": {
   >    "key1": value1"
   >   }
   >  }
   .
   <  202 Accepted
   .
   .
   >  POST <base-url>/<collection>/<version>?start=key2&upto=key3
   >  Authorization:  <hawk auth parameters>
   >  {
   >   "items": {
   >    "key2": "value2"
   >   }
   >  }
   .
   <  202 Accepted
   .
   .
   >  POST <base-url>/<collection>/<version>?start=key3
   >  Authorization:  <hawk auth parameters>
   >  {
   >   "items": {
   >    "key3": "value3"
   >   }
   >  }
   .
   <  201 Created

XXX TODO: There are several error cases that need to be distinguished, possibly by HTTP status code or possibly by some information in the error response body:

• There was a conflicting write, so you can no longer create the requested version
• The requested version is invalid, e.g. wrong sequence number
• The specified "from" version is too old, so we can't use it as the start point of a delta
• The specified "from" version is invalid (e.g. due to lost writes during a rollback)
• The provided batches had holes, or were otherwise invalid
• The server forgot a previous batch and you'll have to start again

Things To Think About

• How do people feel about the separate "login" step. It's providing value to the server since it lets us tunnel some state information, but maybe it's not very nice from the client side?
• Currently there's no explicit way for the server to track the current version held by each client. We could add this in the initial handshake, or intuit it based on their activity.
• Is json the best format for this transfer, or could we come up with a more efficient representation?
• Should we add a way to retrieve specific keys, for real-time updating of just the important bits?

feedback from warner:

 <warner> rfkelly: some random thoughts
 <rfkelly> please :-)
 <warner> there will be "shared collections" and "per-device collections", might be useful to have some metadata indicating which is which
 <warner> something to indicate whether data is stored as class-A or class-B, although we've talked (without conclusion) on how to prevent the storage server from getting to make a downgrade attack
 <warner> might be good to store a key ID with each collection, so clients can discover when a key has been changed (and thus they shouldn't be surprised to get MAC failures when they try to decrypt the records)
 <warner> garbage-collection when the password (and thus kB) is reset, pretty tricky
 <rfkelly> could the keyID also double as the classA/classB indicator?
 <warner> GET base/collection/version?limit= needs a response code to indicate "we're done" versus "more is coming"
 <warner> yeah, probably
 <warner> keyID probably = hash(key)
 <rfkelly> right
 <warner> although, if that, (encKey,hmacKey,keyID) = HKDF(key) would be better
 <rfkelly> is "garbage collection" essentially "delete everything that was created with the old key"
 <rfkelly> ?
 <warner> POSTing batches: first= and upto= sounds good, using "upto not in args" requires that we can always detect a missing message, which might not be the case if we memcache the inbound batch (or if we write it to SQL but then SQL rolls back). Might be worth thinking about that part more than I did in my docs.
 <rfkelly> GET base/collection/version?limit= currently indicates doneness by presence/absence of the "next" key in the body; a response code would be better
 <warner> yeah, GC is that, although we probably need some care to make sure an out-of-date client doesn't manage to delete everything, or get into a delete-fight with a less-out-of-date client
 <warner> (might require seqnums in the keyids)
 <warner> ah, next= is fine, unless REST prefers a response code
 * warner gets down to Things To Think About
 <warner> I think the login step is fine, you probably don't want to be doing pubkey verification with every message
 <warner> it adds one RTT (plus sign, plus verify) per hour, or per whatever lifetime we use on the certs (maybe 12 hours?), which seems pretty reasonable
 <warner> but removes the verify time on every single server message
 <warner> ok, time to chat with chris about native-data stuff
 <warner> rfkelly: looks good overall, I think your list of outstanding questions matches my own

And more:

 <warner> rfkelly: hm, so it might be useful to put the "which keyids do I have data for" list in the verify-signature/issue-token handshake, and then if it changes later, revoke that token, so they must do a new handshake
 <warner> rather than defining error responses for what happens when the data is moved from one class to another (or the class-B data is flushed) in between handshakes
 <rfkelly> interesting
 <rfkelly> basically tie your session to a set of metadata, and if the metadata changes you automatically get your session invaldiated
 <warner> yeah
 <rfkelly> the keyids thing, would it be a distinct keyid per collection, or some additional top-level metadata?
 * warner loves to eliminate error pathways
 <rfkelly> warner++
 <warner> probably one keyid per collection
 <warner> something like, "if you can see this account, you can get/set data for the following collections:.." and "to get the plaintext for collection X, you'll need keyid Y"
 <warner> hm
 <warner> well, the main hope is to not confuse clients who try to use the wrong key
 <warner> basically the only time a client should ever see an HMAC failure is when the server manages to corrupt some data
 <rfkelly> right
 <warner> or if the server is being intentionally malicious
 <warner> so there must be some earlier mechanism to indicate A-vs-B-vs-new-B
 <rfkelly> so I'm thinking of doing a bit more explicit "collection metadata" API; currently the only piece of metadata is hte version number, but now it might be (version, keyid, ...other stuff...?)
 <rfkelly> and let clients explicitly get/set/delete this blog to manage the collection state
 <warner> hm, yeah
 <warner> one thing I think we talked about a while ago was collection discovery
 <rfkelly> warner: I like the idea of using an opaque keyid to distinguish classA/classB, because it could prevent the server from learning what class the data is in
 <warner> yeah
 <rfkelly> the client just tries each key in turn until it finds the one that matches the keyid (like truecrypt does to discover the encryption parameters, IIRC)
 <warner> so, looking at this from the inside of the browser..
 <warner> some component or some plugin tells the PICL client "hey, I have data to sync. My data category is named "bookmarks" and this is a "one shared collection" kind of thing"
 <warner> vs one-collection-per-device
 <warner> also it says "this data is going into class-A" or B, probably according to what the user prefs asked for
 <warner> the data-category is unique to this component/plugin (maybe it's a domain name or URL, or GUID)
 <warner> then the PICL client derives some keys, and computes collection-id = hash(kA, category-name), or maybe hash(kB, category-name), for one-shared-collection types
 <warner> or hash(kA/kB, category-name, device-id) for one-collection-per-device
 <warner> so the server can't actually learn what category-name is, or device-id for that matter
 <warner> and then any one-collection-per-device category also needs a device-id-discovery mechanism
 <rfkelly> can it piggyback that from the devices list in the keyserver/idp/thingo
 <rfkelly> ?
 <warner> something like enc(key=hash(kA,category-name), data=device-id), and the server holds a set of the results
 <warner> hm
 <warner> yeah, that's probably better
 <rfkelly> ISTM that "these are my peer devices" is a higher-level concern than at this storage layer
 <warner> although when you add a device, the existing devices need to learn about it
 <rfkelly> it's no specific to a particular collection or datatype
 * warner nods

But [rfkelly] wonders, if the collection name is derived from a hash of its metadata, whether we need to include an explicit "keyid" at all on the server side. Change the key? Change the name of the collection. Doesn't make garbage-collection any easier though...

need last-written and last-read timestamps, to enable garbage collection in some to-be-defined clever system