Filename: 226-bridgedb-database-improvements.txt
Title: "Scalability and Stability Improvements to BridgeDB: Switching to a
        Distributed Database System and RDBMS"
Author: Isis Agora Lovecruft
Created: 12 Oct 2013
Related Proposals: XXX-social-bridge-distribution.txt
Status: Reserve

*  I.     Overview

   BridgeDB is Tor's Bridge Distribution system, which currently has two major
   Bridge Distribution mechanisms: the HTTPS Distributor and an Email
   Distributor. [0]

   BridgeDB is written largely in Twisted Python, and uses Python2's builtin
   sqlite3 as its database backend.  Unfortunately, this backend system is
   already showing strain through increased times for queries, and sqlite's
   memory usage is not up-to-par with modern, more efficient, NoSQL databases.

   In order to better facilitate the implementation of newer, more complex
   Bridge Distribution mechanisms, several improvements should be made to the
   underlying database system of BridgeDB.  Additionally, I propose that a
   clear distinction in terms, as well as a modularisation of the codebase, be
   drawn between the mechanisms for Bridge Distribution versus the backend
   Bridge Database (BridgeDB) storage system.

   This proposal covers the design and implementation of a scalable NoSQL ―
   Document-Based and Key-Value Relational ― database backend for storing data
   on Tor Bridge relays, in an efficient manner that is ammenable to
   interfacing with the Twisted Python asynchronous networking code of current
   and future Bridge Distribution mechanisms.

*  II.   Terminology

   BridgeDistributor := A program which decides when and how to hand out
                        information on a Tor Bridge relay, and to whom.

   BridgeDB := The backend system of databases and object-relational mapping
               servers, which interfaces with the BridgeDistributor in order
               to hand out bridges to clients, and to obtain and process new,
               incoming ``@type bridge-server-descriptors``,
               ``@type bridge-networkstatus`` documents, and
               ``@type bridge-extrainfo`` descriptors. [3]

   BridgeFinder := A client-side program for an Onion Proxy (OP) which handles
                   interfacing with a BridgeDistributor in order to obtain new
                   Bridge relays for a client.  A BridgeFinder also interfaces
                   with a local Tor Controller (such as TorButton or ARM) to
                   handle automatic, transparent Bridge configuration (no more
                   copy+pasting into a torrc) without being given any
                   additional privileges over the Tor process, [1] and relies
                   on the Tor Controller to interface with the user for
                   control input and displaying up-to-date information
                   regarding available Bridges, Pluggable Transport methods,
                   and potentially Invite Tickets and Credits (a cryptographic
                   currency without fiat value which is generated
                   automatically by clients whose Bridges remain largely
                   uncensored, and is used to purchase new Bridges), should a
                   Social Bridge Distributor be implemented. [2]

*  III.   Databases
** III.A. Scalability Requirements

   Databases SHOULD be implemented in a manner which is ammenable to using a
   distributed storage system; this is necessary because many potential
   datatypes required by future BridgeDistributors MUST be stored permanently.
   For example, in the designs for the Social Bridge Distributor, the list of
   hash digests of spent Credits, and the list of hash digests of redeemed
   Invite Tickets MUST be stored forever to prevent either from being replayed
   ― or double-spent ― by a malicious user who wishes to block bridges faster.
   Designing the BridgeDB backend system such that additional nodes may be
   added in the future will allow the system to freely scale in relation to
   the storage requirements of future BridgeDistributors.

   Additionally, requiring that the implementation allow for distributed
   database backends promotes modularisation the components of BridgeDB, such
   that BridgeDistributors can be separated from the backend storage system,
   BridgeDB, as all queries will be issued through a simplified, common API,
   regardless of the number of nodes system, or the design of future
   BridgeDistributors.

***   1.  Distributed Database System

   A distributed database system SHOULD be used for BridgeDB, in order to
   scale resources as the number of Tor bridge users grows. This database
   system, hereafter referred to as DDBS.

   The DDBS MUST be capable of working within Twisted's asynchronous
   framework. If possible, a Object-Relational Mapper (ORM) SHOULD be used to
   abstract the database backend's structure and query syntax from the Twisted
   Python classes which interact with it, so that the type of database may be
   swapped out for another with less code refactoring.

   The DDBM SHALL be used for persistent storage of complex data structures
   such as the bridges, which MAY include additional information from both the
   `@type bridge-server-descriptor`s and the `@type bridge-extra-info`
   descriptors. [3]

**** 1.a. Choice of DDBS

   CouchDB is chosen for its simple HTTP API, ease of use, speed, and official
   support for Twisted Python applications. [4] Additionally, its
   document-based data model is very similar to the current archetecture of
   tor's Directory Server/Mirror system, in that an HTTP API is used to
   retrieve data stored within virtual directories.  Internally, it uses JSON
   to store data and JavaScript as its query language, both of which are
   likely friendlier to various other components of the Tor Metrics
   infrastructure which sanitise and analyse portions of the Bridge
   descriptors.  At the very least, friendlier than hardcoding raw SQL queries
   as Python strings.

**** 1.b. Data Structures which should be stored in a DDBS:

   - RedactedDB - The Database of Blocked Bridges

     The RedactedDB will hold entries of bridges which have been discovered to
     be unreachable from BridgeDB network vantage point, or have been reported
     unreachable by clients.

   - BridgeDB - The Database of Bridges

     BridgeDB holds information on available Bridges, obtained via bridge
     descriptors and networkstatus documents from the BridgeAuthority. Because
     a Bridge may have multiple `ORPort`s and multiple
     `ServerTransportListenAddress`es, attaching additional data to each of
     these addresses which MAY include the following information on a blocking
     event:
         - Geolocational country code of the reported blocking event
         - Timestamp for when the blocking event was first reported
         - The method used for discovery of the block
         - an the believed mechanism which is causing the block
     would quickly become unwieldy, the RedactedDB and BridgeDB SHOULD be kept
     separate.

   - User Credentials

     For the Social BridgeDistributor, these are rather complex,
     increasingly-growing, concatenations (or tuples) of several datatypes,
     including Non-Interactive Proofs-of-Knowledge (NIPK) of Commitments to
     k-TAA Blind Signatures, and NIPK of Commitments to a User's current
     number of Credits and timestamps of requests for Invite Tickets.

***   2.  Key-Value Relational Database Mapping Server

    For simpler data structures which must be persistently stored, such as the
    list of hashes of previously seen Invite Tickets, or the list of
    previously spent Tokens, a Relational Database Mapping Server (RDBMS)
    SHALL be used for optimisation of queries.

    Redis and Memcached are two examples of RDBMS which are well tested and
    are known to work well with Twisted. The major difference between the two
    is that Memcaches are stored only within volatile memory, while Redis
    additionally supports commands for transferring objects into persistent,
    on-disk storage. 

    There are several support modules for interfacing with both Memcached and
    Redis from Twisted Python, see Twisted's MemCacheProtocol class [5] [6] or
    txyam [7] for Memcached, and txredis [8] or txredisapi [9] for
    Redis. Additionally, numerous big name projects both use Redis as part of
    their backend systems, and also provide helpful documentation on their own
    experience of the process of switching over to the new systems. [17] For
    non-Twisted Python Redis APIs, there is redis-py, which provides a
    connection pool that could likely be interfaced with from Twisted Python
    without too much difficultly. [10] [11]

**** 2.a. Data Structures which should be stored in a RDBMS

    Simple, mostly-flat datatypes, and data which must be frequently indexed
    should be stored in a RDBMS, such as large lists of hashes, or arbitrary
    strings with assigned point-values (i.e. the "Uniform Mapping" for the
    current HTTPS BridgeDistributor).

    For the Social BridgeDistributor, hash digests of the following datatypes
    SHOULD be stored in the RDBMS, in order to prevent double-spending and
    replay attacks:

      - Invite Tickets

        These are anonymous, unlinkable, unforgeable, and verifiable tokens
        which are occasionally handed out to well-behaved Users by the Social
        BridgeDistributor to permit new Users to be invited into the system.
        When they are redeemed, the Social BridgeDistributor MUST store a hash
        digest of their contents to prevent replayed Invite Tickets.

      - Spent Credits

        These are Credits which have already been redeemed for new Bridges.
        The Social BridgeDistributor MUST also store a hash digest of Spent
        Credits to prevent double-spending.

***   3.  Bloom Filters and Other Database Optimisations

    In order to further decrease the need for lookups in the backend
    databases, Bloom Filters can used to eliminate extraneous
    queries. However, this optimization would only be beneficial for string
    lookups, i.e. querying for a User's Credential, and SHOULD NOT be used for
    queries within any of the hash lists, i.e. the list of hashes of
    previously seen Invite Tickets. [14]

****  3.a. Bloom Filters within Redis

    It might be possible to use Redis' GETBIT and SETBIT commands to store a
    Bloom Filter within a Redis cache system; [15] doing so would offload the
    severe memory requirements of loading the Bloom Filter into memory in
    Python when inserting new entries, reducing the time complexity from some
    polynomial time complexity that is proportional to the integral of the
    number of bridge users over the rate of change of bridge users over time,
    to a time complexity of order O(1).

****  3.b. Expiration of Stale Data

    Some types of data SHOULD be safe to expire, such as User Credentials
    which have not been updated within a certain timeframe. This idea should
    be further explored to assess the safety and potential drawbacks to
    removing old data.

    If there is data which SHOULD expire, the PEXPIREAT command provided by
    Redis for the key datatype would allow the RDBMS itself to handle cleanup
    of stale data automatically. [16]

****  4.   Other potential uses of the improved Bridge database system

    Redis provides mechanisms for evaluations to be made on data by calling
    the sha1 for a serverside Lua script. [15] While not required in the
    slightest, it is a rather neat feature, as it would allow Tor's Metrics
    infrastructure to offload some of the computational overhead of gathering
    data on Bridge usage to BridgeDB (as well as diminish the security
    implications of storing Bridge descriptors).

    Also, if Twisted's IProducer and IConsumer interfaces do not provide
    needed interface functionality, or it is desired that other components of
    the Tor software ecosystem be capable of scheduling jobs for BridgeDB,
    there are well-tested mechanisms for using Redis as a message
    queue/scheduling system. [16]

*  References

[0]: https://bridges.torproject.org
     mailto:bridges@bridges.torproject.org
[1]: See proposals 199-bridgefinder-integration.txt at
     https://gitweb.torproject.org/torspec.git/blob/HEAD:/proposals/199-bridgefinder-integration.txt
[2]: See XXX-social-bridge-distribution.txt at
     https://gitweb.torproject.org/user/isis/bridgedb.git/blob/refs/heads/feature/7520-social-dist-design:/doc/proposals/XXX-bridgedb-social-distribution.txt
[3]: https://metrics.torproject.org/formats.html#descriptortypes
[4]: https://github.com/couchbase/couchbase-python-client#twisted-api
[5]: https://twistedmatrix.com/documents/current/api/twisted.protocols.memcache.MemCacheProtocol.html
[6]: http://stackoverflow.com/a/5162203
[7]: http://findingscience.com/twisted/python/memcache/2012/06/09/txyam:-yet-another-memcached-twisted-client.html
[8]: https://pypi.python.org/pypi/txredis
[9]: https://github.com/fiorix/txredisapi
[10]: https://github.com/andymccurdy/redis-py/
[11]: http://degizmo.com/2010/03/22/getting-started-redis-and-python/
[12]: http://www.dr-josiah.com/2012/03/why-we-didnt-use-bloom-filter.html
[13]: http://redis.io/topics/data-types §"Strings"
[14]: http://redis.io/commands/pexpireat
[15]: http://redis.io/commands/evalsha
[16]: http://www.restmq.com/
[17]: https://www.mediawiki.org/wiki/Redis