Name Last Update
api Loading commit data...
client Loading commit data...
ext/rsf Loading commit data...
mapreduce Loading commit data...
partition Loading commit data...
server Loading commit data...
services Loading commit data...
testdata Loading commit data...
ui Loading commit data...
util Loading commit data...
vendor Loading commit data...
.gitignore Loading commit data...
COPYING Loading commit data...
Makefile Loading commit data...
README.rst Loading commit data...
VERSION Loading commit data...
djrandom-cron Loading commit data...
djrandom.init Loading commit data...
mkclientdist.sh Loading commit data...
mkversion.sh Loading commit data...
test.sh Loading commit data...

DJRandom

DJRandom is an online music library application with a web interface, designed for small groups of people who want to share their media collections. On one side, it's a half-joking educational experiment in designing simple distributed systems (which is why it won't support millions of users), but it has also proven to be capable of scaling to a few machines with data in the Terabytes range.

Design

The application is designed do be modular, split into a few key components using minimal interfaces:

  • the database, which offers a plain key/value interface, with range-based scanning (the current implementation is single-hosted but very simple);
  • the index, for full-text search on song metadata;
  • the storage layer (a "blob store"), which holds the song data, again modeled as key/value pairs;
  • a queuing system for background processing.

Each of these components can be replaced by clients to other, better, systems. Many of such systems exists and are open source, and the interfaces can adapt to the specific semantics a large number of them.

But the default djrandom implementation is self-contained and does not depend on any external services. See the file services/README.rst for further details on the various service implementations.

All background jobs are run as mapreduces: again the current implementation is very simple and does not run workers on multiple machines.

Building

Base dependencies:

  • Go (at least version 1.2)
  • LevelDB

Note that the LevelDB version in Debian wheezy is too old with respect to the Go bindings, so you might need to compile it from source.

You'll need to check out the source repository at the right place in your GOPATH:

$ go get -d git.autistici.org/ale/djrandom

Once this is done, choose whether you want to build the full suite or just the clients (which have a much smaller set of dependencies).

Building the clients

The DJRandom client suite need the PortAudio library (version 1.9 or higher) to talk to the audio device, and either ffmpeg or the libav suite (avconv).

On a Debian-based system, apt-get install libav-tools portaudio19-dev should do.

  • ensure that the GOPATH environment variable is set. For instance, the following command will use a directory called go under your $HOME:

    $ export GOPATH=$HOME/go

  • install the client binaries:

    $ cd $GOPATH/src/git.autistici.org/ale/djrandom
    $ make
    $ go install ./client/...
    

You should now have djplay, djupload and djmpd in $GOPATH/bin.

Building the service

Server-side components have quite a few dependencies on third-party libraries, including one that is not packaged in distributions and has to be built manually (imms):

  • a basic C/C++ build environment
  • GNU autotools
  • GNU libtool
  • FFTW (v3)
  • PCRE
  • IMMS (see below)

On a Debian system, install the required packages with:

$ apt-get install build-essential autoconf automake libtool \
      pkg-config libfftw3-dev libpcre3-dev

On OSX, using brew:

$ brew install portaudio leveldb libtool fftw pcre

To build IMMS:

$ go get -d git.autistici.org/ale/imms
$ cd $GOPATH/src/git.autistici.org/ale/imms
$ aclocal ; libtoolize ; automake --foreign --add-missing ; autoconf
$ ./configure && make && sudo make install

(on OSX, you might have to use glibtoolize instead of libtoolize).

You should now be able to build all the server-side tools with:

$ cd ../djrandom
$ make
$ go install ./server/...
$ go install ./mapreduce/...

Running the upload client

Create the ~/.djrandom.conf configuration file with the location of your media directory and your authentication credentials (an API key and a secret), in JSON format:

{
  "music_dir": "/home/user/Music",
  "auth_key": "abcdefghj....",
  "auth_secret": "blahblah...."
}

Then find a way to start djuploader in the background on every login, or whenever your machine starts. It will periodically wake up, check the music_dir, and upload whatever it finds.

It is possible to limit the bandwidth that the uploader is going to use by setting the bw_limit flag to a value in KBytes/sec. For other options just check djuploader --help.

Running the search client

The djplay search client will perform a search on the server and either print out the results in playlist format, or directly attempt to play them to the audio device. It uses the same configuration file as above.

It is best used together with a real audio player of some sort, for instance:

$ djplay --playlist Frank Zappa | vlc

Running the service

A normal deployment usually consists of more than one node (separate machine). Each service runs as a standalone process. The current implementation has two different sets of processes:

  • db_server and task_server must run on a single node (they are in fact not distributed services);
  • index_server, storage_server, djproc and djfe must run on every node.

Generate two partition tables (one for storage, one for the index) with a sufficiently large number of partitions, to ensure an approximately uniform distribution of data, and start the servers using the djrandom.init script.

If you modify the partition tables, for instance to add a new node, you should restart all the processes. Some data will be temporarily unavailable while the nodes automatically rebalance in the background.