In last few weeks, I’ve implemented a minimal, barely compliant1 DICT server called ExTra (also stylized ex.tra). The server implements the protocol as described in the linked specification, as well as reading from existing text-databases used by other servers (dictd and GNU dico). I’ve done this as an exercise for learning Elixir, mainly, and it lacks many features in comparison with the two other existing implementations.
Spec summary
DICT is a simple protocol for looking up dictionaries over
TCP/IP. It include a handful (well, more than
handful, if you include the optional authentication) commands like MATCH
or
DEFINE
to retrieve entries from a dictionary database. The database format
is not defined, so technically, one can make it to work with, say, an SQL
database, but it’s customary to use the format the reference implementation
(dictd) uses.
I largely built this based on Elixir’s guide for building a key-value server
Architecture
This diagram shows a very rough and simplified architecture overview of ExTra. Yea, I know, it looks ugly, but organizing a diagram is hard, so I’ll describe it in details below.
The processes2 are supervised by the application supervisor, and are respawned once they crash. There are three main processes concerned here:
ExTra.Server
: This one accepts connections from the dedicated portTask.Supervisor
: This one is a supervisor that will spawn new processes that will bind to clients until disconnected.ExTra.Dict
is a GenServer that executes commands sent to the server, and reads from dictionaries to generate response. Note that GenServer is not a TCP/IP server.
TCP server
The TCP server is implemented using erlang’s :gen_tcp
:
{:ok, client} = :gen_tcp.accept(socket)
Here the client
is a connection to client. We send and receive data via this
process. To read from this until disconnection, we would put it on an infinite
loop (acceptor
in the diagram). However, that also means the server is
locked to that client and cannot accept another—you probably have learned
this from a network class implementing an echo server in C. This is where the
task supervisor comes in: instead of running directly into the infinite loop,
we spawn a Task
that does it:
Task.Supervisor.start_child(ExTra.ConnSupervisor, fn -> serve_first(client, host) end)
In the loop, commands are parsed and run, something like
ExTra.Command.parse(data)
and ExTra.Command.run(commands)
The ExTra.Command
module then sends these commands to ExTra.Dict
GenServer
to execute it, something like ExTra.Dict.command(ExTra.Dict, command)
.
GenServer
The commands are sent to the GenServer
and handled by some other modules:
@impl true
def handle_call({:define, dictionary, word}, _from, state) do
{:reply, ExTra.Dict.Define.define(dictionary, word), state}
end
def handle_call({:match, dictionary, strategy, word}, _from, state) do
{:reply, ExTra.Dict.Match.match(dictionary, strategy, word), state}
end
This level of abstraction may seems a bit convoluted, but using GenServer here would allow for caching matches and definitions, and separating matches and definitions to a separate modules allow for different search modules depending on config. Not a necessary thing, just for educational purpose.
Matching definitions
The .dict
file stores entries as well as metadata as plain text, while the
.index
file store positions of the entries as:
<entry> <start> <length>
Where <start>
and <length>
are in quartosexagesimal or base 64. Numeral
base 64, not base 64 encoding that is implemented in the standard library.
After a few shortening, the conversion is done in less than 20 lines:
def base64num(num) do
alphabet =
'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/'
|> Stream.with_index()
|> Enum.into(%{})
len = String.length(num)
num
|> String.to_charlist()
|> Stream.with_index()
|> Stream.map(fn {c, i} -> {alphabet[c], len - i - 1} end)
# Left-shift 6 * power bits is equal to multiply by (2^6)^power, but faster
|> Stream.map(fn {digit, power} -> Bitwise.bsl(digit, power * 6) end)
|> Enum.sum()
end
Firstly, I mapped the digits in the alphabet to its respective values, which are also their indices in the char list. The digits in the input string are then mapped to their values based on this map, while their indices are mapped to the power. Finally, these values are powered and summed as the answer.
Upon reading these values, the definition from the dict
files can be
retrieved as simply as:
# the content that comes before what we need
_ = IO.read(file, start)
# has to `binread` to interpret the UTF-8 encoded characters
IO.binread(file, length)
Further
The full implementation can be found on SourceHut. As this is the first application I’ve written in Elixir, I’m sure there’s a lot of stuff I’ve written here isn’t recommended, so if you have some suggestion for improvement, please send me.
As of writing, there are still several features I’d like to implement that I haven’t, such as:
- Proper response for
STATUS
andOPTION MIME
commands - Implement more matching strategies
- Make uses of the GenServer’s state for caching matches
-
The
OPTION MIME
command is not yet compliant, actually. It currently is a no-op command while it should check00-database-mime-header
in the file and respond an empty line if not present. However, none of the clients (that is, onlydict
anddico
, as far as I know) does anything with that field, so it doesn’t cause any trouble. ↩︎ -
To be understood as BEAM processes and not OS processes ↩︎