Skip to content

doubleyou/dps

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DPS - Distributed Pub/Sub

Multichannel distributed pub/sub messaging, with full duplication across several nodes.

Replication has protection from lag: if neighbour is too slow, replication if suspended.

Usage:

Channel creation

dps:new(<<"channel1">>),
dps:new(<<"channel2">>).

Publishing

dps:publish(<<"channel1">>, <<"Message1">>).

Subscribing for WebSockets

TS = 0,
dps:subscribe(<<"channel1">>, TS),
dps:subscribe(<<"channel2">>, TS),
dps:subscribe(<<"channel3">>, TS),
receive
  {dps_msg, Channel, LastTS, Messages} -> 
    reply({json, [{ts,LastTS},{messages,Messages}]})
end.

Subscribing for long-poll comet

multi_fetch will wait for first message from any requested channel. Can return immediately with many messages, if your TS is too old and some messages have arrived after it.

TS = list_to_integer(proplists:get_value("ts",Query, "0")),
{ok, LastTS, Messages} = dps:multi_fetch([<<"channel1">>,<<"channel2">>], TS, 30000),
reply({json, [{ts,LastTS}, {messages,Messages}]}).

Benchmarking

DPS is beleived to handle about 60 000 msg/sec on several channels with replicating on several servers.

How to benchmark:

maxbp:dps max$ make bench
./rebar compile skip_deps=true
==> dps (compile)
erl -pa ebin -pa deps/*/ebin -smp enable -s dps_benchmark run1 -sname bench@localhost -setcookie cookie
Erlang R15B02 (erts-5.9.2) [source] [64-bit] [smp:4:4] [async-threads:0] [hipe] [kernel-poll:false] [dtrace]

Eshell V5.9.2  (abort with ^G)
(bench@localhost)1> Send: 0 msg/s
Send: 31236 msg/s
Send: 43068 msg/s
Send: 66458 msg/s
Send: 54695 msg/s
Send: 57312 msg/s
Send: 57570 msg/s
Send: 57843 msg/s
Send: 58531 msg/s

You will be able to see how many messages are published inside DPS

Now lets compare with Redis pub/sub w. ruby client:

maxbp:dps max$ ./bench_redis/redis.rb 
8710 msg/s
9043 msg/s
9063 msg/s
8971 msg/s
8972 msg/s
8885 msg/s
8852 msg/s

It is not what we are waiting for from C server with replication and without storing messages!

Implementation details

Full replication is used. You should install 2-3 servers in cluster for each independent shard of channels

Replication is using 50 milliseconds delay to group messages

Channel history is a mandatory feature, because message can arrive between http requests

Channel history is strictly limited, because user should always receive some limited amount of data

Server will temporary disable replication if it is going too slowly

TODO:

  • Add more robust nodes discovering
  • Add docs
  • Add channels closing support
  • Compare to replicated redis pub/sub