Lowest Random Weight hashing for neatly rebalancing hashes
Erlang
46
5 commits
updated Feb 2, 2018
This is an Erlang implementation of LRW (equivalent to HRW), an algorithm that allows clients to achieve distributed agreement on which site (or proxy) a given object is to be placed in. It accomplishes the same goal as consistent hashing, using an entirely different method.
This hashing mechanism allows to consistently hash to specific IP addresses, without needing to do bucketing at first, or ever keeping a fixed or ever-increasing amount of entries, while guaranteeing as little key redistribution as possible.
The all/2 function gives an ordering of all nodes:
1> lrw:all(my_item, [{127,0,0,1},{255,0,0,1},{198,2,1,2},{192,198,2,1}]).
[{192,198,2,1},{127,0,0,1},{255,0,0,1},{198,2,1,2}]
2> lrw:all(my_item, [{127,0,0,1},{198,2,1,2},{192,198,2,1},{10,10,100,10}]).
[{192,198,2,1},{127,0,0,1},{198,2,1,2},{10,10,100,10}]
3> lrw:all(my_item, [{127,0,0,1},{192,198,2,1}]).
[{192,198,2,1},{127,0,0,1}]
4> lrw:all(my_item, [{127,0,0,1},{192,18,211,12}, {23,66,77,88}, {252,11,11,11}]).
[{252,11,11,11},{127,0,0,1},{23,66,77,88},{192,18,211,12}]
5> lrw:all("my other item", [{127,0,0,1},{192,18,211,12}, {23,66,77,88}, {252,11,11,11}]).
[{23,66,77,88},{127,0,0,1},{252,11,11,11},{192,18,211,12}]
6> lrw:all(<<"my other item">>, [{127,0,0,1},{192,18,211,12}, {23,66,77,88}, {252,11,11,11}]).
[{252,11,11,11},{23,66,77,88},{127,0,0,1},{192,18,211,12}]
The top/3 functions returns a subset. It is using all/2 in its implementation:
1> lrw:top(12123, [{127,0,0,1},{255,0,0,1},{198,2,1,2},{192,198,2,1}], 2).
[{192,198,2,1},{127,0,0,1}]
The list of nodes can be any data type; not just IPs. For IPv4s, optimized
functions respecting the original algorithm from the paper is implemented in
lrw:all_ip/2 and lrw:top_ip/3.
Custom hashing functions can be passed by calling lrw:all/3 and
lrw:top/4, where the last argument of each function must be
a fun of the form fun(Key, Node) -> Number end.
Pullit in your projects. Compile standalone using rebar (you should have this installed globally on your node already, it's been in use for years, come on!):
$ rebar compile
Verify for type errors and discrepancies:
$ dialyzer --src src/*.erl
Then run the property-based tests (10,000 each) with:
$ rebar get-deps compile eunit --config rebar.tests.config
2.0.1: make hex-compatible and publish2.0.0: all/2-3 and top/3-4 are generic, and all_ip/2 and top_ip/3
are optimized for IPs and respect the original algorithm.1.1.0: Adding all_/2, top_/3, all/3, and top/4 to support
generic hashing.1.0.0: Initial commit5 commits
Erlang
100.0%
Lowest Random Weight hashing for neatly rebalancing hashes
Erlang
46
5 commits
updated Feb 2, 2018
This is an Erlang implementation of LRW (equivalent to HRW), an algorithm that allows clients to achieve distributed agreement on which site (or proxy) a given object is to be placed in. It accomplishes the same goal as consistent hashing, using an entirely different method.
This hashing mechanism allows to consistently hash to specific IP addresses, without needing to do bucketing at first, or ever keeping a fixed or ever-increasing amount of entries, while guaranteeing as little key redistribution as possible.
The all/2 function gives an ordering of all nodes:
1> lrw:all(my_item, [{127,0,0,1},{255,0,0,1},{198,2,1,2},{192,198,2,1}]).
[{192,198,2,1},{127,0,0,1},{255,0,0,1},{198,2,1,2}]
2> lrw:all(my_item, [{127,0,0,1},{198,2,1,2},{192,198,2,1},{10,10,100,10}]).
[{192,198,2,1},{127,0,0,1},{198,2,1,2},{10,10,100,10}]
3> lrw:all(my_item, [{127,0,0,1},{192,198,2,1}]).
[{192,198,2,1},{127,0,0,1}]
4> lrw:all(my_item, [{127,0,0,1},{192,18,211,12}, {23,66,77,88}, {252,11,11,11}]).
[{252,11,11,11},{127,0,0,1},{23,66,77,88},{192,18,211,12}]
5> lrw:all("my other item", [{127,0,0,1},{192,18,211,12}, {23,66,77,88}, {252,11,11,11}]).
[{23,66,77,88},{127,0,0,1},{252,11,11,11},{192,18,211,12}]
6> lrw:all(<<"my other item">>, [{127,0,0,1},{192,18,211,12}, {23,66,77,88}, {252,11,11,11}]).
[{252,11,11,11},{23,66,77,88},{127,0,0,1},{192,18,211,12}]
The top/3 functions returns a subset. It is using all/2 in its implementation:
1> lrw:top(12123, [{127,0,0,1},{255,0,0,1},{198,2,1,2},{192,198,2,1}], 2).
[{192,198,2,1},{127,0,0,1}]
The list of nodes can be any data type; not just IPs. For IPv4s, optimized
functions respecting the original algorithm from the paper is implemented in
lrw:all_ip/2 and lrw:top_ip/3.
Custom hashing functions can be passed by calling lrw:all/3 and
lrw:top/4, where the last argument of each function must be
a fun of the form fun(Key, Node) -> Number end.
Pullit in your projects. Compile standalone using rebar (you should have this installed globally on your node already, it's been in use for years, come on!):
$ rebar compile
Verify for type errors and discrepancies:
$ dialyzer --src src/*.erl
Then run the property-based tests (10,000 each) with:
$ rebar get-deps compile eunit --config rebar.tests.config
2.0.1: make hex-compatible and publish2.0.0: all/2-3 and top/3-4 are generic, and all_ip/2 and top_ip/3
are optimized for IPs and respect the original algorithm.1.1.0: Adding all_/2, top_/3, all/3, and top/4 to support
generic hashing.1.0.0: Initial commit5 commits
Erlang
100.0%