This library implements a stateful consistent hash ring. It's extremely fast (in benchmarks it's faster than all other implementations I've tested against, namely voicelayer/hash-ring and sile/hash_ring), it has no external dependencies, and is written in Elixir.
The algorithm is based on libketama. Nodes on the ring are broken into shards and each one is assigned an integer value in the keyspace, which is the set of integers from 1 to 2^32-1. The distribution of these shards is random, but deterministic.
Keys are then mapped to a shard by converting the key to a binary, hashing it with SHA-256, converting the hash to an integer in the keyspace, then finding the shard which is assigned the next highest value, if there is no next highest value, the lowest integer is used, which is how the "ring" is formed.
This implementation uses a general balanced tree, via Erlang's :gb_tree
module. Each shard
is inserted into the tree, and we use this data structure to efficiently lookup next-highest
key and smallest key. I suspect this is why libring
is faster than other implementations I've
benchmarked against.
Add :libring
to your deps, and run mix deps.get
.
def deps do
[
{:libring, "~> 1.0"}
]
end
You have two choices for managing hash rings in your application:
This API works with the raw ring data structure. It is the fastest implementation, and is best suited for when you have a single process which will need to access the ring, and which can hold the ring in its internal state.
ring = HashRing.new()
|> HashRing.add_node("a")
|> HashRing.add_node("b")
"a" = HashRing.key_to_node(ring, {:myworker, 123})
You can also specify the weight of each node, and add nodes in bulk:
ring = HashRing.new()
|> HashRing.add_nodes(["a", {"b", 64}])
|> HashRing.add_node("c", 200)
"c" = HashRing.key_to_node(ring, {:myworker, 123})
NOTE: Node names do not have to be strings, they can be atoms, tuples, etc.
This API works with rings which are held in the internal state of a GenServer process.
It supports the same API as HashRing
. Because of this, there is a performance overhead
due to the messaging, and the GenServer can be a potential bottleneck. If this is the case
you are better off exploring ways to use the raw HashRing
API. However this API is best suited
for situations where you have multiple processes accessing the ring, or need to maintain multiple
rings.
NOTE: You must have the :libring
application started to use this API.
{:ok, pid} = HashRing.Managed.new(:myring)
:ok = HashRing.Managed.add_node(:myring, "a")
:ok = HashRing.Managed.add_node(:myring, "b", 64)
:ok = HashRing.Managed.add_node(:myring, "c", 200)
"c" = HashRing.Managed.key_to_node(:myring, {:myworker, 123})
You can configure managed rings in config.exs
, and they will be created and initialized
when the :libring
application starts. Configured rings take two shapes, static and dynamic
rings. Static rings are simply those where the nodes are provided up front, although you can
always add/remove nodes manually at runtime; dynamic rings have Erlang node monitoring enabled,
and add or remove nodes on the ring based on cluster membership.
You can whitelist/blacklist nodes when using dynamic rings, so that only those nodes which you actually want to distribute work to are used in calculations. This configuration is shown below as well.
If you provide a whitelist, the blacklist will have no effect, and only nodes matching the whitelist
will be added. If you do not provide a whitelist, the blacklist will be used to filter nodes. If you
do not provide either, a default blacklist containing the ~r/^remsh.*$/
pattern from the example below,
which is a good default to prevent remote shell sessions (at least those done via releases) from causing
the ring to change.
The whitelist and blacklist only have an effect when monitor_nodes: true
.
Below is an example configuration:
config :libring,
rings: [
# A ring which automatically changes based on Erlang cluster membership,
# but does not allow nodes named "a" or "remsh*" to be added to the ring
ring_a: [monitor_nodes: true,
node_type: :visible,
node_blacklist: ["a", ~r/^remsh.*$/]],
# A ring which is composed of three nodes, of which "c" has a non-default weight of 200
# The default weight is 128
ring_b: [nodes: ["a", "b", {"c", 200}]]
]
If you have changes in mind that are significant or potentially time-consuming, please open an RFC-style PR first, where we can discuss your plans first. I don't want you to spend all your time crafting a PR that I ultimately reject because I don't think it's a good fit or is too large for me to review. Not that I plan to reject PRs in general, but I have to be careful to balance features with maintenance burden, or I will quickly be unable to manage the project.
Please ensure that you adhere to a commit style where logically related changes are in a single commit, or broken up in a way that eases review if necessary. Keep commit subject lines informative, but short, and provide additional detail in the extended message text if needed. If you can, mention relevant issue numbers in either the subject or the extended message.
Copyright (c) 2016 Paul Schoenfelder
This library is MIT licensed. See the LICENSE for details.