CPG (CloudI Process Groups)

Purpose

cpg provides a process group interface that is focused on availability and partition tolerance (in the CAP theorem). The pg process group implementation added in Erlang/OTP 23 by WhatsApp Inc. (Meta Platforms Inc. / Facebook Inc.) is based on cpg. The cpg interface is compatible with pg2 (removed in Erlang/OTP 24).

Features (Compare and Contrast)

cpg

• By default, cpg utilizes Erlang strings for group names (list of integers) and provides the ability to set a pattern string as a group name. A pattern string is a string that includes the "*" or "?" wildcard characters (equivalent to a ".+" regex while "**", "??", "*?", and "?*" are forbidden). When a group name is a pattern string, a process can be retrieved by matching the pattern (more information at the CloudI FAQ). To not use this approach for group names, refer to the Usage section below.
• cpg provides its internal state for usage in separate Erlang processes as cached data with the cpg_data module. That approach is more efficient than usage of ets.
• Each cpg scope is an atom used as a locally registered process name for the cpg scope Erlang process. Separate cpg scopes may be used to keep group memberships entirely separate.
• cpg data lookups are done based on the Erlang process being local or remote, or the relative age of the local membership to the group, or with random selection (using the terminology closest, furthest, random, local, remote, oldest, newest). closest prefers local processes if they are present while furthest prefers remote processes if they are present. The oldest process in a group is naturally the most stable process.
• cpg provides an interface for via process registry use (examples are provided in the tests).
• cpg supports hidden node connections (hidden node connections are a way to avoid distributed Erlang node connection limitations by not creating a fully-connected network topology).

pg (>= Erlang/OTP 23) (https://github.com/max-au/spg)

• pg uses one monitor per remote node (it takes longer to update a group after an Erlang process dies and may never remove remote group members).
• pg uses ets while cpg does not (cpg instead provides cached data for more efficient access to the process group data).

pg2 (< Erlang/OTP 24)

• pg2 uses global:trans/2 which is unable to handle network or node failures.
• pg2 uses ets while cpg does not (cpg instead provides cached data for more efficient access to the process group data).

gproc / syn

• Both are focused on consistency with leader election and are unable to be available when suffering network or node failures. Failures can cause unpredictable conflict resolution, in an attempt to achieve consistency.

Design

cpg is a Commutative/Convergent Replicated Data-Type (CRDT) that uses node ownership of Erlang processes to ensure a set of keys has add and remove operations that commute with an internal map data structure. The cpg module provides add and remove operations with the function names join and leave, that may only be called on the node that owns the Erlang process which is the value for the join or leave operation. The key is the process group name which represents a list of Erlang processes (with an single Erlang process being able to be added or removed any number of times).

All cpg join and leave operations change global state as a Commutative Replicated Data-Type (CmRDT) by sending the operation to the associated cpg Erlang process as a distributed Erlang message to all remote nodes after the operation successfully completes on the local node.

cpg also uses distributed Erlang node monitoring to handle netsplits as a Convergent Replicated Data-Type (CvRDT) by sending all of the internal cpg state to remote nodes that have recently connected. The associated cpg Erlang process on the remote node then performs a merge operation to make sure the count of each Erlang pid is consistent with the internal cpg state it received.

The CRDT functionality in cpg is most similar to the POLog (Partially Ordered Log of operations) though the cpg approach would instead be called an "Ordered Log of operations" because it is depending on Erlang messaging on a local node to have causal ordering (no vclocks are necessary to establish causality on the local node with the cpg scope Erlang process message queue providing an "Ordered Log of operations"). After a cpg operation completes successfully on the local node, it is sent to all remote nodes which act as read-only views of the local node.

The cpg scope process on the local node enforces causality by existing as the only read/write store of the local process memberships (i.e., serialized mutability similar to a mutex lock) while the remote nodes obtain the process memberships as soon as possible. If a remote node is down due to a netsplit, it will obtain the local node's state once it reconnects as described above.

Build

rebar get-deps
rebar compile

Usage

If you need non-string (not a list of integers) group names, set the cpg application group_storage env value to a module name that provides a dict module interface (e.g., use dict or mapsd).

Node names that have a prefix of NODE_script_process (where NODE is the current node name) are automatically ignored because they are assumed to be release scripts (e.g., nodetool). To process hidden node membership data, set the cpg application node_type env value to all (instead of visible).

Example

$ erl -sname cpg@localhost -pz ebin/ -pz deps/*/ebin/

(cpg@localhost)1> reltool_util:application_start(cpg).
ok
(cpg@localhost)2> cpg:join(groups_scope1, "Hello", self()).
ok
(cpg@localhost)3> cpg:join(groups_scope1, "World!", self()).
ok
(cpg@localhost)4> cpg:get_local_members(groups_scope1, "Hello").
{ok,"Hello",[<0.39.0>]}
(cpg@localhost)5> cpg:get_local_members(groups_scope1, "World!").
{ok,"World!",[<0.39.0>]}
(cpg@localhost)6> cpg:which_groups(groups_scope1).
["Hello","World!"]
(cpg@localhost)7> cpg:which_groups(groups_scope2).
[]

What does this example mean? The cpg interface allows you to define groups of Erlang processes and each group exists within a scope. A scope is represented as an atom which is used to locally register a cpg Erlang process using start_link/1. For a given cpg scope, any Erlang process can join or leave a group. The group name is a string (list of integers) due to the default usage of the trie data structure, but that can be changed (see the Usage section above). If the scope is not specified, the default scope is used: cpg_default_scope.

In the example, both the process group "Hello" and the process group "World!" are created within the groups_scope1 scope. Within both progress groups, a single Erlang process is added once. If more scopes were required, they could be created automatically by being provided within the cpg application scope list. There is no restriction on the number of process groups that can be created within a scope, and there is nothing limiting the number of Erlang processes that can be added to a single group. A single Erlang process can be added to a single process group in a single scope multiple times to change the probability of returning a particular Erlang process, when only a single process is requested from the cpg interface (e.g., from the get_closest_pid function).

Tests

rebar get-deps
rebar compile
ERL_LIBS="/path/to/proper" rebar eunit

Author

Michael Truog (mjtruog at protonmail dot com)

License

MIT License

References

1. Carlos Baquero, Paulo Sérgio Almeida, Ali Shoker. Making operation-based crdts operation-based. In Proceedings of the First Workshop on Principles and Practice of Eventual Consistency, page 7. ACM, 2014. http://haslab.uminho.pt/ashoker/files/opbaseddais14.pdf
2. Carlos Baquero, Paulo Sérgio Almeida, Ali Shoker. Pure Operation-Based Replicated Data Types. 2017. https://arxiv.org/abs/1710.04469