gen_rpc: A scalable RPC library for Erlang-VM based languages

Copy Markdown View Source

Overview

Build Dependencies

To build this project you need to have the following:

  • Erlang/OTP >= 25.0 (tested up to 29)

  • git >= 1.7

  • GNU make >= 3.80

  • rebar3 >= 3.24 (3.28 and later require OTP 26 or newer)

  • Docker with the Compose plugin, only for the integration tests

Usage

Getting started with gen_rpc is easy. First, add the appropriate dependency line to your rebar.config:

{deps, [
    {gen_rpc, {git, "https://github.com/priestjim/gen_rpc.git", {branch, "main"}}}
]}.

Or if you're using hex.pm/rebar3:

{deps, [
    {gen_rpc, "~> 3.0"}
]}.

Or if you're using Elixir/Mix:

def project do
  [
    deps: [
      {:gen_rpc, "~> 3.0"}
    ]
  ]
end

Then, add gen_rpc as a dependency application to your .app.src/.app file:

{application, my_app, [
    {applications, [kernel, stdlib, gen_rpc]}
]}

Or your mix.exs file:

def application do
  applications: [:gen_rpc]
end

Finally, start a couple of nodes to test it out:

(my_app@127.0.0.1)1> gen_rpc:call('other_node@1.2.3.4', erlang, node, []).
'other_node@1.2.3.4'

API

gen_rpc implements only the subset of the functions of the rpc library that make sense for the problem it's trying to solve. The library's function interface and return values is 100% compatible with rpc with only one addition: Error return values include {badrpc, Error} for RPC-based errors but also {badtcp, Error} for TCP-based errors.

For more information on what the functions below do, run erl -man rpc.

Functions exported

  • call(NodeOrNodeAndKey, Module, Function, Args) and call(NodeOrNodeAndKey, Module, Function, Args, Timeout): A blocking synchronous call, in the gen_server fashion.

  • cast(NodeOrNodeAndKey, Module, Function, Args): A non-blocking fire-and-forget call.

  • async_call(NodeOrNodeAndKey, Module, Function, Args), yield(Key), nb_yield(Key) and nb_yield(Key, Timeout): Promise-based calls. Make a call with async_call and retrieve the result asynchronously, when you need it with yield or nb_yield.

  • multicall(Module, Function, Args), multicall(Nodes, Module, Function, Args), multicall(Module, Function, Args, Timeout) and multicall(NodesOrNodesWithKeys, Module, Function, Args, Timeout): Multi-node version of the call function.

  • abcast(NodesOrNodesWithKeys, Name, Msg) and abcast(Name, Msg): An asynchronous broadcast function, sending the message Msg to the named process Name in all the nodes in NodesOrNodesWithKeys.

  • sbcast(NodesOrNodesWithKeys, Name, Msg) and sbcast(Name, Msg): A synchronous broadcast function, sending the message Msg to the named process Name in all the nodes in NodesOrNodesWithKeys. Returns the nodes in which the named process is alive and the nodes in which it isn't.

  • eval_everywhere(Module, Function, Args) and eval_everywhere(NodesOrNodesWithKeys, Module, Function, Args): Multi-node version of the cast function.

  • nodes(): Returns the list of nodes gen_rpc currently has client connections to.

  • monitor_node(Node, Flag): Subscribes (Flag is true) or unsubscribes (Flag is false) the calling process to connection state changes of Node. The subscribed process receives a {nodeup, Node} message when the first gen_rpc connection to Node is established and a {nodedown, Node} message when the last connection to Node (including the ones opened with per-key sharding) is gone. Please note that in contrast to erlang:monitor_node, calling gen_rpc:monitor_node multiple times will result to only one registration per process and node, and that subscriptions are removed automatically when the subscribed process exits. Always returns true.

Per-Key Sharding

gen_rpc supports multiple outgoing connections per node using a key of arbitrary type to differentiate between connections. To leverage this feature, replace Node in your calls (single-node and multi-node alike) with {Node, Key}. The Key is hashed using erlang:phash2/1, attached to the client process name and a new connection is initiated.

Attention: When using functions that call gen_rpc:nodes/0 implicitly (such as gen_rpc:multicall/3), the channels used to communicate to the nodes are the keyless ones. To leverage the sharded functionality, pre-create your {Node, Key} lists and pass them as the node list in the multi-node function.

Module version control

gen_rpc supports executing RPC calls on remote nodes that are running only specific module versions. To leverage that feature, in place of Module in the section above, use {Module, Version}. If the remote module is not on the version requested a {badrpc,incompatible} will be returned.

Application settings

  • tcp_server_port: The plain TCP port gen_rpc will use for incoming connections or false if you do not want plain TCP enabled.

  • tcp_client_port: The plain TCP port gen_rpc will use for outgoing connections.

  • ssl_server_port: The port gen_rpc will use for incoming SSL connections or false if you do not want SSL enabled.

  • ssl_client_port: The port gen_rpc will use for outgoing SSL connections.

  • ssl_server_options and ssl_client_options: Settings for the ssl interface that gen_rpc will use to connect to a remote gen_rpc server.

  • default_client_driver: The default driver gen_rpc is going to use to connect to remote gen_rpc nodes. It should be either tcp or ssl.

  • client_config_per_node: A map of Node => {Driver, Port} or Node => Port that instructs gen_rpc on the Port and/or Driver to use when connecting to a Node. If you prefer to use an external discovery service to map Nodes to {Driver, Port} tuples, instead of the map, you'll need to define a {Module, Function} tuple instead with a function that takes the Node as its single argument, consumes the external discovery service and returns a {Driver, Port} tuple.

  • rpc_module_control: Set it to blacklist to define a list of modules that will not be exposed to gen_rpc or to whitelist to define the list of modules that will be exposed to gen_rpc. Set it to disabled to disable this feature.

  • rpc_module_list: The list of modules that are going to be blacklisted or whitelisted.

  • authentication_timeout: Default timeout for the authentication state of an incoming connection in milliseconds. Used to protect against half-open connections in a DoS attack.

  • connect_timeout: Default timeout for the initial node-to-node connection in milliseconds.

  • send_timeout: Default timeout for the transmission of a request (call/cast etc.) from the local node to the remote node in milliseconds.

  • call_receive_timeout: Default timeout for the reception of a response in a call in milliseconds.

  • sbcast_receive_timeout: Default timeout for the reception of a response in an sbcast in milliseconds.

  • client_inactivity_timeout: Inactivity period in milliseconds after which a client connection to a node will be closed (and hence have the TCP file descriptor freed).

  • server_inactivity_timeout: Inactivity period in milliseconds after which a server port will be closed (and hence have the TCP file descriptor freed).

  • keepalive_interval: Interval in seconds at which a client connection checks whether any data has been received from the server. If nothing was received, the client sends a ping that resets the server's inactivity timeout, so idle but healthy connections are not closed by server_inactivity_timeout, and a connection that cannot be written to anymore is detected and closed. Keepalive pings do not count as client activity so they do not postpone client_inactivity_timeout. Set to 0 to disable. Defaults to 60.

  • cookie_per_node: Cookie used to authenticate connections with each node, see Cookie per Node. Defaults to {internal, #{}}, which uses the node's Erlang cookie for every node.

  • async_call_inactivity_timeout: Inactivity period in milliseconds after which a pending process holding an async_call return value will exit. This is used for process sanitation purposes so please make sure to set it in a sufficiently high number (or infinity).

  • tcp_server_options and tcp_client_options: Lists of custom gen_tcp socket options for the plain TCP driver, applied to the listening socket and to all outgoing connections respectively (they are global, not per node). They are merged with the defaults in include/tcp.hrl and override them by option name, e.g. {tcp_client_options, [{sndbuf, 262144}, {nodelay, false}]}. Do not override binary, {packet,4} or {active,_}, since gen_rpc depends on them. Defaults to []. SO_REUSEPORT is not enabled by default; to let multiple instances share a listening port, set {tcp_server_options, [{reuseport, true}]} (OTP 26+).

  • named_acceptors: Register each acceptor process under the name gen_rpc.acceptor.<ip>:<port> (the address and port of the connected peer) so that it can be found with whereis/1 and shows up by name in observer. The peer's port changes with every connection and atoms are never garbage collected, so every connection creates a new atom for the lifetime of the VM. Only enable it if you need this observability and can afford the growth of the atom table, for example not with nodes that receive lots of short-lived connections. Client processes are always named after the node (and key) they connect to, since they are looked up by name. Defaults to false.

  • socket_keepalive_idle: Seconds idle after the last packet of data sent to start sending keepalive probes (applies to both drivers, on Linux and macOS only).

  • socket_keepalive_interval: Seconds between keepalive probes.

  • socket_keepalive_count: Probs lost to consider the socket closed

Logging

gen_rpc uses hut for logging. This allows the developer to integrate the logging library of their choice by providing the appropriate definition in their rebar.config. The default logging facility of hut is SASL.

For more information on how to enable gen_rpc to use your own logging facility, consult the README.md of hut.

SSL Configuration

gen_rpc supports SSL for inter-node communication. This allows secure communication and execution over insecure channels such as the internet, essentially allowing a trully globally distributed Erlang/Elixir setup. gen_rpc is very opinionated on how SSL should be configured and the bundled default options include:

  • TLS 1.3 AEAD cipher suites with a forward secret, elliptic curve only key exchange, using hybrid post-quantum ML-KEM groups (x25519mlkem768, secp256r1mlkem768, secp384r1mlkem1024) when the running Erlang/OTP release supports them (see Erlang/OTP versions), followed by x25519, secp384r1 and secp256r1

  • Both server and client-based SSL node CN (Common Name) verification

  • ECDSA only signature algorithms, for both the handshake and the certificate chain: node and CA certificates must be ECDSA (RSA certificates are rejected)

  • TLS 1.3 enforcement (TLS 1.2 and older are not supported, as post-quantum key exchange is only defined for TLS 1.3)

All of these settings can be found in include/ssl.hrl and overriden by redefining the necessary option in ssl_client_options and ssl_server_options. To actually use SSL support, you'll need to define in both ssl_client_options and ssl_server_options:

  • The public and private keys in PEM format, for the node you're running gen_rpc on, using the usual certfile, keyfile options.

  • The public key of the CA that signs the node's key and the public key(s) of CA that gen_rpc should trust, included in the file cacertfile points at.

To generate your own self-signed CA and node certificates, numerous articles can be found online such as this.

Usually, the CA that will be signing your client and server SSL certificates will be the same so a nominal sys.confg that includes SSL support for gen_rpc will look like:

    {gen_rpc, [
        {ssl_client_options, [
            {certfile, "priv/cert.pem"},
            {keyfile, "priv/cert.key"},
            {cacertfile, "priv/ca.pem"}
        ]},
        {ssl_server_options, [
            {certfile, "priv/cert.pem"},
            {keyfile, "priv/cert.key"},
            {cacertfile, "priv/ca.pem"}
        ]}
    ]}

Only EC (ECDSA) certificates are supported, for example on the prime256v1 (secp256r1), secp384r1 or secp521r1 curves. Sample certificates that are used by the test suite can be found in priv/ec_ssl. A node certificate and a CA can be generated with openssl:

openssl ecparam -name prime256v1 -genkey -noout -out ca.key.pem
openssl req -x509 -new -key ca.key.pem -sha256 -days 3650 -subj "/CN=gen_rpc CA" -out ca.cert.pem
openssl ecparam -name prime256v1 -genkey -noout -out node.key.pem
openssl req -new -key node.key.pem -subj "/CN=gen_rpc@10.0.0.1" -out node.csr.pem
openssl x509 -req -in node.csr.pem -CA ca.cert.pem -CAkey ca.key.pem -CAcreateserial -sha256 -days 365 -out node.cert.pem

The certificate Common Name must be the full name of the node, such as gen_rpc@10.0.0.1, as gen_rpc verifies it against the name of the node it connects to.

For multi-site deployments, a performant setup can be provisioned with edge gen_rpc nodes using SSL over the internet and plain TCP for internal data exchange. In that case, non-edge nodes can have {ssl_server_port, false} and {default_client_driver, tcp} and edge nodes can have their plain TCP port firewalled externally and {default_client_driver, ssl}.

Erlang/OTP versions

gen_rpc requires Erlang/OTP 25.0 or later and is continuously tested on OTP 25 to 29 on Ubuntu 24.04 (see .github/workflows/ci.yml). The code is the same on every release, but some behaviour depends on the OTP version you build and run with:

  • Post-quantum key exchange: the hybrid ML-KEM groups are only offered if ssl:groups/0 of the running release lists them (verified on OTP 29). gen_rpc_helper:get_ssl_groups/0 filters the preferred list at runtime, so on older releases the connection transparently falls back to x25519, secp384r1 and secp256r1. Both nodes of a connection must support a common group, and a PQ-capable node can talk to a classical one.

  • TLS 1.3 only: since TLS 1.2 is not enabled, every node that talks over SSL needs TLS 1.3, which all the supported releases provide. There is no fallback for peers running with older settings, upgrade all the nodes together.

  • Tests: the test suites start the second node with the peer module, which exists since OTP 25 (the deprecated slave module is not used).

  • Building: the code builds with warnings_as_errors, so deprecation warnings of newer releases fail the build. Variables that start with an underscore must not be reused once bound (OTP 29) and maybe is a reserved word. If you embed gen_rpc in a project that targets several releases, keep these in mind. The development profile pins lager ~> 3.9.2, which builds on all the supported releases.

  • Docker integration tests: the image is based on erlang:29. To test another release change the FROM line of test/integration/Dockerfile.

By default gen_rpc authenticates connections using the Erlang cookie of the node (erlang:get_cookie()). If you need a different secret for each node you connect to, set cookie_per_node to {internal, Map} where Map is a map of node_name => cookie, for example:

    {gen_rpc, [
        {cookie_per_node, {internal, #{
            'gen_rpc_slave@127.0.0.1' => slave_cookie
        }}}
    ]}

A client sends the cookie configured for the node it connects to, and a server validates that cookie against the one configured for the node that is connecting to it. So for two nodes A and B to talk to each other, A should have B's cookie set to the shared secret and B should have A's cookie set to the same secret. Nodes that are not in the map use the node's Erlang cookie. A connection with a mismatching cookie is rejected and the call returns {badrpc, invalid_cookie}.

If you keep your cookies in an external service, set cookie_per_node to {external, ModuleName}. ModuleName should export get_cookie(Node) which should return the cookie (an atom) to use with Node for both outgoing and incoming connections.

External Source Support

gen_rpc can call an external module to provide driver/port mappings in case you want to use an external discovery service like etcd for node configuration management. The module should export get_config(Node), which takes the Node as an argument and should return either {Driver, Port} (Driver being tcp or ssl and Port being the port the remote node's gen_rpc's driver is listening in) or {error, Reason} (if the service is unavailable). To set it, change client_config_per_node from the default of {internal, #{}} to {external, ModuleName} where ModuleName is the module that exports get_config/1.

Build Targets

gen_rpc bundles a Makefile that makes development straightforward.

To build gen_rpc simply run:

make

To run the full test suite, run:

make test

To run the full test suite, the XRef tool and Dialyzer, run:

make dist

To build the project and drop in a console while developing, run:

make shell-master

or

make shell-slave

If you want to run a "master" and a "slave" gen_rpc nodes to run tests.

To clean every build artifact and log, run:

make distclean

Testing

A full suite of tests has been implemented for gen_rpc. You can run the CT-based test suite, dialyzer and xref by:

make dist

If you have Docker available on your system, you can run dynamic integration tests with "physically" separated hosts/nodes by running the command:

make integration

The CT suites cover the local and remote functionality over the tcp and ssl drivers as well as node monitoring, cookie per node authentication and the client keepalive. The tests start a second Erlang node using the peer module and expect epmd to be running (make epmd takes care of that).

This uses Docker Compose (test/integration/compose.yaml) to launch 3 node containers and 1 master on a private network and runs the integration_SUITE CT test suite on the master. The exit code of make integration is the exit code of the test run.

Rationale

TL;DR: gen_rpc uses a mailbox-per-node architecture and gen_tcp processes to parallelize data reception from multiple nodes without blocking the VM's distributed port.

The reasons for developing gen_rpc became apparent after a lot of trial and error while trying to scale a distributed Erlang infrastructure using the rpc library initially and subsequently erlang:spawn/4 (remote spawn). Both these solutions suffer from very specific issues under a sufficiently high number of requests.

The rpc library operates by shipping data over the wire via Distributed Erlang's ports into a registered gen_server on the other side called rex (Remote EXecution server), which is running as part of the standard distribution. In high traffic scenarios, this allows the inherent problem of running a single gen_server server to manifest: mailbox flooding. As the number of nodes participating in a data exchange with the node in question increases, so do the messages that rex has to deal with, eventually becoming too much for the process to handle (don't forget this is confined to a single thread).

Enter erlang:spawn/4 (remote spawn from now on). Remote spawn dynamically spawns processes on a remote node, skipping the single-mailbox restriction that rex has. The are various libraries written to leverage that loophole (such as Rexi), however there's a catch.

Remote spawn was not designed to ship large amounts of data as part of the call's arguments. Hence, if you want to ship a large binary such as a picture or a transaction log (large can also be small if your network is slow) over remote spawn, sooner or later you'll see this message popping up in your logs if you have subscribed to the system monitor through erlang:system_monitor/2:

{monitor,<4685.187.0>,busy_dist_port,#Port<4685.41652>}

This message essentially means that the VM's distributed port pair was busy while the VM was trying to use it for some other task like Distributed Erlang heartbeat beacons or mnesia synchronization. This of course wrecks havoc in certain timing expectations these subsystems have and the results can be very problematic: the VM might detect a node as disconnected even though everything is perfectly healthy and mnesia might misdetect a network partition.

gen_rpc solves both these problems by sharding data coming from different nodes to different processes (hence different mailboxes) and by using a different gen_tcp port for different nodes (hence not utilizing the Distributed Erlang ports).

Architecture

In order to achieve the mailbox-per-node feature, gen_rpc uses a very specific architecture:

  • Whenever a client needs to send data to a remote node, it will perform a whereis to a process named after the remote node.

  • If the specified client process does not exist, it will request for a new one through the dispatcher process, which in turn will launch it through the appropriate client supervisor. Since this |whereis > request from dispatcher sequence > start client| can happen concurrently by many different processes, serializing it behind a gen_server allows us to avoid race conditions.

  • The dispatcher process launches the new client in a separate process, so that a node that is slow to answer does not hold up the clients of the other nodes. Requests for the same client that arrive in the meantime wait for the same result.

  • That process connects to the remote node's gen_rpc server, authenticates, and only then starts the client process through the client's supervisor, which hands the connection over to it.

  • The gen_rpc server server will ask the acceptor supervisor to launch a new acceptor process and hands it off the new socket connection.

  • The acceptor takes over the new socket and authenticates the client with the cookie configured for the connecting node (the Erlang cookie unless cookie_per_node says otherwise, see Cookie per Node) and any extra protocol-level authentication supported by the selected driver.

  • The client finally encodes the request (call, cast etc.) along with some metadata (the caller's PID and a reference) and sends it over the TCP channel. In case of an async call, the client also launches a process that will be responsible for handing the server's reply to the requester.

  • The acceptor on the other side decodes the TCP message received and spawns a new process that will perform the requested function. By spawning a process external to the server, the acceptor protects itself from misbehaving function calls.

  • As soon as the reply from the server is ready (only needed in async_call and call), the acceptor spawned process messages the server with the reply, the acceptor ships it through the TCP channel to the client and the client send the message back to the requester. In the case of async call, the client messages the spawned worker and the worker replies to the caller with the result.

All gen_tcp processes are properly linked so that any TCP failure will cascade and close the TCP channels and any new connection will allocate a new process and port.

An inactivity timeout has been implemented inside the client and server processes to free unused TCP connections after some time, in case that's needed. Each client also runs a keepalive that periodically pings the server when the connection is idle, so that connections that should stay open are not closed by the server's inactivity timeout and broken connections are detected quickly.

All client connections are registered with the gen_rpc_monitor process, which notifies the processes subscribed with gen_rpc:monitor_node/2 when a node connects or disconnects.

Performance

gen_rpc is being used in production extensively with over 150.000 incoming calls/sec/node on a 8-core Intel Xeon E5 CPU and Erlang 19.1. The median payload size is 500 KB. No stability or scalability issues have been detected in over a year.

Known Issues

  • When shipping an anonymous function over to another node, it will fail to execute because of the way Erlang implements anonymous functions (Erlang serializes the function metadata but not the function body). This issue also exists in both rpc and remote spawn.

Licensing

This project is published and distributed under the Apache License.

Contributing

Please see CONTRIBUTING.md

Contributors: