• Intoduction
• Quick Start
• Use cases and deployment scenarios
• Startup and Discovery
• Operation
• Requests, Tasks and Job Classes
• Failure scenarios
• Configuration

Introduction

NkCLUSTER is a framework for creating clusters of Erlang nodes of any size, and distributing and managing jobs into them. It uses its own cluster management solution, based on NkDIST, riak_core and a custom distribution protocol. NkCLUSTER is one of the core pieces of the upcoming NetComposer platform, but can be used on its own.

Standard Erlang clusters are very convenient and easy to use, but they have some important limitations:

• Since each Erlang node must open a connection to any other node in the cluster, it is usually not practical to scale beyond about 50-100 nodes (hidden nodes are a possible workaround).
• Limited transport options, only TCP is (easily) available. Not very firewall friendly.
• Sending large messages can affect the latency of other, small messages.
• Adding and removing nodes does not redistribute the load on the nodes.
• In practical terms, all nodes must usually belong to the same LAN.

NkCLUSTER allows the creation of clusters with a very large number of nodes. It uses a hybrid approach. Any node in the cluster can have two different roles:

• Control role: They receive, share and process requests from the network, and manage and send jobs to nodes with 'worker' role.
• Worker role: They receive job requests from any node with control role and execute them, and can also receive specialized network traffic to process. They can have also special jobs like being a network router or a disk server.

All of the nodes of the cluster run the same Erlang application (nkcluster), but the first 'N' nodes in the cluster are primary nodes (they have control and worker roles), and the rest of the nodes are secondary nodes (they only have the worker role). N must be power of two, typically 16, 32, 64 or 128. Primary nodes create a riak_core cluster among them. The set of all primary nodes is called the primary cluster, a subset of the whole cluster.

Full nodes talk among them using standard Erlang distribution protocol. From the NkCLUSTER point of view, worker nodes talk only with their node proxy, placed at some control node, using TCP, TLS, SCTP, WS or WSS transports (they can of course talk with other worker nodes or whatever they want to using other means) . NkCLUSTER is designed to allow worker nodes to be deployed at WAN distances from control nodes (for example, at different cloud providers). They can be very firewall-friendly, for example using websockets transports on port 80 or 443, and in some circumstances, without any opened incoming port. However, all control nodes should still belong to the same LAN.

NkCLUSTER uses the concepts of requests, tasks, events and job classes. For each worker node, a node proxy process is started at some specific control node, and can be used to send requests and tasks to its managed worker node. Requests are short-lived RPC calls. Tasks are long-living Erlang processes running at a worker node and managed from a control process at a control node. Events are pieces of information sent from a worker node to its node proxy. All of them must belong to a previously defined job_class.

Any node in the cluster can have a set of labels or metadata associated with it, with an url-like format. For example, core;group=group1;master, meta2;labelA=1;labelB=2 would add two metadata items, core and meta2, each with some keys (master, labelA...) and, optionally, a value the key (group1, 2...). This metadata can be used for any purpose, for example to group nodes or decide where to put an specific Task in the cluster.

NkCLUSTER includes some out-of-the-box tools like metadata management, hot remote loading (or updating) of Erlang code, launching OS commands, Docker management, etc. NkCLUSTER allows on the fly addition and removal of nodes (both control and workers) and it is designed to allow jobs to be developed so that service is not disrupted at any moment.

NkCLUSTER scales seamlessly, from a single machine to a 10-20 control+worker nodes cluster, all the way to a huge cluster, where, for example, 50-100 control nodes manage thousands or tens of thousands of worker nodes. In a future version, NkCLUSTER will allow dynamic creation of nodes in public clouds like Amazon or Azure or private ones like OpenStack.

NkCLUSTER requires Erlang R17+.

Quick Start

git clone https://github.com/Nekso/nkcluster.git
make

Then, you can start five different consoles, and start five nodes, one at each: make dev1, make dev2, make dev3, make dev4 and make dev5.

Nodes 1, 2 and 3 are control/worker nodes. Nodes 4 and 5 are worker nodes. The cluster should discover everything automatically, but you must deploy the cluster plan. At any node of 1, 2 or 3:

> nkcluster_nodes:get_nodes().
[<<"dev1">>,<<"dev2">>,<<"dev3">>,<<"dev4">>,<<"dev5">>]

> nkcluster_nodes:get_node_info(<<"dev4">>).
{ok, #{id=><<"dev4">>, ...}}

> nkcluster:call(<<"dev3">>, erlang, node, [], 5000). 
{reply, '[email protected]'}

Use cases and deployment scenarios

NkCLUSTER is designed to solve an specific type of distributed work problem. It is probably useful for other scenarios, but it is specially well suited for the following case, where any job class is a compound of up to three layers:

• A mandatory, lightweight control, manager, or smart layer, written in Erlang (or any other BEAM language, like Elixir or LFE) that runs at all control nodes, so that any of them can receive a request to start or manage any task associated to this class. Optionally, a helper lightweight OS process or docker container could be started at its co-located worker node (written in any programming language).
• Optionally, a possibly heavyweight set of OS processes, docker containers or pods, running at specific worker nodes, and managed locally by an Erlang application sent from the control cluster over the wire, that talks with the controllers at the control nodes.
• Lastly, an optional, lightweight state associated with each task, also written in Erlang, and running in the same control node than the node proxy process for the worker node where the real job happens is located. Since node proxies are spread evenly among the cluster, your state processes will also be spread automatically.

NkCLUSTER includes full support for managing OS processes (using NkLIB) and for managing Docker containers (using NkDOCKER). The NkCLUSTER recommended approach is having, at the worker nodes, a local Erlang application monitoring processes and containers, and sending high-level, aggregated information to the control cluster, instead of sending raw stdout or similar low-level information. This way, even if the connection is temporarily lost, the worker node can continue working at some extend, while the connection to the control nodes is re-established.

This architecture is quite well suited for many real case scenarios, for example:

• SIP/WebRTC media processing framework, where:
  • NkSIP and WebRTC controllers run at every control node, accepting connections and managing SIP and WebRTC signaling.
  • each specific call or conference is an Erlang process that is evenly distributed at the control cluster.
  • a number of Freeswitch instances are started at worker nodes as docker containers for the heavy media management (transcoding, recording, etc.)
• Software-defined storage network based on Ceph, where:
  • the Ceph control daemons run in all of the control nodes.
  • the disk control daemons run at each worker node offering disks to the network (the middle layer would not be used in this case).
• Highly available Kubernetes cluster, where:
  • etcd and controlling processes run at the control nodes.
  • the real work docker and controllers run at each specific worker node.
• Parallel or streaming processing, where:
  • the control nodes receive streaming information to process, and decide which worker node must do the real processing. * For each task, an equivalent Erlang process is started at the control cluster.
  • The selected worker node receives the piece of information and process it, sending the response back.
• IoT platform, where:
  • the remote devices discover, talking with the control cluster, their associated worker node.
  • the device connects directly to their assigned worker node. Each worker node can handle a big number of devices. In case of failure, they find a new node and try to re-connect to it.

Depending on the configuration, several possible scenarios are supported:

• Standard. Each node listens on a defined port, using any supported transport. Then, you configure all nodes with the addresses of the control cluster (by hand or using NAPTR, SRV or multiple-DNS resolution):
  • The first node is started as a control+worker node.
  • Zero or more nodes are started as control+worker nodes. They automatically discover other control nodes, connect to them and form a riak_core cluster.
  • Zero or more nodes are started as workers only, and, using configuration or DNS, they discover and connect to a random control node. The receiving control node decides the final location for the control process (using riak_core distribution) and the correspoding node proxy process starts there.
  • Each node proxy process starts a connection to its controlled worker node. In case of failure, or if a specific request asks for an exclusive connection (for example for sending big files), a new connection is started.
• No-discovery. If you don't offer a cluster_addr parameter, no discovery should occur. Control nodes must be joined manually, and connections to worker nodes must be explicitly started.
• No-listening. Worker nodes can work without a listening address (for example, behind a firewall not accepting any incoming connection). In this case, the discovery connections are reused from the worker nodes. However, in case of connection failure, control nodes must wait for the next discovery before reconnecting to the nodes.

Startup and Discovery

The same Erlang application (nkcluster) is used at control (control+worker nodes actually) nodes and worker-only nodes. When configuring a new node you must supply:

• A cluster name.
• A set of listening network points, using NkPACKET url-like format, for example nkcluster://localhost;transport=tcp, nkcluster://localhost;transport=wss
• A password.
• If the node is going to be a control node (besides being a worker node) or not.
• Metadata associated with the node.
• Optionally, one or several network addresses to discover control nodes.

If a set of cluster discovery addresses are configured, the node will extract from them all transports, IP addresses and ports, it will randomize the list and try to connect to each one, until one accepts the connection. You should include all planned control nodes, DNS addresses returning several addresses, etc, for example:

{cluster_addr, "nkcluster://my_domain, nkcluster://10.0.0.1:1234;transport=wss"}

It this example, the node will extract all IP addresses and ports from my_domain (using DNS), and will add 10.0.0.1:1234 using wss transport to the list. It will then randomize the list and start trying to connect to each one in turn.

The password for this specific node and tls options can be included in the url:

{cluster_addr, "nkcluster://10.0.0.1:1234;transport=wss;password=pass1;tls_depth=2"}

NkPACKET supports the following DNS records for discovery:

• NAPTR records. If you don't supply port or transport, it will try to find a NAPTR record for the domain. For example, with this record:

example.com NAPTR 10 100 "S" "NKS+D2W" "" _nks._ws.example.com. example.com NAPTR 10 200 "S" "NKS+D2T" "" _nks._tcp.example.com. example.com NAPTR 10 300 "S" "NK+D2W" "" _nk._ws.example.com. example.com NAPTR 10 400 "S" "NK+D2T" "" _nk._tcp.example.com. example.com NAPTR 10 500 "S" "NK+D2S" "" _nk._sctp.example.com. ``` NkCLUSTER will first try WSS transport (resolving _nks._ws.example.com as a SRV domain find IPs and ports), then TLS, then WS, then TCP and finally SCTP as a last resort.

• SRV records. After a NAPTR respone, or if you supplied the desired transport, the corresponding SRV record will be resolved. For example, if tcp was selected, with this record:

_nk._tcp.example.com. 86400 IN SRV 0 5 1972 cluster.example.com. ``` NkCLUSER will try to resolve cluster.example.com, taking all IPs from there and using port `1972`.

• Round-robin DNS. Each time NkCLUSTER must resolve an IP address, if it returns multiple A records, it will randomize the list.

The receiving control node will accept the connection, and both parties will authenticate each other, sending a challenge (the node UUID, a random string auto-generated at boot) that must be signed using the local password, with PBKDF2 protocol. If everything is ok, the control node will select the right node to host the node proxy process and it will start it there. If the selected node is different, the node proxy will try to start a direct connection to the worker, if possible.

Using this node proxy process, you can send requests and start tasks at the worker node. Worker nodes will also send periodic information about its state (status, cpu, memory, etc.). If the connection fails, the control process will try to set it up again. Worker nodes will also try to connect again by themselves if no control node contacts them in a while. In some cases (like a network split) a single worker node could be connected to several proxy processes at different nodes, but this should be a temporary situation, resolved once the cluster converges again.

Operation

Based on riak_core, NkCLUSTER allows the addition and removal of nodes at any moment.

When a new control node is added, it automatically discovers and joins the riak_core cluster. While the cluster is reorganized, node proxy processes are relocated to be evenly distributed among the cluster again. When a node is removed, the opposite happens, also automatically.

Worker nodes can also be added and removed at any moment, and the changes are recognized automatically by the control nodes. Worker nodes can operate at the following states:

• Launching: The node is currently starting.
• Connecting: The cluster is currently trying to connect to the node.
• Ready: The node is ready to receive requests and tasks.
• Standby: The node is ready, but it is not currently accepting any new task.
• Stopping: The node is scheduled to stop as soon as no remaining tasks are running. It does not accept any new task. All tasks are notified to stop as soon as possible. Once all tasks have stopped, the status changes to Stopped automatically.
• Stopped: The node is alive but stopped, not accepting any new task.
• Not Connected: The cluster is not currently able to connect to the node.

Requests are allowed in ready, standby, stopping and stopped states.

Requests, Tasks and Job Classes

Before sending any request or task to a worker node, you must define a job_class module, implementing the nkcluster_job_class behaviour. This callback module must be available a both sides (control and worker). You must implement the following callbacks:

You can send your own Erlang module (or modules) to the remote side, over the wire, using nkcluster:load_module/3 or nkcluster:load_modules/3.

Once defined your job_class module, you can send requests, start tasks and send messages to tasks. Tasks can send events back to the job class.

You can send requests calling nkcluster:request/3,4. The callback request/2 will be called at the remote (worker) side, and your call will block until a response is sent back. You can define a timeout, and also ask NkCLUSTER to start a new, exclusive connection to the worker if possible. If the connection or the node fails, an error will be returned. If you are not asking for a new, exclusive connection, you request processing time must be very short (< 100 msecs). Otherwise, the periodic ping will be delayed and the connection may be dropped.

For long-running jobs, you must start a new task, calling nkcluster:task/3,4. The callback task/2 will be called at the remote side, and it must start a new Erlang process and return its pid() and, optionally, a reply. A job_id is returned to the caller, along with the reply if sent. You can send mesages to any started task calling nkcluster_jobs:command/4,5.

The started task can send events back to the control node at any moment, calling nkcluster_jobs:send_event/2. The event will arrive at the control node, and the callback event/2 will be called for the corresponding job class. NkCLUSTER will also send some automatic events:

Built-in requests

NkCLUSTER offers some standard utility requests out of the box, available in the nkcluster module:

Failure scenarios

Connection failure

When the main connection of a node proxy to its managed worker node fails, the following things happen:

• An {error, connection_failed} error is returned to all pending requests sent over this connection and still waiting for a response. Since the connection has failed, the remote connection (and the request itself, since it is usually running in the same process) have also probably failed (you can't however assume this for sure).
• An {nkcluster, {node_status, not_connected}} event is sent to all classes that the control node has seen (all classes that have received at last one event). Job classes must then assume that some events may have been lost, and must try to recover its state once the connection is alive again.
• The control node will try to connect immediately to the remote node (only if it published one or several listening points). If it fails, it will try again periodically. Meanwhile, all new requests will fail.
• If the remote worker didn't publish any listening address, the control process exists. The remote node should try to discover nodes again periodically, and a new control process will then be started.

Requests can start a secondary, exclusive connection. In this case, the failure of the main connection does not affect them (but affect their events). If this exclusive connection is the one who fails, the request will fail in the same manner described above.

Node failure

If the node proxy process fails (because of the failure of the whole control node or a bug), all pending requests will also fail (since they are using gen_server:call/3 under the hood). The worker node will detect this and try to announce itself again. The control cluster will then start a new node proxy process at the same (if again alive) or other node.

If the worker node process fails (because of the failure of the whole worker node or a bug), the control process will enter into not_connected state, notifying all detected job classes as described before. If the worker node no longer exists, you must call nkcluster_nodes:stop(NodeId) to remove it.

Configuration

NkCLUSTER uses standard Erlang application environment variables. The same Erlang application is used for agents and controllers.