evernode-client-cluster-manager

Drop-in npm module that adds 14 cluster management handlers to any HotPocket contract running on Evernode. Designed for use with the Evernode Cluster Manager client tool but works with any HP JS client.

What's new in 1.3.4

mesh.idle_timeout via hp.cfg.override. When hp.cfg.override contains mesh.idle_timeout, the upgrade handler stores it in contract.config and post_exec.sh applies it to both patch.cfg and hp.cfg. Required for roundtimes exceeding mesh.idle_timeout × stage_slice% (default safe max 480000ms). Takes effect on next container restart. A hpcore PR (EvernodeXRPL/hpcore#414) has been raised to make this dynamic without restart.

What's new in 1.3.2

Dynamic round time via hp.cfg.override. When hp.cfg.override contains contract.consensus.roundtime, the upgrade handler stores it in contract.config and post_exec.sh applies it to patch.cfg via jq. hpcore reads consensus.roundtime dynamically from patch.cfg each ledger — confirmed taking effect within one consensus round on a live cluster without any container restart.

Dynamic log level via hp.cfg.override. When hp.cfg.override contains log.log_level, the upgrade handler now passes it through to contract.config and post_exec.sh applies it dynamically. Previously post_exec.sh had log level hardcoded to "dbg". Note: log level takes effect on next container restart on current hpcore versions — a hpcore PR has been raised (EvernodeXRPL/hpcore) to allow dynamic log level change without restart.

DEP0128 deprecation warning fixed. Corrected "main" field in package.json from "dist/index.js" to "index.js" and added "type": "commonjs". The contract build script copies the ncc bundle to the package root — not into a dist/ subdirectory — so the main path was pointing to the wrong location. Node.js 20 flagged this every round. Warning is now gone.

What's new in 1.3.0

MATURED flow now works correctly. The checkAndSendMatured function uses hotpocket-js-client to connect to existing UNL nodes and send the MATURED signal. Previously this was a dynamic inline require() inside the function body — ncc cannot statically bundle dynamic requires, so the module was silently missing from the compiled output and the function failed every round without error. Moved to a top-level static require so ncc bundles it correctly.

Full peer mesh maintenance. On every addNode, removeNode, and node promotion, the full UNL peer list is written to patch.cfg mesh.known_peers via ctx.updateConfig(). Previously nodes only had their single bootstrap peer in known_peers, meaning a cold restart with a dead bootstrap peer left the node completely isolated.

cluster.info written with full UNL peer list. Previously cluster.info contained only the anchor node and the new node. Now it contains all current UNL nodes, giving new nodes multiple peers to try when connecting and sending MATURED.

checkAndPromoteMatured updates full peer list via ctx.updatePeers. After promoting a new node to UNL, all existing UNL nodes receive the complete peer list as a live update — not just the single new peer as before.

What 1.2.2 fixed

Removed the purgePeers handler. It used hpcore's OVERWRITE mode (ctx.updatePeers(peers, "*")) which clears the entire req_known_remotes table and closes all live peer sessions at once. When sent simultaneously to every UNL node — which is what consensus does — every node tore down every peer connection in the same round, collapsing the cluster.

The ghost-peer cleanup it was designed to address is now handled correctly and automatically by removeNode and removePeer (since 1.2.1), which use FORCE mode on a single peer at a time. See CHANGELOG.md for full details.

What 1.2.1 fixed

Ghost peer retries after node removal. handleRemoveNode now flushes hpcore's req_known_remotes via ctx.updatePeers([], [peerStr]) so removed nodes don't generate endless Trying to connect <removed-node> retries every ~4 seconds. handleRemovePeer does the same for manual cleanup of stale entries.

Project setup

The package is a require() inside your contract code. Your contract lives in its own directory with a package.json and index.js. From an empty directory:

mkdir my-contract && cd my-contract
npm init -y
npm install [email protected] evernode-client-cluster-manager

The npm init -y step is mandatory. Without a package.json in the current directory, npm walks up looking for a parent project, finds one, and silently installs nothing — node_modules/ is never created.

Final layout:

my-contract/
├── package.json
├── index.js
└── node_modules/
    ├── hotpocket-nodejs-contract/
    └── evernode-client-cluster-manager/

Usage

'use strict';
const HotPocket      = require('hotpocket-nodejs-contract');
const ClusterManager = require('evernode-client-cluster-manager');

const VERSION = '1.0.0';

const contract = async (ctx) => {
    // One line — registers all handlers.
    // Returns true if a management command was handled — return early so your
    // business logic is skipped for that round.
    if (await ClusterManager.init(ctx, VERSION)) return;

    // Your business logic here
};

const hpc = new HotPocket.Contract();
hpc.init(contract);

Deployment

You have two paths to a running cluster:

Using the CLI tool (recommended): evernode-cluster-manager handles multi-node acquisition, bundling, deployment, live upgrades and ongoing cluster management.

Manual deployment with evdevkit:

evdevkit bundle my-contract <pubkey> /usr/bin/node -a index.js
evdevkit deploy bundle.zip <domain> <user_port>

Handlers

Readonly (9)

| Type | Description | |------|-------------| | status | Ledger info, contract ID, public key, version, readonly flag | | readCfg | Full running HP config from /contract/cfg/hp.cfg — includes mesh, user, node sections and known_peers | | readPatchCfg | Contract override config via ctx.getConfig() | | readEnvVars | Host environment variables from /contract/env.vars — external ports, quotas, security config | | readLog | Last N lines of hp.log | | readContractLog | Last N lines of rw.stdout.log or rw.stderr.log | | getBootstrapPeer | Returns the most stable peer for a new node to bootstrap from — prioritises original deploy nodes, then promoted nodes by seniority | | readClusterJson | Returns cluster.json from contract state — node membership, statuses, promotion history | | readAuthorizedPubkey | Returns authorized_pubkey.txt — the management key authorized to submit consensus inputs |

Consensus (5)

| Type | Description | |------|-------------| | upgrade | Deploy new contract bundle via base64, runs post_exec.sh. Supports dynamic roundtime and log level via hp.cfg.override | | addNode | Register a new node as non-UNL pending MATURED signal. Writes full UNL peer list to patch.cfg known_peers for cold restart resilience | | removeNode | Remove pubkey from UNL, update patch.cfg.known_peers, and flush hpcore's req_known_remotes to stop retry spam | | removePeer | Manual cleanup: flush a peer from patch.cfg.known_peers and req_known_remotes without UNL changes. Useful for orphan entries | | matured | Received from a non-UNL node when it has synced — marks it as acknowledged. Promoted to UNL after MATURITY_LCL_THRESHOLD ledgers |

Readonly handlers run on each node independently — no consensus required. Consensus handlers require all UNL nodes to agree before executing.

Autonomous (no input required)

These run on every consensus round inside init():

• checkAndPromoteMatured — promotes acknowledged nodes to UNL after stability threshold. Writes full peer list to patch.cfg via ctx.updateConfig() and broadcasts full peer list via ctx.updatePeers() after each promotion. Also prunes nodes stuck in status: created for more than 5 moments (never acknowledged, definitively failed).
• checkAndSendMatured — runs on non-UNL nodes. Reads cluster.info (full UNL peer list) to find existing nodes, connects and sends MATURED signal when synced. Retries up to 3 times.

Input format

All inputs are JSON strings sent via submitContractReadRequest (readonly) or submitContractInput (consensus) from the HP JS client.

// Readonly
{ "type": "status" }
{ "type": "readCfg" }
{ "type": "readPatchCfg" }
{ "type": "readEnvVars" }
{ "type": "readLog", "lines": 100 }
{ "type": "readContractLog", "lines": 100, "logFile": "stdout" }
{ "type": "getBootstrapPeer" }
{ "type": "readClusterJson" }
{ "type": "readAuthorizedPubkey" }

// Consensus
{ "type": "upgrade", "bundle": "<base64 encoded bundle.zip>" }
{ "type": "addNode", "pubkey": "ed...", "ip": "host.example.com", "peerPort": 22865, "userPort": 26865, "existingNodes": [...] }
{ "type": "removeNode", "pubkey": "ed...", "ip": "host.example.com", "peerPort": 22865 }
{ "type": "removePeer", "peerIp": "host.example.com", "peerPort": 22865 }
{ "type": "matured", "data": "ed..." }

Critical rules

• Never use non-deterministic values (Date.now, Math.random) in consensus handler outputs
• Always keep a VERSION constant and bump it on every upgrade so the cluster manager can track versions
• Never remove any handler — the cluster manager client depends on them
• The upgrade handler expects a valid bundle.zip containing the contract files
• ctx.updatePeers(peers, "*") is not safe to use from any consensus handler — it triggers hpcore OVERWRITE mode which closes all live peer sessions on every UNL node simultaneously. This is why 1.2.2 removed the purgePeers handler entirely

Requirements

• hotpocket-nodejs-contract >= 0.7.4
• Node.js >= 16
• Deployed on Evernode (requires HotPocket Docker environment with /contract/cfg/hp.cfg and /contract/env.vars)

Related

• Evernode Cluster Manager — the client tool that uses this package
• Evernode Host API — host discovery API for finding deployment targets
• CHANGELOG.md — version history