connect-qos

Connect middleware that helps maintain a high quality of service during heavy traffic. The basic idea is to identify bad actors and not penalize legitimate traffic more than necessary until proper mitigation can be activated.

Warning

While this library provides some basic HTTP (Layer 7) flood attack protection, it does NOT remove the need for proper multi-layered DDoS defenses.

It's recommended to monitor for 5xx errors and alarm if threshold exceeded -- otherwise you may face an attack and not know about it.

Getting Started with Connect/Express

Using Connect or Express?

const
	connect = require("connect"),
	http = require("http");
const { ConnectQOS } = require("connect-qos");

var app = connect()
	.use(new ConnectQOS().getMiddleware())
	.use(function(req, res) {
		res.end("Hello World!");
	});

http.createServer(app).listen(8392);

Getting Started with HTTP

Real coders don't use middleware? We've got you covered too...

const http = require("http");
const { ConnectQOS } = require("connect-qos");

var qos = new ConnectQOS();
var qosMiddleware = qos.getMiddleware();
http.createServer(function(req, res) {
	qosMiddleware(req, res, function() {
		res.end("Hello World!");
	});
}).listen(8392);

Middleware Options

• beforeThrottle(qosInstance, req, reason) - If a function is provided it will be invoked prior to throttling a request in case a decision is desired. Only if the function explicitly returns false will the throttle request be denied, not resulting in a 503 status.
• destroySocket (default: true) - If denying bad actor also destroy the socket to prevent reuse.

Additional Methods

Users may also invoke methods isBadHost(host), isBadIp(ip), or isBadSubnet(subnet) on the qos instance to check the status of a given host, IP, or subnet. These methods will return true or false indicating whether the actor is currently considered to be a bad actor. This can be done for TLS/SNI to provide additional layer 5 mitigations.

Subnet keys match the format produced by resolveSubnetFromIp — the full 4-octet network address with host bits zeroed (e.g. 103.142.223.0 for a /24). IPv4-mapped IPv6 (::ffff:a.b.c.d) is unwrapped to IPv4 first; pure IPv6 is used as-is.

Goals

1. Identify potential bad actors
2. Respond with 503 (BUSY) during heavy traffic for bad actors only
3. Very light weight, no complex algorithms

Options

For you tweakers out there, here's some levers to pull:

• minLag (default: 70) - Lag time in milliseconds before throttling kicks in. Default should typically suffice unless you support cpu-intensive operations.
• maxLag (default: 300) - The highest lag threshold which will block the greatest amount of traffic determined by maxBadHostThreshold or maxBadIpThreshold.
• minHostRate (default: 20) - Minimum rate if lag is >= maxLag. Disable rate limiting by setting to 0.
• maxHostRate (default: 40) - Maximum rate if lag is <= minLag.
• maxHostRatio (default: 0) - If a given host receives the specified threshold (0.1 = 10%) a hostViolation will be returned. This prevents a single host from accounting for excessive traffic and is an effective method for combating very large attacks.
• minIpRate (default: 0) - Minimum rate if lag is >= maxLag. Disable rate limiting by setting to 0.
• maxIpRate (default: 0) - Maximum rate if lag is <= minLag.
• maxIpRateHostViolation (default: 0) - Maximum rate if target host is currently exceeding the configured maxHostRatio. This can be used to increase IP throttling if a particular host is being targeted by a large number of IPs. Requests hitting this max rate will receive hostViolation while requests below the rate threshold but hitting the target host will not be flagged.
• subnetMaskBits (default: 24, allowed: 20–30) - CIDR prefix length used to derive subnet keys from IP addresses. /24 groups up to 256 IPs; /20 groups up to 4,096. /30 is the finest granularity (4 IPs per group). Useful for catching distributed attacks spread across many IPs in the same subnet.
• minSubnetRate (default: 0) - Minimum subnet request rate (req/s) before a subnet can be flagged. Setting to 0 disables subnet tracking entirely.
• maxSubnetRate (default: 0) - Maximum subnet rate (req/s) if lag is <= minLag. Disable by setting to 0.
• maxSubnetRateHostViolation (default: 0) - Maximum subnet rate (req/s) when the target host is exceeding maxHostRatio. When set, only subnets exceeding this rate against a violated host receive hostViolation; subnets below the threshold are not flagged. Disable by setting to 0.
• minSubnetUniqueIps (default: 0.5) - Minimum number of distinct source IPs within a subnet required before the subnet can be flagged as a bad actor. This prevents a single abusive IP from triggering a block for its entire subnet. Accepts two forms:
  • Fraction (0 < value < 1): resolved as a percentage of the subnet's host count. 0.5 on a /24 (256 hosts) requires at least 128 unique IPs.
  • Absolute (value >= 1): used as-is (rounded to the nearest integer). Set to 1 to restore legacy behaviour where any rate-exceeding subnet is immediately blocked. Requires subnet tracking to be enabled (minSubnetRate > 0). Note: the gate only applies when a source IP is available. Direct calls such as isBadSubnet('1.2.3.0') without a request object have no IP context, so the gate is skipped and the subnet is evaluated by rate alone.
• subnetWhitelist Set<string>([]) - Subnet keys that are never flagged as bad actors. Key format matches the derived subnet key (e.g. 103.142.223.0 for a /24).
• errorStatusCode (default: 503) - The HTTP status code to return if the request has been throttled.
• errorResponseDelay (default: 0) - Number of milliseconds to delay sending an error response to bad actors. A value of 0 will result in the response being sent synchronously before returning from the middleware.
• historySize (default: 200) - The LRU history size to use in tracking bad actors. Hosts and IPs both get their own dedicated LRU.
• maxAge (default: 10000) - Time (in ms) before history is purged. 10 seconds is generally more than adequate to capture an accurate hit rate.
• hostWhitelist Set<string>(['localhost']) - If provided will never flag hosts as bad actors.
• ipWhitelist Set<string>([]) - If provided will never flag IPs as bad actors.
• httpBehindProxy (default: false) - x-forwarded-for header only supported if this option is set to true.
• httpsBehindProxy (default: false) - x-forwarded-for header only supported if this option is set to true.

Cluster-Wide Rate Limiting

connect-qos supports cluster-wide rate limiting via Redis for multi-node deployments behind a load balancer. Without cluster mode, each node tracks request rates independently, so a distributed attacker spreading requests across nodes evades per-node limits. With cluster mode enabled, all nodes share counts via a dedicated Redis/Valkey instance.

Enabling Cluster Mode

const Redis = require('ioredis');
const { ConnectQOS } = require('connect-qos');

const redisClient = new Redis.Cluster([{ host: 'redis.example.com', port: 6379 }]);

const qos = new ConnectQOS({
  minIpRate: 8,
  maxIpRate: 15,
  maxAge: 10000,
  cluster: {
    redis: { client: redisClient, keyPrefix: 'qos:' },
    syncIntervalMs: 2000,
    maxTrackedActors: 50000,
    clusterMaxIpRate: 50,
    clusterMaxSubnetRate: 200,
    clusterMaxHostRatio: 0.20,
    clusterMaxIpRateHostViolation: 10,
    clusterMaxSubnetRateHostViolation: 30,
    onSync: (stats) => console.log('Cluster sync:', stats),
    onError: (err) => console.error('Cluster sync error', err.message),
  }
});

process.on('SIGTERM', () => qos.destroy());

How It Works

• Zero hot-path latency: all throttle decisions use in-process state; Redis is never on the request path
• Background sync: every syncIntervalMs, accumulated per-actor hit deltas are published to Redis sorted sets and cluster-wide counts are read back
• Sliding window: rate is computed from the current + previous time buckets to avoid boundary spikes
• Graceful degradation: if Redis is unavailable, the previous blocklist remains active; the node falls back to per-node limiting until the next successful sync
• Delta retry: if publishing to Redis fails, deltas are merged back into the local accumulator for the next cycle — no counts are lost

Cluster Options

| Option | Default | Description | |---|---|---| | redis.client | required | ioredis Redis or Cluster instance | | redis.keyPrefix | 'qos:' | Key prefix for all cluster Redis keys | | syncIntervalMs | 2000 | How often to sync with Redis (ms) | | maxTrackedActors | 50000 | Max actors tracked per sorted set window; lowest-traffic actors are removed when exceeded | | clusterMaxIpRate | 0 (disabled) | Cluster-wide IP rate limit (req/s); 0 disables | | clusterMaxSubnetRate | 0 (disabled) | Cluster-wide /24 subnet rate limit (req/s) | | clusterMaxHostRatio | 0 (disabled) | Block host if it exceeds this fraction of total cluster traffic | | clusterMaxIpRateHostViolation | 0 (disabled) | Per-IP rate limit when the target host has a cluster violation. Requires local IP tracking to be enabled (minIpRate > 0); enforced via local Metrics history, so it has no effect when local tracking is disabled. | | clusterMaxSubnetRateHostViolation | 0 (disabled) | Per-subnet rate limit when the target host has a cluster violation. Requires local subnet tracking to be enabled (minSubnetRate > 0); same constraint as above. | | clusterMinSubnetUniqueIps | 0.5 | Minimum unique source IPs observed by any single node before a subnet can be cluster-blocked. Each node tracks and publishes its own observed count; Redis retains the maximum across nodes via ZADD GT (requires Redis ≥ 6.2 / Valkey ≥ 7.2). The gate uses sliding-window continuity: diversity seen in the previous window is carried forward proportionally (max(current, floor(prev × prevWeight))), so a transiently quiet node does not lose its tracking data at window boundaries. Accepts the same fraction/absolute semantics as the local minSubnetUniqueIps option: 0.5 resolves to 50 % of the subnet's host count (128 for /24); any value >= 1 is used as an absolute count. Set to 1 to disable the gate (any single rate-exceeding subnet is blocked). | | onSync | — | Callback after each sync cycle; receives ClusterSyncStats | | onError | — | Callback on Redis sync errors (non-fatal; node continues operating) |

Redis Eviction Safety

QOS sorted-set keys have no TTL. Under volatile-* eviction policies, only keys with TTLs are eligible for eviction — keys without TTLs are never evicted but also never safe: once maxmemory is exhausted, Redis returns OOM errors for write commands instead of evicting. To avoid OOM errors, either:

• Use a dedicated Redis/Valkey instance with enough memory that eviction is never needed (recommended), or
• Use maxmemory-policy allkeys-lru so Redis can evict the lowest-traffic QOS keys when memory pressure is high (graceful degradation: old windows are evicted first, which is acceptable behavior).

Lifecycle

Call qos.destroy() to stop the background sync interval. The interval is created with .unref() so it will not prevent the Node.js process from exiting, but explicit cleanup is still recommended for clean shutdown.

onError Callback

onError fires on both sync failures (publish/read pipeline errors) and background cleanup errors. These are non-fatal; the node continues operating with its previous blocklist. Log all onError calls at warn level.

Testing

Unit tests require no external dependencies:

npm test

Integration tests verify cluster mode against a real Redis instance. Start Redis locally, then:

npm run test:integration

The tests connect to redis://localhost:6379 by default. To use a different URL:

REDIS_URL=redis://myhost:6379 npm run test:integration

If Redis is unreachable the suite fails immediately with a clear error rather than hanging.

Performance

With quality of service being the entire purpose of this library needless to say performance is a critical influence in every decision. You can expect connect-qos to never be a bottleneck of any kind. Local tests on modest laptop easily exceeds 3.5K req/sec on a hello world http server. See for yourself with npm run bench.