@optimystic/db-core 0.13.4 → 0.14.0

package/src/cohort-topic/antidos/replay-guard.ts

@@ -0,0 +1,109 @@
+/**
+ * Cohort-topic substrate — correlation-id replay guard (anti-DoS).
+ *
+ * Transcribed from `docs/cohort-topic.md` §Anti-DoS bullet 3. Every `RegisterV1` carries a
+ * `correlationId` (16 random bytes) and a signature over `(topicId, tier, correlationId, timestamp)`.
+ * The signature (verified elsewhere — the participant-key check) proves authorship; this guard adds
+ * **freshness**: it drops a registration whose `timestamp` is outside the accepted skew window
+ * (stale, or implausibly far in the future) and drops a `correlationId` it has already seen inside the
+ * window (a captured-and-replayed registration).
+ *
+ * Because a stale registration is rejected outright, the guard only needs to remember correlation ids
+ * for one `maxAgeMs` window: an id older than that would be rejected on timestamp alone, so its record
+ * can be pruned. Pruning runs on access, bounding memory to the live window's worth of registrations.
+ */
+
+import { bytesKey } from "../registration/bytes.js";
+import { createLogger } from "../../logger.js";
+
+const log = createLogger("cohort-topic:antidos");
+
+/** Default acceptance window: a registration timestamped older than this (relative to `now`) is stale. */
+export const DEFAULT_REPLAY_MAX_AGE_MS = 60_000;
+/** Default tolerated forward clock skew: a timestamp this far past `now` is rejected as implausible. */
+export const DEFAULT_REPLAY_MAX_FUTURE_SKEW_MS = 5_000;
+
+export interface CorrelationReplayGuardConfig {
+	/** A timestamp older than `now − maxAgeMs` is stale. Default {@link DEFAULT_REPLAY_MAX_AGE_MS}. */
+	maxAgeMs?: number;
+	/** A timestamp newer than `now + maxFutureSkewMs` is rejected. Default {@link DEFAULT_REPLAY_MAX_FUTURE_SKEW_MS}. */
+	maxFutureSkewMs?: number;
+}
+
+/** Freshness + anti-replay gate over registration `correlationId`s and timestamps. */
+export interface CorrelationReplayGuard {
+	/**
+	 * Accept `correlationId` (from `peerId`, stamped `timestamp`) evaluated at `now`. Returns `false`
+	 * — and records nothing — when the timestamp is stale or implausibly future, or when this
+	 * `correlationId` was already accepted inside the window (replay). Returns `true` and remembers the
+	 * id on first sight of a fresh registration.
+	 */
+	accept(correlationId: Uint8Array, peerId: Uint8Array, timestamp: number, now: number): boolean;
+}
+
+/** A remembered acceptance, kept until its timestamp ages out of the window. */
+interface SeenEntry {
+	timestamp: number;
+	peer: string;
+}
+
+class WindowedReplayGuard implements CorrelationReplayGuard {
+	private readonly seen = new Map<string, SeenEntry>();
+	private readonly maxAgeMs: number;
+	private readonly maxFutureSkewMs: number;
+	/** `now` of the last prune, so pruning amortizes rather than scanning on every call. */
+	private lastPruneAt = -Infinity;
+
+	constructor(config: CorrelationReplayGuardConfig = {}) {
+		this.maxAgeMs = config.maxAgeMs ?? DEFAULT_REPLAY_MAX_AGE_MS;
+		if (!(this.maxAgeMs > 0)) {
+			throw new RangeError(`maxAgeMs must be > 0, got ${this.maxAgeMs}`);
+		}
+		this.maxFutureSkewMs = config.maxFutureSkewMs ?? DEFAULT_REPLAY_MAX_FUTURE_SKEW_MS;
+		if (!(this.maxFutureSkewMs >= 0)) {
+			throw new RangeError(`maxFutureSkewMs must be >= 0, got ${this.maxFutureSkewMs}`);
+		}
+	}
+
+	accept(correlationId: Uint8Array, peerId: Uint8Array, timestamp: number, now: number): boolean {
+		if (timestamp < now - this.maxAgeMs) {
+			log("replay-guard reject: stale timestamp age=%d > maxAge=%d", now - timestamp, this.maxAgeMs);
+			return false; // stale
+		}
+		if (timestamp > now + this.maxFutureSkewMs) {
+			log("replay-guard reject: future timestamp skew=%d > maxSkew=%d", timestamp - now, this.maxFutureSkewMs);
+			return false; // implausibly future
+		}
+		this.maybePrune(now);
+		const key = bytesKey(correlationId);
+		const prior = this.seen.get(key);
+		if (prior !== undefined) {
+			// Attribute the replay to the *original* accepter: the correlationId (not the peer id) is the
+			// anti-replay key, so a replayer that spoofs a different source id is still caught here.
+			log("replay-guard reject: replayed correlationId (first seen from peer=%s)", prior.peer);
+			return false; // replay
+		}
+		this.seen.set(key, { timestamp, peer: bytesKey(peerId) });
+		return true;
+	}
+
+	/** Forget ids whose timestamp has aged past the window — they would be rejected as stale anyway. */
+	private maybePrune(now: number): void {
+		// Amortize: prune at most once per window rather than scanning on every accept.
+		if (now - this.lastPruneAt < this.maxAgeMs) {
+			return;
+		}
+		this.lastPruneAt = now;
+		const cutoff = now - this.maxAgeMs;
+		for (const [key, entry] of this.seen) {
+			if (entry.timestamp < cutoff) {
+				this.seen.delete(key);
+			}
+		}
+	}
+}
+
+/** Build a {@link CorrelationReplayGuard} over the configured staleness window and forward skew. */
+export function createCorrelationReplayGuard(config: CorrelationReplayGuardConfig = {}): CorrelationReplayGuard {
+	return new WindowedReplayGuard(config);
+}

package/src/cohort-topic/antidos/topic-budget.ts

@@ -0,0 +1,145 @@
+/**
+ * Cohort-topic substrate — per-cohort topic budget (anti-DoS).
+ *
+ * Transcribed from `docs/cohort-topic.md` §Anti-DoS bullet 2. A cohort holds forwarder state for at
+ * most `topics_max` (default 2048) topics. Once the budget is full a request to instantiate a **new**
+ * topic is refused (the caller answers `UnwillingCohort`); topics already in the budget continue to
+ * serve. To keep room for genuinely active topics, the budget evicts cold ones: eviction order is
+ * **LRU by participant count**, with **zero-recent-registration topics dropped first** — a topic that
+ * carries no direct participants is the cheapest to shed and is shed before any populated topic.
+ *
+ * `admit` only ever evicts a *zero-participant* topic. If every resident topic still carries
+ * participants the budget is genuinely saturated and the new instantiation is refused — an attacker
+ * cannot, by flooding new topic ids, evict a topic that real participants are using. LRU ordering
+ * uses a monotonic access counter bumped on every `admit`/`touch`, so it is deterministic and needs
+ * no wall clock.
+ */
+
+import { bytesKey } from "../registration/bytes.js";
+import { createLogger } from "../../logger.js";
+
+const log = createLogger("cohort-topic:antidos");
+
+/** Default per-cohort forwarder-topic ceiling — `topics_max`. */
+export const DEFAULT_TOPICS_MAX = 2048;
+
+export interface TopicBudgetConfig {
+	/** Max topics with forwarder state. Default {@link DEFAULT_TOPICS_MAX}. */
+	topicsMax?: number;
+}
+
+/** Per-cohort cap on the number of topics with forwarder state, with LRU eviction of cold topics. */
+export interface TopicBudget {
+	/**
+	 * Admit `topicId` to the forwarder-state budget. Returns `true` if it is already resident or there
+	 * is (or can be freed) room; `false` when the budget is full of populated topics, in which case the
+	 * caller answers `UnwillingCohort` for the new instantiation. Admitting a *new* topic when full
+	 * evicts the coldest zero-participant resident first.
+	 */
+	admit(topicId: Uint8Array): boolean;
+	/**
+	 * Record the current direct-participant count for a resident topic (LRU bookkeeping). Updates the
+	 * eviction key and bumps the topic's recency; a no-op for a topic that is not resident.
+	 */
+	touch(topicId: Uint8Array, participantCount: number): void;
+	/** Whether `topicId` currently holds forwarder state. */
+	has(topicId: Uint8Array): boolean;
+	/**
+	 * The direct-participant count last reported for a resident `topicId` (the eviction key), or
+	 * `undefined` when the topic is not resident. A resident reporting `0` is cold/evictable. Read-only
+	 * introspection for tests/diagnostics — eviction policy reads the same field internally.
+	 */
+	participantCount(topicId: Uint8Array): number | undefined;
+	/** Number of resident topics. */
+	readonly size: number;
+}
+
+/** Per-resident eviction bookkeeping. */
+interface ResidentState {
+	/** Direct-participant count last reported via `touch` (eviction primary key; ascending = colder). */
+	participantCount: number;
+	/** Monotonic access sequence (LRU tiebreaker; lower = least-recently used). */
+	seq: number;
+}
+
+class LruTopicBudget implements TopicBudget {
+	private readonly residents = new Map<string, ResidentState>();
+	private readonly topicsMax: number;
+	private seqCounter = 0;
+
+	constructor(config: TopicBudgetConfig = {}) {
+		this.topicsMax = config.topicsMax ?? DEFAULT_TOPICS_MAX;
+		if (!Number.isInteger(this.topicsMax) || this.topicsMax <= 0) {
+			throw new RangeError(`topicsMax must be a positive integer, got ${this.topicsMax}`);
+		}
+	}
+
+	get size(): number {
+		return this.residents.size;
+	}
+
+	has(topicId: Uint8Array): boolean {
+		return this.residents.has(bytesKey(topicId));
+	}
+
+	participantCount(topicId: Uint8Array): number | undefined {
+		return this.residents.get(bytesKey(topicId))?.participantCount;
+	}
+
+	admit(topicId: Uint8Array): boolean {
+		const key = bytesKey(topicId);
+		const existing = this.residents.get(key);
+		if (existing !== undefined) {
+			existing.seq = ++this.seqCounter; // already resident — refresh recency, always allowed
+			return true;
+		}
+		if (this.residents.size < this.topicsMax) {
+			this.residents.set(key, { participantCount: 0, seq: ++this.seqCounter });
+			return true;
+		}
+		// Full: only a zero-participant resident may be evicted to make room for a new topic.
+		const victim = this.coldestEvictable();
+		if (victim === undefined) {
+			log("topic-budget full size=%d max=%d — refuse new topic", this.residents.size, this.topicsMax);
+			return false;
+		}
+		this.residents.delete(victim);
+		this.residents.set(key, { participantCount: 0, seq: ++this.seqCounter });
+		log("topic-budget evicted cold topic to admit new (size=%d max=%d)", this.residents.size, this.topicsMax);
+		return true;
+	}
+
+	touch(topicId: Uint8Array, participantCount: number): void {
+		const key = bytesKey(topicId);
+		const state = this.residents.get(key);
+		if (state === undefined) {
+			return;
+		}
+		state.participantCount = participantCount;
+		state.seq = ++this.seqCounter;
+	}
+
+	/**
+	 * The coldest evictable resident key: among zero-participant topics, the least-recently used.
+	 * `undefined` when every resident still carries participants (nothing may be evicted for a new topic).
+	 */
+	private coldestEvictable(): string | undefined {
+		let victim: string | undefined;
+		let victimSeq = Infinity;
+		for (const [key, state] of this.residents) {
+			if (state.participantCount > 0) {
+				continue; // populated topics continue — never evicted for a new instantiation
+			}
+			if (state.seq < victimSeq) {
+				victim = key;
+				victimSeq = state.seq;
+			}
+		}
+		return victim;
+	}
+}
+
+/** Build a {@link TopicBudget} over the configured `topics_max`. */
+export function createTopicBudget(config: TopicBudgetConfig = {}): TopicBudget {
+	return new LruTopicBudget(config);
+}

package/src/cohort-topic/antiflood/index.ts

@@ -0,0 +1,2 @@
+export * from "./jitter.js";
+export * from "./invariants.js";

package/src/cohort-topic/antiflood/invariants.ts

@@ -0,0 +1,108 @@
+/**
+ * Cohort-topic substrate — anti-flood structural invariants (centralized for the e2e suite).
+ *
+ * Transcribed from `docs/cohort-topic.md` §Anti-flood properties. Three of the five named floods are
+ * defended *by construction* in the walk / promotion machinery already; this module does not
+ * re-implement that machinery — it provides pure, dependency-free predicates over a recorded walk
+ * trace so the unit and e2e suites can assert the invariants hold on the real engine's behaviour,
+ * the same way `packages/substrate-simulator/src/walk-metrics.ts` instruments
+ * `outwardMovesArePromoted` / `unwillingRetriesRestartAtDMax` on the simulator side.
+ *
+ * The caller records one {@link WalkProbe} per probe RPC the {@link import("../walk.js").WalkEngine}
+ * issues (its `treeTier` and the reply `result`), grouping the probes of one
+ * {@link import("../walk.js").WalkEngine.register} call into a {@link WalkTrace}. The predicates then
+ * check the anti-flood discipline:
+ *
+ * - **claim 3 — no speculative outward probe:** {@link outwardMovesArePromoted}. The only tier
+ *   *increase* between consecutive probes is one taken immediately after a `promoted` reply.
+ * - **walk discipline — inward only on `no_state`:** {@link inwardStepsFollowNoState}. Every single
+ *   tier *decrease* is the response to a `no_state` (no inward move on any other reply).
+ * - **claim 4 — inward retry restarts at `d_max`:** {@link retriesRestartAtDMax}. A fresh walk (the
+ *   restart after an `unwilling_cohort` / cohort-level back-off) begins at `d_max`, never re-hitting
+ *   the declined coord.
+ * - **claim 5 — sticky promotion:** {@link DEFAULT_T_PROMOTE_STICKY_MS} re-exported as the canonical
+ *   sticky window; the promotion lifecycle (`promotion.ts`) enforces it and `promotion.spec.ts`
+ *   covers the no-flap behaviour. {@link stickyHolds} is the predicate the e2e suite uses to confirm
+ *   a freshly-promoted cohort refuses demotion within the window.
+ */
+
+import type { RegisterResult } from "../wire/types.js";
+import { DEFAULT_T_PROMOTE_STICKY_MS } from "../promotion.js";
+
+export { DEFAULT_T_PROMOTE_STICKY_MS };
+
+/** One probe RPC in a walk: the tier it was issued at and the reply result it drew. */
+export interface WalkProbe {
+	/** Walk position `d` the probe was issued at. */
+	readonly treeTier: number;
+	/** The cohort reply classification. */
+	readonly result: RegisterResult;
+}
+
+/** The ordered probe log of a single {@link import("../walk.js").WalkEngine.register} call. */
+export interface WalkTrace {
+	/** `d_max` the walk started from. */
+	readonly dMax: number;
+	/** Probes in issue order. */
+	readonly probes: readonly WalkProbe[];
+}
+
+/**
+ * Claim 3 (no speculative outward probe): every tier *increase* between consecutive probes is taken
+ * only in response to a `promoted` redirect on the preceding probe. Any outward move not preceded by
+ * `promoted` is a speculative deeper probe — the flood this invariant forbids.
+ */
+export function outwardMovesArePromoted(trace: WalkTrace): boolean {
+	const { probes } = trace;
+	for (let i = 1; i < probes.length; i++) {
+		const moveOutward = probes[i]!.treeTier > probes[i - 1]!.treeTier;
+		if (moveOutward && probes[i - 1]!.result !== "promoted") {
+			return false;
+		}
+	}
+	return true;
+}
+
+/**
+ * Walk discipline: every single inward step (`treeTier` decreased by exactly 1, or reset to 0 at the
+ * root cold-start) is the response to a `no_state`. The walk never moves toward the root on any other
+ * reply, so inward traffic is one-hop-per-`no_state`, not a storm.
+ */
+export function inwardStepsFollowNoState(trace: WalkTrace): boolean {
+	const { probes } = trace;
+	for (let i = 1; i < probes.length; i++) {
+		const moveInward = probes[i]!.treeTier < probes[i - 1]!.treeTier;
+		if (moveInward && probes[i - 1]!.result !== "no_state") {
+			return false;
+		}
+	}
+	return true;
+}
+
+/**
+ * Claim 4 (inward retry restarts at `d_max`): each fresh walk in a re-registration sequence starts at
+ * its `d_max`, so a participant that drew `unwilling_cohort` backs off in *time* and re-walks from the
+ * top rather than re-hitting the declined coord. `traces` are consecutive `register` calls by one
+ * participant; the first probe of each must sit at that walk's `d_max`.
+ */
+export function retriesRestartAtDMax(traces: readonly WalkTrace[]): boolean {
+	for (const trace of traces) {
+		const first = trace.probes[0];
+		if (first === undefined) {
+			continue;
+		}
+		if (first.treeTier !== trace.dMax) {
+			return false;
+		}
+	}
+	return true;
+}
+
+/**
+ * Claim 5 (sticky promotion): a cohort promoted at `promotedAt` must not be reconsidered for demotion
+ * before `promotedAt + stickyMs`. Returns whether `now` is still inside the sticky window — the e2e
+ * suite asserts the promotion lifecycle refuses demotion exactly while this is `true`.
+ */
+export function stickyHolds(promotedAt: number, now: number, stickyMs: number = DEFAULT_T_PROMOTE_STICKY_MS): boolean {
+	return now - promotedAt < stickyMs;
+}

package/src/cohort-topic/antiflood/jitter.ts

@@ -0,0 +1,117 @@
+/**
+ * Cohort-topic substrate — re-registration jitter (anti-flood claim 2).
+ *
+ * Transcribed from `docs/cohort-topic.md` §Anti-flood properties (claim 2: "Re-registration storm
+ * after cohort failure") and folded back from the simulator-validated
+ * `packages/substrate-simulator/src/walk.ts` (`rejoinStagger` / `rateLimitedStagger`). When a cohort
+ * fails, every participant it served re-registers at once; without staggering, the recovering or
+ * replacement cohort takes the whole wave in a single instant and is shoved straight past
+ * `cap_promote`. Spreading the wave over `T_rejoin_jitter` (default 30 s, widened with the observed
+ * FRET cohort-failure rate) bounds the recovering cohort's inbound rate to
+ * `cap_promote / T_rejoin_jitter`.
+ *
+ * Two staggering forms, both exposed here:
+ * - {@link RejoinJitter.scheduleRejoin} — a single participant draws a uniform offset over the
+ *   jitter window. Decorrelates one participant's retry from its peers (the production API: a
+ *   participant that loses its cohort calls this once). Random, so the bound holds *in expectation*.
+ * - {@link RejoinJitter.scheduleWave} — a whole wave of `count` rejoiners is placed at a fixed
+ *   interval `windowMs / cap_promote`, so **any** `windowMs`-long window contains at most
+ *   `cap_promote` arrivals *by construction*. This is the hard-bound form the e2e suite asserts
+ *   against; it widens the span past one window when `count > cap_promote` so the rate ceiling
+ *   `cap_promote / T_rejoin_jitter` is never exceeded.
+ *
+ * This module is FRET-free and side-effect-free apart from the injected RNG: the caller schedules a
+ * timer at the returned timestamp.
+ */
+
+import { createLogger } from "../../logger.js";
+
+const log = createLogger("cohort-topic:antiflood");
+
+/** Default jitter window `T_rejoin_jitter` (ms) — `docs/cohort-topic.md` §Configuration. */
+export const DEFAULT_T_REJOIN_JITTER_MS = 30_000;
+/** Default `cap_promote` — sets the inbound-rate ceiling `cap_promote / T_rejoin_jitter`. */
+export const DEFAULT_REJOIN_CAP_PROMOTE = 64;
+
+export interface RejoinJitterConfig {
+	/** Base jitter window `T_rejoin_jitter` (ms). Default {@link DEFAULT_T_REJOIN_JITTER_MS}. */
+	tRejoinJitterMs?: number;
+	/** Acceptance ceiling per window; the rate bound is `capPromote / T_rejoin_jitter`. Default 64. */
+	capPromote?: number;
+	/**
+	 * Multiplier applied to the base window to track the observed FRET cohort-failure rate (claim 2:
+	 * "scaled with cohort failure rate observed from FRET"). A higher failure rate → a wider window so
+	 * the larger expected re-registration wave still respects the rate ceiling. Default 1 (steady FRET).
+	 */
+	failureRateScale?: number;
+	/** Uniform RNG in `[0, 1)`. Injected for determinism in tests; defaults to `Math.random`. */
+	random?: () => number;
+}
+
+/** Staggers post-cohort-failure re-registration so the recovering cohort isn't stormed. */
+export interface RejoinJitter {
+	/**
+	 * Jittered re-registration timestamp for **one** participant: `now + ⌊U[0, windowMs)⌋`. Random,
+	 * so a wave of independent callers spreads roughly uniformly over the window.
+	 */
+	scheduleRejoin(now: number): number;
+	/**
+	 * Rate-bounded timestamps for a whole re-registration wave of `count` participants. Arrivals are
+	 * evenly spaced at `windowMs / capPromote`, so any `windowMs`-long sliding window holds at most
+	 * `capPromote` of them — the `cap_promote / T_rejoin_jitter` ceiling, exactly rather than in
+	 * expectation. Returns one timestamp per participant, ascending.
+	 */
+	scheduleWave(count: number, now: number): number[];
+	/** Effective jitter window (ms) = base `T_rejoin_jitter` × `failureRateScale`. */
+	readonly windowMs: number;
+	/** Per-window acceptance ceiling (`capPromote`); the inbound-rate bound's numerator. */
+	readonly capPromote: number;
+}
+
+class StaggeredRejoinJitter implements RejoinJitter {
+	readonly windowMs: number;
+	readonly capPromote: number;
+	private readonly random: () => number;
+
+	constructor(config: RejoinJitterConfig = {}) {
+		const base = config.tRejoinJitterMs ?? DEFAULT_T_REJOIN_JITTER_MS;
+		if (!(base > 0)) {
+			throw new RangeError(`tRejoinJitterMs must be > 0, got ${base}`);
+		}
+		const scale = config.failureRateScale ?? 1;
+		if (!(scale >= 1)) {
+			throw new RangeError(`failureRateScale must be >= 1, got ${scale}`);
+		}
+		this.capPromote = config.capPromote ?? DEFAULT_REJOIN_CAP_PROMOTE;
+		if (!Number.isInteger(this.capPromote) || this.capPromote <= 0) {
+			throw new RangeError(`capPromote must be a positive integer, got ${this.capPromote}`);
+		}
+		this.windowMs = base * scale;
+		this.random = config.random ?? ((): number => Math.random());
+	}
+
+	scheduleRejoin(now: number): number {
+		const offset = Math.floor(this.random() * this.windowMs);
+		return now + offset;
+	}
+
+	scheduleWave(count: number, now: number): number[] {
+		if (!Number.isInteger(count) || count < 0) {
+			throw new RangeError(`count must be a non-negative integer, got ${count}`);
+		}
+		// Even spacing at windowMs/capPromote guarantees ≤ capPromote arrivals in any windowMs window;
+		// for count > capPromote the wave necessarily spans more than one window, holding the rate.
+		const interval = this.windowMs / this.capPromote;
+		const out = new Array<number>(count);
+		for (let i = 0; i < count; i++) {
+			out[i] = now + Math.floor(i * interval);
+		}
+		log("scheduleWave count=%d window=%d cap=%d span=%d", count, this.windowMs, this.capPromote, count > 0 ? out[count - 1]! - now : 0);
+		return out;
+	}
+}
+
+/** Build a {@link RejoinJitter} over the configured window, rate ceiling, and (test-injectable) RNG. */
+export function createRejoinJitter(config: RejoinJitterConfig = {}): RejoinJitter {
+	return new StaggeredRejoinJitter(config);
+}