@peerbit/shared-log 13.2.24 → 13.2.26

package/src/sync/sync-peer-state.ts

@@ -0,0 +1,107 @@
+// Per-peer sync slot state shared by the simple and rateless synchronizers
+// (stage-4 sync unification). One row per peer carries the response/lookup
+// slot accounting whose release closures may settle long after the peer is
+// gone; both synchronizers use the same registry and the same
+// capture-the-row identity pattern.
+//
+// Per-field lifetime policy (review this table when adding fields):
+//
+// | field                        | created            | cleared                    |
+// |------------------------------|--------------------|----------------------------|
+// | lookups / responses / active | first slot acquire | physical release settles;  |
+// |                              |                    | row DETACH only removes    |
+// |                              |                    | current peer-generation    |
+// |                              |                    | quota (global quota stays  |
+// |                              |                    | charged until settlement)  |
+// | pendingResponseHashes        | consume accepts an | lease release settles      |
+// | (simple synchronizer only)   | authorization      | while attached; row        |
+// |                              | (custody moves     | DETACHES at peer           |
+// |                              | batch -> row)      | disconnect (budget returns |
+// |                              |                    | in aggregate)              |
+// | activeReleases               | active-response or | logical release settles    |
+// |                              | coordinate         | (self-removal); detach     |
+// |                              | dispatch starts    | invokes outstanding        |
+// |                              |                    | logical releases so        |
+// |                              |                    | retained work drains       |
+//
+// Global physical-work accounting deliberately SURVIVES both close()/open()
+// generation rotation and peer disconnect: the underlying storage resolvers
+// and transport sends are not universally abortable, so cross-generation
+// work stays charged until it settles. Disconnect only detaches the row from
+// the current peer generation and releases logical lifecycle ownership.
+//
+// Deliberately host-side this stage (stage-5 folds them into the host peer
+// sessions): dispatch epochs and dispatch-target sets (benchmark-gated hot
+// path), pending-claim rows (owned by the pending-sync record store),
+// admission per-peer indexes (owned by the record store's reservation
+// objects) and the recently-sent exchange-head dedupe rows.
+
+export type SyncPeerSlotRow = {
+	peer: string;
+	attached: boolean;
+	lookups: number;
+	responses: number;
+	active: number;
+	// Portion of the simple synchronizer's global pending-response hash budget
+	// held by this peer's ACTIVE (consumed) authorizations. Batch-resident
+	// hashes stay charged against the batch; consume moves custody here so a
+	// disconnect can return the budget even when the response ship never
+	// settles. Always 0 on rateless rows.
+	pendingResponseHashes: number;
+	// Outstanding disconnect-time LOGICAL releases. Both synchronizers put
+	// active-response lifecycle cleanup here; Simple also puts coordinate
+	// dispatch completion here. Detach invokes the remainder so retained work
+	// and dispatch listeners drain. Each eventual async completion separately
+	// returns its global physical permit exactly once.
+	activeReleases: Set<() => void>;
+};
+
+export class SyncPeerSlotRegistry {
+	readonly rows = new Map<string, SyncPeerSlotRow>();
+
+	ensure(peer: string): SyncPeerSlotRow {
+		let row = this.rows.get(peer);
+		if (!row) {
+			row = {
+				peer,
+				attached: true,
+				lookups: 0,
+				responses: 0,
+				active: 0,
+				pendingResponseHashes: 0,
+				activeReleases: new Set(),
+			};
+			this.rows.set(peer, row);
+		}
+		return row;
+	}
+
+	// Detaches and returns the peer's row (or undefined when the peer holds no
+	// slots). The caller drops only current-generation/per-peer accounting and
+	// logical ownership. Physical release closures retain this row by identity
+	// and return global quota only when the underlying work actually settles.
+	detach(peer: string): SyncPeerSlotRow | undefined {
+		const row = this.rows.get(peer);
+		if (!row) {
+			return undefined;
+		}
+		row.attached = false;
+		this.rows.delete(peer);
+		return row;
+	}
+
+	// Rows whose slots all settled while attached are dropped eagerly so the
+	// registry never accumulates idle peers between disconnects.
+	maybeDropIdle(row: SyncPeerSlotRow): void {
+		if (
+			row.attached &&
+			row.lookups === 0 &&
+			row.responses === 0 &&
+			row.active === 0 &&
+			row.pendingResponseHashes === 0 &&
+			row.activeReleases.size === 0
+		) {
+			this.rows.delete(row.peer);
+		}
+	}
+}