sync.ts

1import type { Database } from "bun:sqlite";2import {3	clearMirror,4	knownEventIds,5	knownObject,6	knownReleases,7	knownSave,8	land,9	mirrorState,10	openMirror,11	type StoredObject,12	storeObjects,13} from "../mirror/store.ts";14import { hiddenPaths, recordSync, registerRepo } from "../server/db.ts";15import { RESERVED_NAMES } from "../server/reserved.ts";16import { dropTarballs } from "../tarball/cache.ts";17import type { Bucket } from "./reader.ts";18import {19	BucketError,20	checkFormat,21	type EventRecord,22	escapePath,23	isSafePath,24	listEventIds,25	listReleaseVersions,26	listRepos,27	OversizedObject,28	type ReleaseRecord,29	readEvent,30	readHead,31	readObject,32	readRelease,33	readReleasesLatest,34	readSave,35	type SaveRecord,36} from "./repo.ts";3738const INTERVAL_MS = 2 * 60 * 1000;39// Compressed bytes a sync holds before landing them; sized for a 2 GB host.40const LANDING_BYTES = 32 * 1024 * 1024;41const LANDING_SAVES = 100;4243export type SyncTarget = {44	bucket: Bucket;45	kepr: Database;46	mirrorDir: string;47	/** Called after each repo that synced without error. */48	synced?: (repo: string) => void;49};5051let running: Promise<void> | null = null;5253/** Syncs every repo in the bucket; the admin console's "sync now" calls this too. Never overlaps itself. */54export function syncAll(target: SyncTarget): Promise<void> {55	running ??= syncEach(target).finally(() => {56		running = null;57	});58	return running;59}6061/** Rebuilds one repo's mirror from the bucket, once any sync under way has finished. */62export async function rebuildRepo(target: SyncTarget, repo: string): Promise<void> {63	await syncAll(target);64	clearMirror(openMirror(target.mirrorDir, repo));65	await syncAll(target);66}6768/**69 * Applies a changed list of names ignored in every repo. Once any sync under way has finished (its70 * later batches would land with the old list), a longer list is applied to each mirror in place; a71 * name taken off needs its files back, so every mirror is rebuilt.72 */73export async function reapplyHidden(target: SyncTarget, rebuild: boolean): Promise<void> {74	const repos = await listRepos(target.bucket);75	await syncAll(target);76	for (const repo of repos) {77		const mirror = openMirror(target.mirrorDir, repo);78		const head = mirrorState(mirror, "head");79		if (rebuild) clearMirror(mirror);80		else if (head) {81			land(mirror, {82				head,83				releasesLatest: mirrorState(mirror, "releases_latest") ?? "",84				saves: [],85				objects: new Map(),86				events: [],87				releases: [],88				quarantined: [],89				hidden: hiddenPaths(target.kepr, repo),90			});91		}92	}93	if (rebuild) await syncAll(target);94}9596export function startSyncLoop(target: SyncTarget): void {97	const tick = () => syncAll(target).catch((error) => console.error(`kepr: sync failed: ${error}`));98	tick();99	setInterval(tick, INTERVAL_MS);100}101102async function syncEach(target: SyncTarget): Promise<void> {103	for (const repo of await listRepos(target.bucket)) {104		registerRepo(target.kepr, repo);105		let error: string | null = null;106		if (RESERVED_NAMES.has(repo)) {107			error = `not synced: the name "${repo}" clashes with Kepr's own /${repo} route; rename the repo`;108		} else {109			try {110				const mirror = openMirror(target.mirrorDir, repo);111				const events = await syncRepo(target.bucket, mirror, repo, hiddenPaths(target.kepr, repo));112				if (events.some((event) => event.kind === "strip")) await dropTarballs(target.bucket, repo);113			} catch (cause) {114				error = cause instanceof Error ? cause.message : String(cause);115			}116		}117		recordSync(target.kepr, repo, Math.floor(Date.now() / 1000), error);118		if (error === null) target.synced?.(repo);119	}120}121122/**123 * Koh's steal flow (§9.4) against a mirror: walk back from the remote head to a save the mirror has,124 * fetch what is missing, and land it, in batches oldest first. Events are listed every time, since a125 * flag or a strip arrives without moving `head`; releases only when `head` or `releases/latest`126 * moved (§18), since publishing one moves both. A path §5.5 refuses, or one whose object inflates127 * past Koh's file limit, is quarantined: the save lands without it, as if stripped. Returns the128 * events it landed.129 */130export async function syncRepo(131	bucket: Bucket,132	mirror: Database,133	repo: string,134	hidden: string[] = [],135	landingBytes = LANDING_BYTES,136): Promise<EventRecord[]> {137	await checkFormat(bucket, repo);138	const head = await readHead(bucket, repo);139	if (head === null) return [];140	const unknown = new Set<string>();141	for (let cursor = head; cursor !== "" && !knownSave(mirror, cursor); ) {142		if (unknown.has(cursor)) throw new BucketError(`${repo}: save ${cursor} is its own ancestor`);143		unknown.add(cursor);144		cursor = (await readSave(bucket, repo, cursor)).parent;145	}146	// Oldest first, landing whenever the batch passes `landingBytes` of objects or LANDING_SAVES147	// records: a first sync of a big repo never holds its whole history in memory, and one cut short148	// resumes from the last batch. Records are read again here rather than kept from the walk, since149	// each lists every file of its tree.150	let saves: SaveRecord[] = [];151	let objects = new Map<string, StoredObject>();152	let quarantined: string[] = [];153	let fetched = 0;154	const oversized = new Set<string>();155	const landBatch = (batchHead: string) => {156		land(mirror, {157			head: batchHead,158			releasesLatest: mirrorState(mirror, "releases_latest") ?? "",159			saves,160			objects,161			events: [],162			releases: [],163			quarantined,164			hidden,165		});166		saves = [];167		objects = new Map();168		quarantined = [];169		fetched = 0;170	};171	for (const id of [...unknown].reverse()) {172		const save = await readSave(bucket, repo, id);173		quarantined.push(...save.quarantined);174		// One save can hold more than a batch, as a first import does: its objects are stored ahead175		// of it, after the saves before it have landed, so no landing's cleanup sees them unreferenced.176		const ahead = new Map<string, StoredObject>();177		for (const file of save.files) {178			const { hash } = file;179			if (180				objects.has(hash) ||181				ahead.has(hash) ||182				oversized.has(hash) ||183				knownObject(mirror, hash)184			) {185				continue;186			}187			try {188				const { compressed, bytes } = await readObject(bucket, repo, hash);189				ahead.set(hash, { compressed, size: bytes.length });190				fetched += compressed.length;191			} catch (cause) {192				if (!(cause instanceof OversizedObject)) throw cause;193				oversized.add(hash);194			}195			if (fetched >= landingBytes) {196				const last = saves.at(-1);197				if (last) landBatch(last.id);198				storeObjects(mirror, ahead);199				ahead.clear();200				fetched = 0;201			}202		}203		for (const [hash, object] of ahead) objects.set(hash, object);204		quarantined.push(205			...save.files.filter((f) => oversized.has(f.hash)).map((f) => escapePath(f.path)),206		);207		save.files = save.files.filter((file) => !oversized.has(file.hash));208		saves.push(save);209		if (saves.length >= LANDING_SAVES && save.id !== head) landBatch(save.id);210	}211	const known = knownEventIds(mirror);212	const events: EventRecord[] = [];213	for (const id of await listEventIds(bucket, repo)) {214		if (known.has(id)) continue;215		const event = await readEvent(bucket, repo, id);216		// Koh ignores it too: no save can hold an illegal path (§5.5).217		if (event.kind !== "strip" || isSafePath(event.path)) events.push(event);218	}219	const releasesLatest = (await readReleasesLatest(bucket, repo)) ?? "";220	const moved =221		mirrorState(mirror, "head") !== head ||222		mirrorState(mirror, "releases_latest") !== releasesLatest;223	const releases = moved ? await newReleases(bucket, mirror, repo) : [];224	if (!moved && events.length === 0) return [];225	land(mirror, {226		head,227		releasesLatest,228		saves,229		objects,230		events,231		releases,232		quarantined,233		hidden,234	});235	return events;236}237238async function newReleases(239	bucket: Bucket,240	mirror: Database,241	repo: string,242): Promise<ReleaseRecord[]> {243	const known = knownReleases(mirror);244	const releases: ReleaseRecord[] = [];245	for (const version of await listReleaseVersions(bucket, repo)) {246		if (!known.has(version)) releases.push(await readRelease(bucket, repo, version));247	}248	return releases;249}