store.ts

1import { Database } from "bun:sqlite";2import { mkdirSync } from "node:fs";3import { join } from "node:path";4import { inflateSync } from "node:zlib";5import type { EventRecord, ReleaseRecord, SaveRecord } from "../bucket/repo.ts";6import { isTestPath } from "../describe/links.ts";78// Koh 3's schema (§5.1, `src/storage/storage.v`) minus `file_cache`, which only a working tree needs,9// plus `releases`: each version's release.json (§11.2) as read from the bucket, since that is bucket10// data too and pages and feeds need it on every request.11const SCHEMA_VERSION = 2;12const SCHEMA = `13CREATE TABLE objects (14	hash TEXT PRIMARY KEY,15	size INTEGER NOT NULL,16	data BLOB NOT NULL17);18CREATE TABLE saves (19	id            TEXT PRIMARY KEY,20	parent        TEXT REFERENCES saves(id),21	message       TEXT NOT NULL,22	created_at    INTEGER NOT NULL,23	machine       TEXT NOT NULL,24	is_checkpoint INTEGER NOT NULL DEFAULT 0,25	size_bytes    INTEGER NOT NULL DEFAULT 0,26	agent         TEXT,27	session       TEXT,28	restores      TEXT,29	replayed_from TEXT30);31CREATE TABLE save_files (32	save_id    TEXT NOT NULL REFERENCES saves(id),33	path       TEXT NOT NULL,34	hash       TEXT NOT NULL REFERENCES objects(hash),35	executable INTEGER NOT NULL DEFAULT 0,36	PRIMARY KEY (save_id, path)37);38CREATE TABLE events (39	id         TEXT PRIMARY KEY,40	kind       TEXT NOT NULL,41	save_id    TEXT,42	path       TEXT,43	reason     TEXT,44	created_at INTEGER NOT NULL,45	machine    TEXT NOT NULL46);47CREATE TABLE state (48	key   TEXT PRIMARY KEY,49	value TEXT NOT NULL50);51CREATE TABLE releases (52	version TEXT PRIMARY KEY,53	record  TEXT NOT NULL54);55CREATE INDEX saves_parent ON saves(parent);56CREATE INDEX save_files_hash ON save_files(hash);57`;5859/** An object as the bucket stores it: zlib bytes, already checked against its hash. */60export type StoredObject = { compressed: Uint8Array; size: number };6162/** What one sync brings in, landed in one transaction. Saves are oldest first. */63export type Landing = {64	head: string;65	releasesLatest: string;66	saves: SaveRecord[];67	objects: Map<string, StoredObject>;68	events: EventRecord[];69	releases: ReleaseRecord[];70	/** Escaped names of the paths these saves arrived without (§5.5, A3). */71	quarantined: string[];72};7374// A recount inflates every file at the newest save; a file this large is data, not lines of code.75const MAX_COUNTED = 8 * 1024 * 1024;7677// Display names for the header's language, by extension; data and docs formats never count.78const LANGUAGES: Record<string, string> = {79	ts: "TypeScript",80	tsx: "TypeScript",81	js: "JavaScript",82	mjs: "JavaScript",83	zig: "Zig",84	v: "V",85	rs: "Rust",86	go: "Go",87	py: "Python",88	c: "C",89	h: "C",90	cpp: "C++",91	cc: "C++",92	hpp: "C++",93	swift: "Swift",94	kt: "Kotlin",95	java: "Java",96	rb: "Ruby",97	lua: "Lua",98	sh: "Shell",99	nix: "Nix",100	css: "CSS",101	html: "HTML",102};103104// Vendored code is someone else's: counting Koh 3's bundled SQLite made it read as 290k lines of C.105const VENDORED = /(^|\/)(vendor|third_party|node_modules)\//;106// Bumped when what counts as a line changes, so every mirror recounts once when it opens.107const STATS_RULE = "source-only";108109const open = new Map<string, Database>();110111/** The mirror of one repo at `<dir>/<repo>.db`, created on first use and kept open. */112export function openMirror(dir: string, repo: string): Database {113	const path = join(dir, `${repo}.db`);114	const cached = open.get(path);115	if (cached) return cached;116	mkdirSync(dir, { recursive: true });117	const db = new Database(path, { strict: true });118	db.run("PRAGMA journal_mode = WAL");119	db.run("PRAGMA foreign_keys = ON");120	db.run("PRAGMA busy_timeout = 5000");121	const { user_version } = db.query<{ user_version: number }, []>("PRAGMA user_version").get() ?? {122		user_version: 0,123	};124	if (user_version !== 0 && user_version !== SCHEMA_VERSION) {125		db.close();126		throw new Error(127			`${path} has mirror schema version ${user_version}; this Kepr reads version ${SCHEMA_VERSION}. ` +128				"The mirror is disposable: delete it and the next sync rebuilds it from the bucket.",129		);130	}131	if (user_version === 0) {132		db.transaction(() => {133			db.run(SCHEMA);134			db.run(`PRAGMA user_version = ${SCHEMA_VERSION}`);135		})();136	}137	// Mirrors counted under an older rule are recounted now, not at the next save.138	if (mirrorState(db, "head") && mirrorState(db, "stats_rule") !== STATS_RULE) refreshStats(db);139	open.set(path, db);140	return db;141}142143export function mirrorState(db: Database, key: string): string | null {144	const row = db145		.query<{ value: string }, [string]>("SELECT value FROM state WHERE key = ?")146		.get(key);147	return row?.value ?? null;148}149150export function knownSave(db: Database, id: string): boolean {151	return db.query("SELECT 1 FROM saves WHERE id = ?").get(id) !== null;152}153154export function knownObject(db: Database, hash: string): boolean {155	return db.query("SELECT 1 FROM objects WHERE hash = ?").get(hash) !== null;156}157158export function knownEventIds(db: Database): Set<string> {159	const rows = db.query<{ id: string }, []>("SELECT id FROM events").all();160	return new Set(rows.map((row) => row.id));161}162163export function knownReleases(db: Database): Set<string> {164	const rows = db.query<{ version: string }, []>("SELECT version FROM releases").all();165	return new Set(rows.map((row) => row.version));166}167168/** Lands a sync like Koh's steal: objects, events, saves, then every strip over every save. */169export function land(db: Database, landing: Landing): void {170	const putObject = db.query("INSERT OR IGNORE INTO objects (hash, size, data) VALUES (?, ?, ?)");171	const putEvent = db.query(172		"INSERT OR IGNORE INTO events (id, kind, save_id, path, reason, created_at, machine) VALUES " +173			"(?, ?, ?, ?, ?, ?, ?)",174	);175	const putSave = db.query(176		`INSERT OR IGNORE INTO saves (id, parent, message, created_at, machine, is_checkpoint, size_bytes,177			agent, session, restores, replayed_from) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,178	);179	const putFile = db.query(180		"INSERT OR IGNORE INTO save_files (save_id, path, hash, executable) VALUES (?, ?, ?, ?)",181	);182	const putRelease = db.query("INSERT OR IGNORE INTO releases (version, record) VALUES (?, ?)");183	const setState = db.query("INSERT OR REPLACE INTO state (key, value) VALUES (?, ?)");184	db.transaction(() => {185		for (const [hash, object] of landing.objects)186			putObject.run(hash, object.size, object.compressed);187		for (const e of landing.events) {188			putEvent.run(e.id, e.kind, e.save_id, e.path, e.reason, e.created_at, e.machine);189		}190		for (const s of landing.saves) {191			putSave.run(192				s.id,193				s.parent === "" ? null : s.parent,194				s.message,195				s.created_at,196				s.machine,197				s.is_checkpoint ? 1 : 0,198				s.size_bytes,199				s.agent,200				s.session,201				s.restores,202				s.replayed_from,203			);204			for (const f of s.files) putFile.run(s.id, f.path, f.hash, f.executable ? 1 : 0);205		}206		for (const r of landing.releases) putRelease.run(r.version, JSON.stringify(r));207		applyStrips(db);208		setState.run("head", landing.head);209		setState.run("releases_latest", landing.releasesLatest);210		const quarantined = new Set([...quarantinedPaths(db), ...landing.quarantined]);211		setState.run("quarantined", JSON.stringify([...quarantined].sort()));212	})();213	// Outside the transaction: the recount can inflate megabytes and needs no lock to be correct.214	refreshStats(db);215}216217/** Escaped names of every path this mirror's saves arrived without. */218export function quarantinedPaths(db: Database): string[] {219	return JSON.parse(mirrorState(db, "quarantined") ?? "[]");220}221222// §10.2: a strip of one save reaches that save; a strip of every save bans its path from all of223// history, whenever a save was made or arrives. Re-applying every strip is harmless, which is how a224// save that arrives after its strip event is covered too.225function applyStrips(db: Database): void {226	const strips = db227		.query<{ save_id: string; path: string }, []>(228			"SELECT save_id, path FROM events WHERE kind = 'strip'",229		)230		.all();231	if (strips.length === 0) return;232	const remove = db.query(233		`DELETE FROM save_files234		WHERE (path = $path OR substr(path, 1, length($dir)) = $dir)235		AND ($save = '' OR save_id = $save)`,236	);237	for (const strip of strips) {238		// Koh compares the prefix exactly, never with LIKE, so `_` and `%` in a path match only themselves.239		const dir = `${strip.path.replace(/\/+$/, "")}/`;240		remove.run({ path: strip.path, dir, save: strip.save_id });241	}242	db.run("DELETE FROM objects WHERE hash NOT IN (SELECT hash FROM save_files)");243}244245// Lines and languages never change between syncs, so they are counted once here, not per view. They246// come from the newest real save, not `head`: a checkpoint's files are not public history (§21.4).247// Lines are source lines: files in a known language, tests left out; docs and data never count.248function refreshStats(db: Database): void {249	const save =250		db251			.query<{ id: string }, []>(252				"SELECT id FROM saves WHERE is_checkpoint = 0 ORDER BY rowid DESC LIMIT 1",253			)254			.get()?.id ?? "";255	const files = db256		.query<{ path: string; size: number; data: Uint8Array }, [string]>(257			"SELECT f.path, o.size, o.data FROM save_files f JOIN objects o ON o.hash = f.hash WHERE f.save_id = ?",258		)259		.all(save);260	let lines = 0;261	const byLanguage = new Map<string, number>();262	for (const file of files) {263		const language = LANGUAGES[file.path.split(".").pop() ?? ""];264		// Tests are not the program: counted, Koh 3's read as 18k lines of V where it has 11k.265		if (!language || VENDORED.test(file.path) || isTestPath(file.path)) continue;266		if (file.size > MAX_COUNTED) continue;267		const bytes = inflateSync(file.data);268		if (bytes.includes(0)) continue;269		let count = 0;270		for (const byte of bytes) if (byte === 10) count++;271		if (bytes.length > 0 && bytes[bytes.length - 1] !== 10) count++;272		lines += count;273		byLanguage.set(language, (byLanguage.get(language) ?? 0) + count);274	}275	const languages = [...byLanguage].sort((a, b) => b[1] - a[1]);276	const setState = db.query("INSERT OR REPLACE INTO state (key, value) VALUES (?, ?)");277	setState.run("lines", String(lines));278	setState.run("language", languages[0]?.[0] ?? "");279	setState.run("languages", JSON.stringify(languages));280	setState.run("stats_save", save);281	setState.run("stats_rule", STATS_RULE);282}