@mailwoman/corpus 9.4.0 → 10.1.0

package/lib/tools/fetch/download/network.ts

@@ -0,0 +1,250 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Download and manifest utilities for `mailwoman corpus fetch <source>`.
+ */
+
+import { openWriteStream, pipeline, Readable } from "@mailwoman/core/fs/streams"
+import { writeLocalFile } from "@mailwoman/core/fs/writers"
+import { sleep } from "@mailwoman/core/utils/sleep"
+import type { PathBuilder, PathBuilderLike } from "path-ts"
+
+/**
+ * Rate limited — retryable, the server is asking us to back off.
+ */
+const HTTP_TOO_MANY_REQUESTS = 429
+
+/**
+ * Lowest 5xx status.
+ *
+ * Server-side failures are retryable. 4xx are not.
+ */
+const HTTP_SERVER_ERROR_MIN = 500
+
+/**
+ * Highest 5xx status.
+ */
+const HTTP_SERVER_ERROR_MAX = 599
+
+/**
+ * The option base every `mailwoman corpus fetch <source>` module extends.
+ */
+/**
+ * How long a failed transfer waits before the next attempt.
+ */
+export const DEFAULT_RETRY_DELAY_MS = 5000
+
+export interface BaseFetchOptions {
+	/**
+	 * Destination root for downloaded source data.
+	 *
+	 * Each source writes its own subdirectory.
+	 */
+	outRoot: PathBuilder
+	/**
+	 * Pause between transfer retries, in milliseconds.
+	 *
+	 * Defaults to {@linkcode DEFAULT_RETRY_DELAY_MS}.
+	 *
+	 * A test that exercises the failure path pays this delay once per retry in real time.
+	 * Measured at 20.1 s for the two failing-transfer cases in `geonames-postal.test.ts`,
+	 * which is the whole cost of that file.
+	 * The retry count is the behavior under test there.
+	 *
+	 * The pause between attempts is not, so it is a caller's to shorten.
+	 */
+	retryDelayMs?: number
+}
+
+/**
+ * The per-run result every fetch module returns.
+ * The command maps `failed > 0` to exit code 1.
+ */
+export interface FetchSummary {
+	fetched: number
+	skipped: number
+	failed: number
+	failedCodes: string[]
+}
+
+/**
+ * The sibling `manifest.json` shape the single-file fetch modules write: origin URL +
+ * fetch timestamp + byte count + sha256, so downstream adapters can verify provenance.
+ */
+export interface SourceManifest {
+	source_url: string
+	downloaded_at: string
+	filename: string
+	sha256: string
+	bytes: number
+}
+
+/**
+ * A status worth retrying: rate limiting or a server-side failure.
+ */
+/**
+ * An HTTP failure that includes its status, so callers branch on `error.status`
+ * rather than on message prose.
+ *
+ * The prose route shipped a real flake: a caller classified "not published upstream"
+ * with `message.includes("404")`, and the message contains the URL.
+ * An ephemeral test-server port such as `:40453` satisfies it while the actual status is 500.
+ *
+ * Roughly 1–2% of ephemeral ports contain the substring.
+ * That collision can fail a CI run and vanish locally.
+ */
+export class HTTPStatusError extends Error {
+	readonly status: number
+
+	constructor(status: number, message: string) {
+		super(message)
+		this.name = "HTTPStatusError"
+		this.status = status
+	}
+}
+
+export function isTransientStatus(status: number): boolean {
+	return status === HTTP_TOO_MANY_REQUESTS || (status >= HTTP_SERVER_ERROR_MIN && status <= HTTP_SERVER_ERROR_MAX)
+}
+
+export interface DownloadOptions {
+	url: string
+	dest: PathBuilderLike
+	/**
+	 * Per-attempt timeout.
+	 *
+	 * Default 10 minutes — these are multi-GB government dumps.
+	 */
+	timeoutMs?: number
+	/**
+	 * Extra attempts after the first, taken only on transient statuses or network errors.
+	 *
+	 * Default 0.
+	 */
+	retries?: number
+	/**
+	 * Delay between attempts.
+	 *
+	 * Default 5s.
+	 */
+	retryDelayMs?: number
+	headers?: Record<string, string>
+	report?: (line: string) => void
+}
+
+/**
+ * Download `url` to `dest` with per-attempt timeout and transient-status retry.
+ *
+ * @throws On a non-transient http status or once retries are exhausted.
+ * Returns the byte count written.
+ */
+export async function downloadToFile(options: DownloadOptions): Promise<{ bytes: number }> {
+	const {
+		url,
+		dest,
+		timeoutMs = 600_000,
+		retries = 0,
+		retryDelayMs = DEFAULT_RETRY_DELAY_MS,
+		headers,
+		report,
+	} = options
+
+	let lastError: unknown
+
+	for (let attempt = 0; attempt <= retries; attempt++) {
+		if (attempt > 0) {
+			report?.(`retry ${attempt}/${retries} after ${retryDelayMs}ms — ${url}`)
+			await sleep(retryDelayMs)
+		}
+
+		let res: Response
+
+		try {
+			// Raw `fetch`, deliberately: this is the shared file downloader and the body is piped to disk below.
+			// `APIClient` is the repo default for API requests with small bodies and repeated calls.
+			// It buffers non-stream responses in memory.
+			// A multi-gigabyte transfer cannot use that path.
+			res = await fetch(url, { headers, signal: AbortSignal.timeout(timeoutMs) })
+		} catch (error) {
+			// AbortSignal timeouts and network-level failures are retryable.
+			lastError = error
+
+			continue
+		}
+
+		if (!res.ok) {
+			const error = new HTTPStatusError(res.status, `HTTP ${res.status} ${res.statusText} — ${url}`)
+
+			if (!isTransientStatus(res.status)) throw error
+			lastError = error
+
+			continue
+		}
+
+		try {
+			const buffer = Buffer.from(await res.arrayBuffer())
+			await writeLocalFile(buffer, dest)
+
+			return { bytes: buffer.byteLength }
+		} catch (error) {
+			// A mid-stream abort while reading the body is retryable too.
+			lastError = error
+		}
+	}
+
+	throw lastError instanceof Error ? lastError : new Error(String(lastError))
+}
+
+export interface StreamDownloadOptions {
+	headers?: Record<string, string>
+	timeoutMs: number
+	retries: number
+	retryDelayMs: number
+}
+
+/**
+ * Stream an http download to disk, returning the final http status (0 on network error after retries).
+ *
+ * Follows redirects (the Census and OpenAddresses endpoints both 302 to their real hosts).
+ *
+ * Kept separate from {@link downloadToFile} on purpose: this one streams a multi-GB body to disk
+ * (the buffered helper reads via `arrayBuffer()`) and returns the HTTP status instead of throwing.
+ * Per-file result collectors and two-URL fallback ladders consume that result.
+ */
+export async function streamDownload(url: string, dest: PathBuilderLike, opts: StreamDownloadOptions): Promise<number> {
+	for (let attempt = 0; attempt <= opts.retries; attempt++) {
+		try {
+			const res = await fetch(url, {
+				headers: opts.headers ?? {},
+				redirect: "follow",
+				signal: AbortSignal.timeout(opts.timeoutMs),
+			})
+
+			if (res.ok && res.body) {
+				await pipeline(Readable.fromWeb(res.body), openWriteStream(dest))
+
+				return res.status
+			}
+
+			if (attempt < opts.retries && isTransientStatus(res.status)) {
+				await sleep(opts.retryDelayMs)
+
+				continue
+			}
+
+			return res.status
+		} catch {
+			if (attempt < opts.retries) {
+				await sleep(opts.retryDelayMs)
+
+				continue
+			}
+
+			return 0
+		}
+	}
+
+	return 0
+}

package/lib/tools/fetch/download/retry.ts

@@ -0,0 +1,174 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ * @file Retries and resumes interrupted corpus downloads.
+ */
+
+import { tryStat } from "@mailwoman/core/fs/readers"
+import { openWriteStream, pipeline, Readable } from "@mailwoman/core/fs/streams"
+import { movePath } from "@mailwoman/core/fs/writers"
+import { sleep } from "@mailwoman/core/utils/sleep"
+import type { PathBuilderLike } from "path-ts"
+
+import { DEFAULT_RETRY_DELAY_MS, HTTPStatusError } from "#tools/fetch/download/network"
+
+/**
+ * The answer to a range that starts at or past the end of the body:
+ * the resume point is already the whole file.
+ */
+const HTTP_RANGE_NOT_SATISFIABLE = 416
+
+/**
+ * The answer to a satisfiable `Range` request.
+ *
+ * Anything else means the host ignored the range and would send the whole body again.
+ */
+const HTTP_PARTIAL_CONTENT = 206
+
+/**
+ * Run one transfer up to `1 + retries` times, pausing `retryDelayMs` between attempts.
+ *
+ * A portal that drops a long connection would otherwise lose a whole collection fetch to one file.
+ * The last error is rethrown so the caller can name the file it lost.
+ */
+export async function withRetries<T>(
+	transfer: () => Promise<T>,
+	options: { retries?: number; retryDelayMs?: number; report?: (line: string) => void; label?: string } = {}
+): Promise<T> {
+	const { retries = 3, retryDelayMs = DEFAULT_RETRY_DELAY_MS, report, label = "transfer" } = options
+	let lastError: unknown
+
+	for (let attempt = 0; attempt <= retries; attempt++) {
+		if (attempt > 0) {
+			report?.(`  retry ${attempt}/${retries} after ${retryDelayMs}ms — ${label}`)
+			await sleep(retryDelayMs)
+		}
+
+		try {
+			return await transfer()
+		} catch (error) {
+			lastError = error
+		}
+	}
+
+	throw lastError instanceof Error ? lastError : new Error(String(lastError))
+}
+
+async function bytesOnDisk(path: string): Promise<number> {
+	const stat = await tryStat(path)
+
+	return stat?.size ?? 0
+}
+
+/**
+ * Download `url` to `dest` in ranges, resuming from whatever the `.tmp` sibling already holds.
+ *
+ * The total comes from the first `Content-Range`, and the loop gives up after `maxConnections`
+ * connections so a host that keeps answering 206 with no bytes cannot spin.
+ *
+ * Answers the byte count written and throws when the host answers anything but 206 for a range.
+ */
+export async function resumableDownload(options: {
+	url: string
+	dest: PathBuilderLike
+	headers?: Record<string, string>
+	maxConnections?: number
+	retryDelayMs?: number
+	report?: (line: string) => void
+}): Promise<number> {
+	const { url, dest, headers = {}, maxConnections = 400, retryDelayMs = DEFAULT_RETRY_DELAY_MS, report } = options
+	const tmp = dest + ".tmp"
+	let have = await bytesOnDisk(tmp)
+	let total: number | undefined
+
+	for (let connection = 0; connection < maxConnections; connection++) {
+		if (total !== undefined && have >= total) break
+
+		let res: Response
+
+		try {
+			res = await fetch(url, { headers: { ...headers, range: `bytes=${have}-` }, signal: AbortSignal.timeout(600_000) })
+		} catch {
+			await sleep(retryDelayMs)
+
+			continue
+		}
+
+		if (res.status === HTTP_RANGE_NOT_SATISFIABLE && have > 0) {
+			// The file on disk is already the whole body when no bytes follow `have`.
+			// `Content-Range: bytes */<total>` gives its length.
+			const whole = /\*\/(\d+)/.exec(res.headers.get("content-range") ?? "")?.[1]
+			total = whole ? Number(whole) : have
+			await res.body?.cancel()
+
+			break
+		}
+
+		if (res.status !== HTTP_PARTIAL_CONTENT) {
+			throw new HTTPStatusError(res.status, `HTTP ${res.status} for a range request — ${url}`)
+		}
+
+		const range = /bytes \d+-\d+\/(\d+)/.exec(res.headers.get("content-range") ?? "")?.[1]
+
+		if (range) {
+			total = Number(range)
+		}
+
+		const before = have
+
+		try {
+			await pipeline(
+				Readable.fromWeb(res.body as Parameters<typeof Readable.fromWeb>[0]),
+				openWriteStream(tmp, { flags: "a" })
+			)
+		} catch {
+			// The connection dropped mid-body.
+			// The received bytes remain on disk.
+			// The next range starts after them.
+		}
+
+		have = await bytesOnDisk(tmp)
+
+		if (have === before) {
+			await sleep(retryDelayMs)
+		} else {
+			report?.(`  ${(have / 1024 / 1024).toFixed(1)} MB${total ? ` of ${(total / 1024 / 1024).toFixed(1)} MB` : ""}`)
+		}
+	}
+
+	if (total === undefined || have < total) {
+		throw new Error(`resumable download stalled at ${have} of ${total ?? "?"} bytes — ${url}`)
+	}
+
+	await movePath(tmp, dest)
+
+	return have
+}
+
+/**
+ * The `Set-Cookie` values of a response folded into one `Cookie` header,
+ * so a second request reuses the session the first opened.
+ */
+export function cookieHeader(res: Response): string {
+	return res.headers
+		.getSetCookie()
+		.map((cookie) => cookie.split(";")[0])
+		.join("; ")
+}
+
+/**
+ * The file name a `Content-Disposition: attachment` header names, RFC 5987 form (`filename*=utf-8''…`)
+ * first, plain `filename="…"` second, or `fallback` when the header includes neither.
+ */
+export function attachmentFilename(res: Response, fallback: string): string {
+	const disposition = res.headers.get("content-disposition") ?? ""
+	const extended = /filename\*=(?:utf-8|UTF-8)'[^']*'([^;]+)/.exec(disposition)?.[1]
+
+	if (extended) return decodeURIComponent(extended.trim())
+	const plain = /filename="?([^";]+)"?/.exec(disposition)?.[1]
+
+	if (plain) return decodeURIComponent(plain.trim())
+
+	return fallback
+}