#289 stream downloads rather than staging to temp disk

This commit is contained in:
Alan Grainger
2026-09-04 11:20:45 +12:00
parent 02b30e0176
commit 9c72829c4f
7 changed files with 428 additions and 446 deletions
-1
View File
@@ -8,7 +8,6 @@
"maxZoomQuality": "preview",
"downloadedFilename": 0,
"allowDownload": 0,
"downloadFromImmichConcurrencyLimit": 20,
"allowSlugLinks": true,
"showHomePage": true,
"showMetadata": {
+3 -6
View File
@@ -16,7 +16,7 @@ import {
import { buildAssetMetadata } from './gallery/metadata'
import crypto from 'crypto'
import { assetBuffer } from './stream/asset'
import { downloadAssets, sweepStaleStagingDirs } from './stream/download'
import { downloadAssets } from './stream/download'
import dayjs from 'dayjs'
import { NextFunction, Request, Response } from 'express-serve-static-core'
import { Asset, AssetType, ImageSize, KeyType, SharedLink } from './types'
@@ -342,8 +342,8 @@ process.on('uncaughtException', (err) => {
process.exit(1)
})
// Log-only: with asyncHandler routing request errors into errorHandler, a
// stray rejection from a background task (version check, staging-dir sweep,
// etc.) is not worth killing every in-flight request for.
// stray rejection from a background task (the version check, etc.) is not
// worth killing every in-flight request for.
process.on('unhandledRejection', (reason, promise) => {
console.error('Unhandled Rejection at:', promise, 'reason:', reason)
})
@@ -361,7 +361,4 @@ const server = app.listen(port, () => {
// than silently serving broken album shares. Unknown/unreachable is
// tolerated (logs a warning and continues) - see enforceMinimumImmichVersion.
enforceMinimumImmichVersion().catch(e => console.error('Immich version check failed:', e))
// Clean up any zip-download staging dirs left behind by a previous
// run that crashed before its finally block could run
sweepStaleStagingDirs().catch(e => console.error('sweepStaleStagingDirs failed:', e))
})
+149 -178
View File
@@ -1,22 +1,19 @@
import { assetFetchUrl, authHeadersForAsset } from '../immich'
import { Response } from 'express-serve-static-core'
import { Asset, SharedLink } from '../types'
import { getNumericConfigOption } from '../config/access'
import { log } from '../utils/log'
import archiver, { Archiver } from 'archiver'
import { sanitize } from '../utils/sanitize'
import { pipeline } from 'stream/promises'
import { promises as fs, createWriteStream } from 'fs'
import { tmpdir } from 'os'
import { join } from 'path'
import { resolveDownloadEndpoint, ImageEndpoint } from '../gallery/sizing'
import { title } from '../share'
import { getFilename } from '../gallery/filename'
import { createLimiter } from '../utils/limiter'
import { createIdleTimeoutStream } from '../utils/idleTimeoutStream'
import { readableFromWeb } from '../utils/webStream'
import { respondToInvalidRequest } from '../invalidRequestHandler'
const STAGING_DIR_PREFIX = 'ipp-zip-'
/** Attempts to get response headers from Immich for one asset before giving up. */
const MAX_ATTEMPTS = 3
/** How long to wait for Immich's response headers on each attempt. */
const HEADER_TIMEOUT_MS = 20_000
/**
* Download all assets in a share as a zip file.
@@ -25,173 +22,171 @@ export async function downloadAll (res: Response, share: SharedLink) {
await downloadAssets(res, share, share.assets)
}
/**
* Delete staging directories left over from a prior run that crashed before
* its `finally` block could clean up. Intended to run once at startup.
*
* A live download holds its staging dir open for the duration of the zip;
* `maxAgeMs` should comfortably exceed any realistic download time so we
* never delete a dir from a still-running download in another worker.
*/
export async function sweepStaleStagingDirs (maxAgeMs = 60 * 60 * 1000) {
const root = tmpdir()
const cutoff = Date.now() - maxAgeMs
const entries = await fs.readdir(root).catch(() => [] as string[])
for (const name of entries) {
if (!name.startsWith(STAGING_DIR_PREFIX)) continue
const path = join(root, name)
const stat = await fs.stat(path).catch(() => null)
if (!stat || stat.mtimeMs >= cutoff) continue
await fs.rm(path, { recursive: true, force: true }).catch(e => {
log.warn(`Failed to sweep stale staging dir ${path}: ${e instanceof Error ? e.message : String(e)}`)
})
}
}
type StagedAsset = { tempfile: string, asset: Asset, endpoint: ImageEndpoint, servedMime?: string }
type FetchedAsset = { response: globalThis.Response, asset: Asset, endpoint: ImageEndpoint, servedMime?: string, url: string }
type Failure = { asset: Asset, url: string, status?: number, error?: unknown }
type StageOutcome = StagedAsset | { failure: Failure } | null
type StagingOptions = {
stagingDir: string
concurrency: number
maxAttempts: number
headerTimeoutMs: number
idleTimeoutMs: number
}
type FetchOutcome = FetchedAsset | { failure: Failure } | null
/**
* Stream the given assets back as a zip file.
*
* Immich's own download service (server/src/services/download.service.ts)
* zips files directly from local disk - no HTTP, retries, or timeouts needed.
* We're a proxy fetching over HTTP, so the shape diverges:
* zips files straight off local disk. We're a proxy fetching over HTTP, so
* each asset is fetched from Immich and its body piped directly into the
* archive - nothing is staged to disk or held in memory, so a zip of any
* size needs only a few stream buffers #289.
*
* - Bounded concurrency against the upstream Immich server.
* - Two-phase fetch per asset: get headers (retried with linear backoff,
* bounded by a 20s header-receipt timeout), then stream the body to a
* temp file under an idle timeout that resets on every chunk -
* slow-but-steady downloads survive, truly stalled ones fail fast.
* - Stage everything to disk first so we can detect failure BEFORE any
* bytes hit the client. Once we've started streaming the zip we can't
* recover gracefully; aborting means killing the response socket and
* leaving a visibly broken download. This is the alternative to silently
* omitting failed assets - the user gets a clear "broken" signal instead
* of a zip quietly missing files.
* Assets are fetched strictly one at a time. Archiver writes entries
* serially anyway, and an unread response body would hold an idle
* connection open on Immich for as long as the current entry takes to reach
* the visitor - long enough for a reverse proxy to close it. The cost is
* Immich's per-request overhead between entries, which is negligible next
* to the visitor's link speed.
*
* The only thing we share with Immich here is using zip STORE (no
* compression), since photos and videos are already compressed.
* Failure handling:
* - Headers fail (after retries) before anything has been sent: an
* ordinary 404, so the visitor sees an error page rather than an empty
* download. The real reason goes to the server log.
* - Anything fails once the zip is on the wire: the socket is destroyed
* and the visitor gets a visibly broken download. That's deliberate; the
* alternative is a zip quietly missing files.
* - Visitor disconnects: the upstream fetch is aborted so Immich stops
* streaming, and the archive is torn down so it releases the in-flight
* entry #284.
*
* There is no body-level idle timeout: with the visitor as the sink, a slow
* or paused client stalls the flow just like a stalled upstream would, so a
* timer on the body can't tell them apart. The header timeout still bounds
* how long Immich may take to start answering.
*
* Zip entries use STORE (no compression), since photos and videos are
* already compressed.
*/
export async function downloadAssets (res: Response, share: SharedLink, assets: Asset[]) {
const archive = archiver('zip', { store: true })
// Without a listener, an archiver 'error' emission would crash the process.
archive.on('error', e => log(`Archiver error for share ${share.key}: ${e.message}`))
const controller = new AbortController()
let clientGone = false
let resolveClosed!: () => void
const resClosed = new Promise<void>(resolve => { resolveClosed = resolve })
const onClose = () => {
if (res.writableFinished) return
clientGone = true
controller.abort()
resolveClosed()
}
res.once('close', onClose)
if (res.closed) onClose()
// Headers and piping are deferred until the first asset has arrived, so a
// failure before then can still be answered with a normal error response.
let piped = false
for (const asset of assets) {
if (controller.signal.aborted) break
const fetched = await fetchOne(share, asset, controller.signal)
if (fetched === null) break // aborted while waiting on Immich
if ('failure' in fetched) {
if (!piped) {
archive.abort()
respondToInvalidRequest(res, 404, describeFailure(share, fetched.failure))
return
}
abortDownload(archive, res, share, fetched.failure)
return
}
if (!piped) {
startZipResponse(res, share, archive)
piped = true
}
const entry = await appendEntry(archive, fetched)
if (entry !== 'done') {
if (clientGone) break
controller.abort()
abortDownload(archive, res, share, { asset: fetched.asset, url: fetched.url, error: entry.error })
return
}
}
if (clientGone) {
log(`Zip download for share ${share.key} cancelled by client`)
teardownArchive(archive, res)
return
}
if (!piped) startZipResponse(res, share, archive) // empty selection: still a valid (empty) zip
// finalize() resolves when archiver has finished writing the zip output.
// Raced against client disconnect because finalize() never settles once
// the response is destroyed. The inline rejection handler also stops a
// late finalize failure becoming an unhandled rejection after a lost race.
const finished = archive.finalize().then(() => 'done' as const, () => 'error' as const)
const outcome = await Promise.race([finished, resClosed.then(() => 'closed' as const)])
if (outcome !== 'done') {
if (outcome === 'closed') log(`Zip download for share ${share.key} cancelled by client`)
// 'error' was already logged by the archiver error listener
teardownArchive(archive, res)
}
}
function startZipResponse (res: Response, share: SharedLink, archive: Archiver) {
res.setHeader('Content-Type', 'application/zip')
let filename = (sanitize(title(share)) || 'photos') + '.zip'
filename = encodeURI(filename)
res.setHeader('Content-Disposition', `attachment; filename*=UTF-8''${filename}`)
// Hint to intermediate proxies (Nginx, etc.) not to buffer this response.
res.setHeader('X-Accel-Buffering', 'no')
const archive = archiver('zip', { store: true })
// Without a listener, an archiver 'error' emission would crash the process.
archive.on('error', e => log(`Archiver error for share ${share.key}: ${e.message}`))
archive.pipe(res)
}
const options: StagingOptions = {
stagingDir: await fs.mkdtemp(join(tmpdir(), STAGING_DIR_PREFIX)),
concurrency: Math.max(1, getNumericConfigOption('ipp.downloadFromImmichConcurrencyLimit', 20)),
maxAttempts: 3,
headerTimeoutMs: 20_000,
idleTimeoutMs: 20_000
}
const controller = new AbortController()
let clientGone = false
let resolveClosed!: () => void
const resClosed = new Promise<void>(resolve => { resolveClosed = resolve })
res.once('close', () => {
if (res.writableFinished) return
clientGone = true
controller.abort()
resolveClosed()
/**
* Pipe one fetched body into the archive and wait for the entry to be fully
* written (archiver's 'entry' event fires after the data descriptor, which
* is later than the body's own end).
*
* On a body error compress-commons forwards the error to archiver's 'error'
* event and then moves on to the next queued entry, so the caller must abort
* the download itself. The body listener covers the tick between append()
* and archiver attaching its own handler, when an unhandled 'error' would
* otherwise crash the process.
*/
function appendEntry (archive: Archiver, fetched: FetchedAsset): Promise<'done' | { error: unknown }> {
return new Promise(resolve => {
if (!fetched.response.body) {
resolve({ error: new Error('Upstream response has no body') })
return
}
const body = readableFromWeb(fetched.response.body)
const cleanup = () => {
archive.off('entry', onEntry)
archive.off('error', onError)
body.off('error', onError)
}
const onEntry = () => { cleanup(); resolve('done') }
const onError = (error: unknown) => { cleanup(); resolve({ error }) }
archive.once('entry', onEntry)
archive.once('error', onError)
body.once('error', onError)
archive.append(body, { name: getFilename(fetched.asset, fetched.endpoint.servedSize, fetched.servedMime) })
})
try {
const stages = stageAssetsToDisk(share, assets, options, controller.signal)
const failure = await archiveStaged(archive, stages, controller)
if (clientGone) {
log(`Zip download for share ${share.key} cancelled by client`)
teardownArchive(archive, res)
return
}
if (failure) {
abortDownload(archive, res, share, failure)
return
}
// finalize() resolves when archiver has finished writing the zip output,
// which means every queued tempfile has been read. Safe to delete after.
// Raced against client disconnect because finalize() never settles once
// the response is destroyed. The inline rejection handler also stops a
// late finalize failure becoming an unhandled rejection after a lost race.
const finished = archive.finalize().then(() => 'done' as const, () => 'error' as const)
const outcome = await Promise.race([finished, resClosed.then(() => 'closed' as const)])
if (outcome !== 'done') {
if (outcome === 'closed') log(`Zip download for share ${share.key} cancelled by client`)
// 'error' was already logged by the archiver error listener
teardownArchive(archive, res)
}
} finally {
await fs.rm(options.stagingDir, { recursive: true, force: true }).catch(() => { /* best effort */ })
}
}
/**
* Kick off staging for every asset at once; the limiter caps how many run
* concurrently. Returns the promise array in input order so the consumer can
* archive in order.
*/
function stageAssetsToDisk (share: SharedLink, assets: Asset[], options: StagingOptions, signal: AbortSignal): Promise<StageOutcome>[] {
const limit = createLimiter(options.concurrency)
return assets.map((asset, index) => limit(() => stageOne(share, asset, index, options, signal)))
}
/**
* Walk the staged assets in input order, adding each to the archive as soon
* as it lands. The first failure aborts the download signal (so unstarted
* stages bail out and in-flight fetches are cancelled) and is returned to the
* caller. We wait for in-flight stages to settle before returning so the
* staging dir is safe to delete.
*/
async function archiveStaged (archive: Archiver, stages: Promise<StageOutcome>[], controller: AbortController): Promise<Failure | null> {
let failure: Failure | null = null
for (const stage of stages) {
const result = await stage
if (result === null) break // aborted by an earlier stage
if ('failure' in result) {
controller.abort()
failure = result.failure
break
}
// archive.file queues the entry; archiver lazily opens and reads it.
archive.file(result.tempfile, { name: getFilename(result.asset, result.endpoint.servedSize, result.servedMime) })
}
// After abort, wait for any in-flight stages to settle before we delete
// the staging dir. stageOne never rejects (errors become failure objects),
// so a plain Promise.all is safe.
if (controller.signal.aborted) await Promise.all(stages)
return failure
}
function abortDownload (archive: Archiver, res: Response, share: SharedLink, failure: Failure) {
function describeFailure (share: SharedLink, failure: Failure): string {
const detail = failure.status !== undefined
? `HTTP ${failure.status}`
: (failure.error instanceof Error ? failure.error.message : String(failure.error))
log(`Aborting zip download for share ${share.key}: failed to fetch asset ${failure.asset.id} from ${failure.url} (${detail})`)
return `Zip download for share ${share.key}: failed to fetch asset ${failure.asset.id} from ${failure.url} (${detail})`
}
function abortDownload (archive: Archiver, res: Response, share: SharedLink, failure: Failure) {
log('Aborting ' + describeFailure(share, failure))
teardownArchive(archive, res)
}
/**
* Tear down an aborted archive so it releases its resources. `abort()` alone
* is not enough; it kills the queue but leaves any in-flight entry paused,
* holding an open read fd on its tempfile forever #284
* holding its source stream open forever #284
*/
function teardownArchive (archive: Archiver, res: Response) {
archive.abort()
@@ -201,22 +196,20 @@ function teardownArchive (archive: Archiver, res: Response) {
}
/**
* Fetch one asset's upstream bytes into a temp file. Two phases:
* 1. fetchHeadersWithRetry - get the response headers, retried on
* transient failure.
* 2. streamBodyToTempFile - drain the body to disk under an idle timeout.
* Fetch one asset's response headers from Immich, retried on transient
* failure. The body is left unread for the caller to pipe.
*
* Returns the staged asset on success, a wrapped Failure on error, or null
* if we observed the abort flag and bailed without doing work.
* Returns the fetched asset on success, a wrapped Failure on error, or null
* if the download was aborted before we got an answer.
*/
async function stageOne (share: SharedLink, asset: Asset, index: number, options: StagingOptions, signal: AbortSignal): Promise<StageOutcome> {
async function fetchOne (share: SharedLink, asset: Asset, signal: AbortSignal): Promise<FetchOutcome> {
if (signal.aborted) return null
const endpoint = resolveDownloadEndpoint(asset, share.allowDownload !== false)
const url = assetFetchUrl(asset, endpoint.subpath, endpoint.sizeQueryParam)
const reqAuthHeaders = await authHeadersForAsset(asset)
const fetched = await fetchHeadersWithRetry(url, reqAuthHeaders, options.maxAttempts, options.headerTimeoutMs, signal, asset)
const fetched = await fetchHeadersWithRetry(url, reqAuthHeaders, MAX_ATTEMPTS, HEADER_TIMEOUT_MS, signal, asset)
if (fetched === null) return null
if ('failure' in fetched) return { failure: { ...fetched.failure, asset, url } }
@@ -224,7 +217,7 @@ async function stageOne (share: SharedLink, asset: Asset, index: number, options
// getFilename would fall back to an id-based name. The `/original` response
// carries the real name in Content-Disposition and the mime in Content-Type,
// so recover them from the headers we already fetched - no extra calls.
const stagedAsset = asset.originalFileName ? asset : enrichFromHeaders(asset, fetched.response)
const namedAsset = asset.originalFileName ? asset : enrichFromHeaders(asset, fetched.response)
// Playback-fallback downloads serve Immich's transcode; carry the response
// content-type so the zip entry's extension matches the actual bytes.
@@ -232,12 +225,7 @@ async function stageOne (share: SharedLink, asset: Asset, index: number, options
? (fetched.response.headers.get('content-type') || '').split(';')[0].trim() || undefined
: undefined
// Use the array index in the path so we never collide on duplicate IDs.
const tempfile = join(options.stagingDir, `${index}-${asset.id}`)
const streamed = await streamBodyToTempFile(fetched.response, tempfile, options.idleTimeoutMs)
if ('failure' in streamed) return { failure: { asset, url, error: streamed.failure } }
return { tempfile, asset: stagedAsset, endpoint, servedMime }
return { response: fetched.response, asset: namedAsset, endpoint, servedMime, url }
}
/**
@@ -290,8 +278,7 @@ type HeaderFetchOutcome =
* because the signal we pass to fetch stays bound to the response body - if
* the timeout fires after headers arrive but while the body is still
* streaming, the body read errors out. We clear the header timer as soon
* as we have a response; downstream the idle-timeout transform guards
* against stalled bodies.
* as we have a response.
*
* The download-wide `signal` is combined into the fetch signal so that an
* aborted download (asset failure or client disconnect) cancels the header
@@ -333,19 +320,3 @@ async function fetchHeadersWithRetry (
}
return { failure: { status: lastStatus, error: lastError } }
}
/**
* Stream `response.body` to `tempfile`, guarded by an idle timeout. The
* idle stream destroys itself if no chunk arrives within `idleMs`, killing
* the pipeline with a clear error.
*/
async function streamBodyToTempFile (response: globalThis.Response, tempfile: string, idleMs: number): Promise<{ ok: true } | { failure: unknown }> {
if (!response.body) return { failure: new Error('Upstream response has no body') }
const body = readableFromWeb(response.body)
try {
await pipeline(body, createIdleTimeoutStream(idleMs), createWriteStream(tempfile))
return { ok: true }
} catch (e) {
return { failure: e }
}
}
-192
View File
@@ -1,192 +0,0 @@
import { describe, it, expect, vi, beforeAll, afterAll, afterEach } from 'vitest'
import { promises as fs } from 'fs'
import { Writable } from 'stream'
import type { Response } from 'express-serve-static-core'
import { downloadAssets } from '../src/stream/download'
import { Asset, AssetType, KeyType, SharedLink } from '../src/types'
/*
Isolate this suite's staging dirs in a dedicated tmp dir so the
before/after assertions can't race other test files that also create
ipp-zip-* dirs under the real os.tmpdir().
*/
const { TEST_TMP } = vi.hoisted(() => {
return { TEST_TMP: (process.env.TMPDIR || '/tmp') + '/ipp-download-abort-test-' + process.pid }
})
vi.mock('os', async importOriginal => {
const orig = await importOriginal<typeof import('os')>()
return { ...orig, tmpdir: () => TEST_TMP }
})
beforeAll(async () => {
await fs.mkdir(TEST_TMP, { recursive: true })
})
afterAll(async () => {
await fs.rm(TEST_TMP, { recursive: true, force: true })
})
afterEach(() => {
vi.unstubAllGlobals()
})
async function stagingDirs (): Promise<string[]> {
const entries = await fs.readdir(TEST_TMP).catch(() => [] as string[])
return entries.filter(name => name.startsWith('ipp-zip-'))
}
/**
* Open fds pointing at deleted staging tempfiles. An aborted archiver used to
* keep the in-flight entry's read stream open forever; on tmpfs that pins the
* deleted file's space (issue #284). Linux-only introspection - returns []
* elsewhere, making the assertions no-ops.
*/
async function deletedStagingFds (): Promise<string[]> {
const fds = await fs.readdir('/proc/self/fd').catch(() => [] as string[])
const leaked: string[] = []
for (const fd of fds) {
const target = await fs.readlink('/proc/self/fd/' + fd).catch(() => '')
if (target.includes('ipp-zip-') && target.includes('(deleted)')) leaked.push(target)
}
return leaked
}
/** Wait for archiver's post-abort drain to release fds, then return leftovers. */
async function settledDeletedStagingFds (timeoutMs = 3000): Promise<string[]> {
const deadline = Date.now() + timeoutMs
let leaked = await deletedStagingFds()
while (leaked.length > 0 && Date.now() < deadline) {
await new Promise(resolve => setTimeout(resolve, 100))
leaked = await deletedStagingFds()
}
return leaked
}
function makeAsset (id: string): Asset {
return {
id,
key: 'testkey',
keyType: KeyType.key,
type: AssetType.image,
isTrashed: false,
originalFileName: id + '.jpg',
originalMimeType: 'image/jpeg'
}
}
const share: SharedLink = {
key: 'testkey',
keyType: KeyType.key,
type: 'ALBUM',
description: 'Test album',
assets: []
}
/*
Minimal stand-in for the Express response: a real Writable (so 'close',
writableFinished and destroy behave like the real thing) plus the header
setters downloadAssets calls.
*/
class FakeRes extends Writable {
received = 0
constructor () {
super()
// Node's http stack handles response stream errors internally; a bare
// Writable would crash the test process on destroy(err) instead.
this.on('error', () => { /* swallowed, like http.ServerResponse */ })
}
setHeader (_name: string, _value: string) { return this }
_write (chunk: Buffer, _enc: string, cb: (error?: Error | null) => void) {
this.received += chunk.length
cb()
}
}
function asResponse (res: FakeRes): Response {
return res as unknown as Response
}
/** Fetch stub that returns the full body immediately. */
function instantFetch (bytes: number) {
return vi.fn(async () => new globalThis.Response(new Uint8Array(bytes), { status: 200 }))
}
/**
* Fetch stub whose body sends one chunk then stays open forever, erroring
* only when the passed-in signal aborts - like a real streaming download
* that gets cancelled. Records each request's signal for assertions.
*/
function trickleFetch (signals: AbortSignal[]) {
return vi.fn(async (_url: unknown, init?: RequestInit) => {
const signal = init?.signal as AbortSignal
signals.push(signal)
const stream = new ReadableStream({
start (controller) {
controller.enqueue(new Uint8Array(1024))
signal.addEventListener('abort', () => {
try { controller.error(signal.reason ?? new Error('aborted')) } catch { /* already errored */ }
})
}
})
return new globalThis.Response(stream, { status: 200 })
})
}
describe('downloadAssets abort handling', () => {
it('streams a zip and cleans up the staging dir on the happy path', async () => {
vi.stubGlobal('fetch', instantFetch(2048))
const res = new FakeRes()
await downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2')])
expect(res.received).toBeGreaterThan(0)
expect(res.writableFinished).toBe(true)
expect(await stagingDirs()).toEqual([])
})
it('resolves, cleans up and cancels upstream fetches when the client aborts mid-staging', async () => {
const signals: AbortSignal[] = []
vi.stubGlobal('fetch', trickleFetch(signals))
const res = new FakeRes()
const done = downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2')])
// Give staging a moment to start, then drop the client connection
await new Promise(resolve => setTimeout(resolve, 50))
res.destroy(new Error('client aborted'))
await done
expect(signals.length).toBeGreaterThan(0)
expect(signals.every(s => s.aborted)).toBe(true)
expect(await stagingDirs()).toEqual([])
expect(await settledDeletedStagingFds()).toEqual([])
}, 10_000)
it('resolves and cleans up when the client aborts while the zip is streaming', async () => {
// Large enough that archiver can't flush everything into a destroyed
// response; without disconnect handling finalize() would hang forever.
vi.stubGlobal('fetch', instantFetch(5 * 1024 * 1024))
const res = new FakeRes()
const origWrite = res._write.bind(res)
let first = true
res._write = (chunk, enc, cb) => {
origWrite(chunk, enc, cb)
if (first) {
first = false
setImmediate(() => res.destroy(new Error('client aborted')))
}
}
await downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2')])
expect(await stagingDirs()).toEqual([])
// The regression from issue #284's follow-up: the staging dir was removed
// but archiver still held an open fd on a deleted tempfile.
expect(await settledDeletedStagingFds()).toEqual([])
}, 10_000)
it('cleans up and destroys the response when an upstream fetch keeps failing', async () => {
vi.stubGlobal('fetch', vi.fn(async () => new globalThis.Response('nope', { status: 500 })))
const res = new FakeRes()
await downloadAssets(asResponse(res), share, [makeAsset('a1')])
expect(res.destroyed).toBe(true)
expect(await stagingDirs()).toEqual([])
expect(await settledDeletedStagingFds()).toEqual([])
}, 10_000)
})
+276
View File
@@ -0,0 +1,276 @@
import { describe, it, expect, vi, afterEach } from 'vitest'
import { Writable } from 'stream'
import { once } from 'events'
import type { Response } from 'express-serve-static-core'
import { downloadAssets } from '../src/stream/download'
import { Asset, AssetType, KeyType, SharedLink } from '../src/types'
/*
"Download all" streams each asset from Immich straight into the zip on the
wire - no disk staging, no buffering of whole files (issue #289). These tests
cover the shape of that pipeline: backpressure, one upstream body at a time,
and the three ways it can end early (visitor leaves, Immich refuses, body
dies mid-stream).
*/
afterEach(() => {
vi.unstubAllGlobals()
vi.restoreAllMocks()
})
function makeAsset (id: string): Asset {
return {
id,
key: 'testkey',
keyType: KeyType.key,
type: AssetType.image,
isTrashed: false,
originalFileName: id + '.jpg',
originalMimeType: 'image/jpeg'
}
}
const share: SharedLink = {
key: 'testkey',
keyType: KeyType.key,
type: 'ALBUM',
description: 'Test album',
assets: []
}
/*
Minimal stand-in for the Express response: a real Writable (so 'close',
writableFinished and destroy behave like the real thing) plus the header
and status helpers downloadAssets and respondToInvalidRequest call.
*/
class FakeRes extends Writable {
received = 0
chunks: Buffer[] = []
statusCode = 200
headers: Record<string, string> = {}
/** Delay applied to every write, to simulate a slow visitor. */
writeDelayMs = 0
constructor () {
super()
// Node's http stack handles response stream errors internally; a bare
// Writable would crash the test process on destroy(err) instead.
this.on('error', () => { /* swallowed, like http.ServerResponse */ })
}
setHeader (name: string, value: string) { this.headers[name] = value; return this }
status (code: number) { this.statusCode = code; return this }
send () { this.end(); return this }
_write (chunk: Buffer, _enc: string, cb: (error?: Error | null) => void) {
this.received += chunk.length
this.chunks.push(chunk)
if (this.writeDelayMs > 0) setTimeout(cb, this.writeDelayMs)
else cb()
}
get output () { return Buffer.concat(this.chunks) }
}
function asResponse (res: FakeRes): Response {
return res as unknown as Response
}
/** Fetch stub that returns the full body immediately. */
function instantFetch (bytes: number) {
return vi.fn(async () => new globalThis.Response(new Uint8Array(bytes), { status: 200 }))
}
/**
* Fetch stub whose body sends one chunk then stays open forever, erroring
* only when the passed-in signal aborts - like a real streaming download
* that gets cancelled. Records each request's signal for assertions.
*/
function trickleFetch (signals: AbortSignal[]) {
return vi.fn(async (_url: unknown, init?: RequestInit) => {
const signal = init?.signal as AbortSignal
signals.push(signal)
const stream = new ReadableStream({
start (controller) {
controller.enqueue(new Uint8Array(1024))
signal.addEventListener('abort', () => {
try { controller.error(signal.reason ?? new Error('aborted')) } catch { /* already errored */ }
})
}
})
return new globalThis.Response(stream, { status: 200 })
})
}
/**
* Fetch stub with a pull-based body: chunks are only produced on demand, and
* `pulled` counts how many bytes have been handed over so far.
*/
function pullFetch (totalBytes: number, chunkBytes: number, counter: { pulled: number }) {
return vi.fn(async () => {
let sent = 0
const stream = new ReadableStream({
pull (controller) {
if (sent >= totalBytes) {
controller.close()
return
}
const size = Math.min(chunkBytes, totalBytes - sent)
controller.enqueue(new Uint8Array(size))
sent += size
counter.pulled = sent
}
})
return new globalThis.Response(stream, { status: 200 })
})
}
/**
* Parse the central directory of a zip buffer: one {name, size} per entry.
* Just enough of the format to check the output is a real zip with the
* expected entries; extractors read the central directory, so it's the part
* that matters.
*/
function centralDirectory (zip: Buffer): Array<{ name: string, size: number }> {
const entries: Array<{ name: string, size: number }> = []
let offset = zip.indexOf('PK\x01\x02', 0, 'binary')
while (offset !== -1) {
const size = zip.readUInt32LE(offset + 24)
const nameLength = zip.readUInt16LE(offset + 28)
const name = zip.subarray(offset + 46, offset + 46 + nameLength).toString('utf8')
entries.push({ name, size })
offset = zip.indexOf('PK\x01\x02', offset + 46 + nameLength, 'binary')
}
return entries
}
function countOf (zip: Buffer, signature: string): number {
let count = 0
let offset = zip.indexOf(signature, 0, 'binary')
while (offset !== -1) {
count++
offset = zip.indexOf(signature, offset + 1, 'binary')
}
return count
}
describe('downloadAssets', () => {
it('streams a valid zip with one entry per asset', async () => {
vi.stubGlobal('fetch', instantFetch(2048))
const res = new FakeRes()
await downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2')])
expect(res.writableFinished).toBe(true)
expect(res.headers['Content-Type']).toBe('application/zip')
const zip = res.output
expect(countOf(zip, 'PK\x03\x04')).toBe(2) // local file headers
expect(countOf(zip, 'PK\x05\x06')).toBe(1) // end of central directory
expect(centralDirectory(zip)).toEqual([
{ name: 'a1.jpg', size: 2048 },
{ name: 'a2.jpg', size: 2048 }
])
})
it('honours backpressure from a slow visitor instead of buffering the body', async () => {
const total = 32 * 1024 * 1024
const counter = { pulled: 0 }
vi.stubGlobal('fetch', pullFetch(total, 64 * 1024, counter))
const res = new FakeRes()
res.writeDelayMs = 1
let worstLead = 0
const origWrite = res._write.bind(res)
res._write = (chunk, enc, cb) => {
worstLead = Math.max(worstLead, counter.pulled - res.received)
origWrite(chunk, enc, cb)
}
await downloadAssets(asResponse(res), share, [makeAsset('a1')])
if (!res.writableFinished) await once(res, 'finish')
expect(counter.pulled).toBe(total)
// The lead is bounded by the stream buffers between the fetch body and
// the response: archiver and zip-stream each hold up to 1 MB per side,
// so roughly 4 MB in total. Without backpressure it would approach the
// whole 32 MB.
expect(worstLead).toBeLessThan(8 * 1024 * 1024)
}, 20_000)
it('fetches one asset at a time and cancels the upstream fetch when the client aborts mid-body', async () => {
const signals: AbortSignal[] = []
const fetchMock = trickleFetch(signals)
vi.stubGlobal('fetch', fetchMock)
const res = new FakeRes()
const done = downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2'), makeAsset('a3')])
// Give the first body a moment to start; nothing else should be in flight
await new Promise(resolve => setTimeout(resolve, 50))
expect(fetchMock).toHaveBeenCalledTimes(1)
res.destroy(new Error('client aborted'))
await done
expect(fetchMock).toHaveBeenCalledTimes(1)
expect(signals.every(s => s.aborted)).toBe(true)
}, 10_000)
it('resolves when the client aborts while the zip is streaming', async () => {
// Large enough that archiver can't flush everything into a destroyed
// response; without disconnect handling finalize() would hang forever.
vi.stubGlobal('fetch', instantFetch(5 * 1024 * 1024))
const res = new FakeRes()
const origWrite = res._write.bind(res)
let first = true
res._write = (chunk, enc, cb) => {
origWrite(chunk, enc, cb)
if (first) {
first = false
setImmediate(() => res.destroy(new Error('client aborted')))
}
}
await downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2')])
expect(res.destroyed).toBe(true)
}, 10_000)
it('answers 404 (not a broken zip) when the first asset fails before anything is sent', async () => {
vi.spyOn(console, 'log').mockImplementation(() => {})
vi.stubGlobal('fetch', vi.fn(async () => new globalThis.Response('nope', { status: 500 })))
const res = new FakeRes()
await downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2')])
expect(res.statusCode).toBe(404)
expect(res.destroyed).toBe(false)
expect(res.writableEnded).toBe(true)
expect(res.received).toBe(0)
expect(res.headers['Content-Type']).toBeUndefined()
}, 10_000)
it('destroys the response when a later asset keeps failing', async () => {
vi.spyOn(console, 'log').mockImplementation(() => {})
let calls = 0
vi.stubGlobal('fetch', vi.fn(async () => {
calls++
return calls === 1
? new globalThis.Response(new Uint8Array(2048), { status: 200 })
: new globalThis.Response('nope', { status: 500 })
}))
const res = new FakeRes()
await downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2')])
expect(res.received).toBeGreaterThan(0)
expect(res.destroyed).toBe(true)
expect(res.writableFinished).toBe(false)
}, 10_000)
it('destroys the response and stops fetching when a body errors mid-stream', async () => {
vi.spyOn(console, 'log').mockImplementation(() => {})
const signals: AbortSignal[] = []
const fetchMock = vi.fn(async (_url: unknown, init?: RequestInit) => {
signals.push(init?.signal as AbortSignal)
const stream = new ReadableStream({
start (controller) {
controller.enqueue(new Uint8Array(1024))
setTimeout(() => controller.error(new Error('upstream died')), 20)
}
})
return new globalThis.Response(stream, { status: 200 })
})
vi.stubGlobal('fetch', fetchMock)
const res = new FakeRes()
await downloadAssets(asResponse(res), share, [makeAsset('a1'), makeAsset('a2'), makeAsset('a3')])
expect(res.destroyed).toBe(true)
expect(fetchMock).toHaveBeenCalledTimes(1)
expect(signals.every(s => s.aborted)).toBe(true)
}, 10_000)
})
-63
View File
@@ -1,63 +0,0 @@
import { describe, it, expect, afterEach } from 'vitest'
import { promises as fs } from 'fs'
import { tmpdir } from 'os'
import { join } from 'path'
import { sweepStaleStagingDirs } from '../src/stream/download'
const PREFIX = 'ipp-zip-'
async function makeDir (name: string, mtimeMsAgo: number) {
const path = join(tmpdir(), name)
await fs.mkdir(path, { recursive: true })
await fs.writeFile(join(path, 'inside.bin'), 'x')
const mtime = new Date(Date.now() - mtimeMsAgo)
await fs.utimes(path, mtime, mtime)
return path
}
async function exists (path: string) {
try {
await fs.access(path)
return true
} catch {
return false
}
}
const created: string[] = []
afterEach(async () => {
await Promise.all(created.splice(0).map(p => fs.rm(p, { recursive: true, force: true }).catch(() => { /* best effort */ })))
})
describe('sweepStaleStagingDirs', () => {
it('deletes stale staging dirs older than the cutoff', async () => {
const stale = await makeDir(PREFIX + 'stale-' + process.pid + '-' + Math.random().toString(36).slice(2), 2 * 60 * 60 * 1000)
created.push(stale)
await sweepStaleStagingDirs(60 * 60 * 1000)
expect(await exists(stale)).toBe(false)
})
it('preserves staging dirs younger than the cutoff', async () => {
const fresh = await makeDir(PREFIX + 'fresh-' + process.pid + '-' + Math.random().toString(36).slice(2), 10 * 60 * 1000)
created.push(fresh)
await sweepStaleStagingDirs(60 * 60 * 1000)
expect(await exists(fresh)).toBe(true)
})
it('ignores directories that do not match the staging prefix', async () => {
const unrelated = await makeDir('not-ipp-zip-' + process.pid + '-' + Math.random().toString(36).slice(2), 24 * 60 * 60 * 1000)
created.push(unrelated)
await sweepStaleStagingDirs(60 * 60 * 1000)
expect(await exists(unrelated)).toBe(true)
})
it('uses its default 1h cutoff when called with no argument', async () => {
const stale = await makeDir(PREFIX + 'default-' + process.pid + '-' + Math.random().toString(36).slice(2), 2 * 60 * 60 * 1000)
const fresh = await makeDir(PREFIX + 'default2-' + process.pid + '-' + Math.random().toString(36).slice(2), 30 * 60 * 1000)
created.push(stale, fresh)
await sweepStaleStagingDirs()
expect(await exists(stale)).toBe(false)
expect(await exists(fresh)).toBe(true)
})
})
-6
View File
@@ -75,12 +75,6 @@ The bulk-zip and per-asset buttons can be toggled independently once downloads a
>
> To guarantee full-quality downloads of everything, leave the share's download permission on in Immich.
## `downloadFromImmichConcurrencyLimit`
**Type:** `int` · **Default:** `20`
Maximum number of assets IPP will fetch from your Immich server in parallel when building a "download all" zip. Lower this if your Immich server is slow or you see download timeouts on large albums; raise it for faster downloads if your server can handle the load.
## `allowSlugLinks`
**Type:** `bool`