fix: implement atomic append for compilation cache (#26830)

Note: this reverts commit ffd6cf60eb0fc3ea567802da7ed0e6de17f2133f.

Fixes #26769
This commit is contained in:
Andrey Lushnikov 2023-08-31 17:55:13 -07:00 committed by GitHub
parent 4948920437
commit 54e4e5caca
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 31 additions and 15 deletions

View File

@ -15,10 +15,8 @@
*/ */
import fs from 'fs'; import fs from 'fs';
import type { WriteFileOptions } from 'fs';
import path from 'path'; import path from 'path';
import { rimraf } from '../utilsBundle'; import { rimraf } from '../utilsBundle';
import { createGuid } from './crypto';
export const existsAsync = (path: string): Promise<boolean> => new Promise(resolve => fs.stat(path, err => resolve(!err))); export const existsAsync = (path: string): Promise<boolean> => new Promise(resolve => fs.stat(path, err => resolve(!err)));
@ -55,11 +53,3 @@ export async function copyFileAndMakeWritable(from: string, to: string) {
export function sanitizeForFilePath(s: string) { export function sanitizeForFilePath(s: string) {
return s.replace(/[\x00-\x2C\x2E-\x2F\x3A-\x40\x5B-\x60\x7B-\x7F]+/g, '-'); return s.replace(/[\x00-\x2C\x2E-\x2F\x3A-\x40\x5B-\x60\x7B-\x7F]+/g, '-');
} }
export function writeFileSyncAtomic(aPath: string, data: Buffer | string, options: WriteFileOptions) {
const dirName = path.dirname(aPath);
const fileName = path.basename(aPath);
const tmpPath = path.join(dirName, fileName + '-' + createGuid());
fs.writeFileSync(tmpPath, data, options);
fs.renameSync(tmpPath, aPath);
}

View File

@ -18,7 +18,6 @@ import fs from 'fs';
import os from 'os'; import os from 'os';
import path from 'path'; import path from 'path';
import { sourceMapSupport } from '../utilsBundle'; import { sourceMapSupport } from '../utilsBundle';
import { writeFileSyncAtomic } from 'playwright-core/lib/utils';
export type MemoryCache = { export type MemoryCache = {
codePath: string; codePath: string;
@ -74,6 +73,22 @@ function _innerAddToCompilationCache(filename: string, options: { codePath: stri
memoryCache.set(filename, options); memoryCache.set(filename, options);
} }
// Each worker (and runner) process compiles and caches client code and source maps.
// There are 2 levels of caching:
// 1. Memory Cache: per-process, single threaded.
// 2. SHARED Disk Cache: helps to re-use caching across processes (worker re-starts).
//
// Now, SHARED Disk Cache might be accessed at the same time by different workers, trying
// to write/read concurrently to it. We tried to implement "atomic write" to disk cache, but
// failed to do so on Windows. See context: https://github.com/microsoft/playwright/issues/26769#issuecomment-1701870842
//
// Under further inspection, it turns out that our Disk Cache is append-only, so instead of a general-purpose
// "atomic write" it will suffice to have "atomic append". For "atomic append", it is sufficient to:
// - make sure there are no concurrent writes to the same file. This is implemented using the `wx` flag to the Node.js `fs.writeFile` calls.
// - have a signal that guarantees that file is actually finished writing. We use marker files for this.
//
// The following method implements the "atomic append" principles for the disk cache.
//
export function getFromCompilationCache(filename: string, hash: string, moduleUrl?: string): { cachedCode?: string, addToCache?: (code: string, map?: any) => void } { export function getFromCompilationCache(filename: string, hash: string, moduleUrl?: string): { cachedCode?: string, addToCache?: (code: string, map?: any) => void } {
// First check the memory cache by filename, this cache will always work in the worker, // First check the memory cache by filename, this cache will always work in the worker,
// because we just compiled this file in the loader. // because we just compiled this file in the loader.
@ -85,7 +100,8 @@ export function getFromCompilationCache(filename: string, hash: string, moduleUr
const cachePath = calculateCachePath(filename, hash); const cachePath = calculateCachePath(filename, hash);
const codePath = cachePath + '.js'; const codePath = cachePath + '.js';
const sourceMapPath = cachePath + '.map'; const sourceMapPath = cachePath + '.map';
if (fs.existsSync(codePath)) { const markerFile = codePath + '-marker';
if (fs.existsSync(markerFile)) {
_innerAddToCompilationCache(filename, { codePath, sourceMapPath, moduleUrl }); _innerAddToCompilationCache(filename, { codePath, sourceMapPath, moduleUrl });
return { cachedCode: fs.readFileSync(codePath, 'utf8') }; return { cachedCode: fs.readFileSync(codePath, 'utf8') };
} }
@ -93,9 +109,19 @@ export function getFromCompilationCache(filename: string, hash: string, moduleUr
return { return {
addToCache: (code: string, map: any) => { addToCache: (code: string, map: any) => {
fs.mkdirSync(path.dirname(cachePath), { recursive: true }); fs.mkdirSync(path.dirname(cachePath), { recursive: true });
if (map) try {
writeFileSyncAtomic(sourceMapPath, JSON.stringify(map), 'utf8'); if (map)
writeFileSyncAtomic(codePath, code, 'utf8'); fs.writeFileSync(sourceMapPath, JSON.stringify(map), { encoding: 'utf8', flag: 'wx' });
fs.writeFileSync(codePath, code, { encoding: 'utf8', flag: 'wx' });
// NOTE: if the worker crashes RIGHT HERE, before creating a marker file, we will never be able to
// create it later on. As a result, the entry will never be added to the disk cache.
//
// However, this scenario is EXTREMELY unlikely, so we accept this
// limitation to reduce algorithm complexity.
fs.closeSync(fs.openSync(markerFile, 'w'));
} catch (error) {
// Ignore error that is triggered by the `wx` flag.
}
_innerAddToCompilationCache(filename, { codePath, sourceMapPath, moduleUrl }); _innerAddToCompilationCache(filename, { codePath, sourceMapPath, moduleUrl });
} }
}; };