Repository files navigation

A lean and fast 'fs' for the browser

I wanted to see if I could make something faster thanBrowserFS orfiler that still implements enough of thefs API to run theisomorphic-git test suite in browsers.

This library does not even come close to implementing the fullfs API.Instead, it only implementsthe subset used by isomorphic-git 'fs' plugin interface plus thefs.promises versions of those functions.

Unlike BrowserFS, which has a dozen backends and is highly configurable,lightning-fs has a single configuration that should Just Work for most users.

1. needs to work in all modern browsers
2. needs to work with large-ish files and directories
3. needs to persist data
4. needs to enable performant web apps

Req #3 excludes pure in-memory solutions. Req #4 excludeslocalStorage because it blocks the DOM and cannot be run in a webworker. Req #1 excludes WebSQL and Chrome's FileSystem API. So that leaves us with IndexedDB as the only usable storage technology.

1. speed (time it takes to execute file system operations)
2. bundle size (time it takes to download the library)
3. memory usage (will it work on mobile)

In order to get improve #1, I ended up making a hybrid in-memory / IndexedDB system:

• mkdir,rmdir,readdir,rename, andstat are pure in-memory operations that take 0ms
• writeFile,readFile, andunlink are throttled by IndexedDB

The in-memory portion of the filesystem is persisted to IndexedDB with a debounce of 500ms.The files themselves are not currently cached in memory, because I don't want to waste a lot of memory.Applications can alwaysadd an LRU cache on top oflightning-fs - if I add one internally and it isn't tuned well for your application, it might be much harder to work around.

Multiple tabs (and web workers) can share a filesystem. However, because SharedArrayBuffer is still not available in most browsers, the in-memory cache that makes LightningFS fast cannot be shared. If each thread was allowed to update its cache independently, then you'd have a complex distributed system and would need a fancy algorithm to resolve conflicts. Instead, I'm counting on the fact that your multi-threaded applications will NOT be IO bound, and thus a simpler strategy for sharing the filesystem will work. Filesystem access is bottlenecked by a mutex (implemented via polling and an atomic compare-and-replace operation in IndexedDB) to ensure that only one thread has access to the filesystem at a time. If the active thread is constantly using the filesystem, no other threads will get a chance. However if the active thread's filesystem goes idle - no operations are pending and no new operations are started - then after 500ms its in-memory cache is serialized and saved to IndexedDB and the mutex is released. (500ms was chosen experimentally such that anisomorphic-gitclone operation didn't thrash the mutex.)

While the mutex is being held by another thread, any fs operations will be stuck waiting until the mutex becomes available. If the mutex is not available even after ten minutes then the filesystem operations will fail with an error. This could happen if say, you are trying to write to a log file every 100ms. You can overcome this by making sure that the filesystem is allowed to go idle for >500ms every now and then.

First, create or open a "filesystem". (The name is used to determine the IndexedDb store name.)

importFSfrom'@isomorphic-git/lightning-fs';constfs=newFS("testfs")

Note: It is better not to create multipleFS instances using the same name in a single thread. Memory usage will be higher as each instance maintains its own cache, and throughput may be lower as each instance will have to compete over the mutex for access to the IndexedDb store.

Options object:

You can procrastinate initializing the FS object until later.And, if you're really adventurous, you canre-initialize it with a different name to switch between IndexedDb databases.

importFSfrom'@isomorphic-git/lightning-fs';constfs=newFS()// Some time later...fs.init(name,options)// Some time later...fs.init(different_name,different_options)

Make directory

Options object:

Remove directory

Read directory

The callback return value is an Array of strings. NOTE:To save time, it is NOT SORTED. (Fun fact: Node.js'readdir output is not guaranteed to be sorted either. I learned that the hard way.)

data should be a string of a Uint8Array.

Ifopts is a string, it is interpreted as{ encoding: opts }.

Options object:

The result value will be a Uint8Array or (ifencoding is'utf8') a string.

Ifopts is a string, it is interpreted as{ encoding: opts }.

Options object:

Delete a file

Rename a file or directory

The result is a Stat object similar to the one used by Node but with fewer and slightly different properties and methods.The included properties are:

• type ("file" or "dir")
• mode
• size
• ino
• mtimeMs
• ctimeMs
• uid (fixed value of 1)
• gid (fixed value of 1)
• dev (fixed value of 1)

The included methods are:

• isFile()
• isDirectory()
• isSymbolicLink()

Likefs.stat except that paths to symlinks return the symlink stats not the file stats of the symlink's target.

Create a symlink atfilepath that points totarget.

Read the target of a symlink.

Create or change the stat data for a file backed by HTTP. Size is fetched with a HEAD request. Useful when using an HTTP backend withouturlauto set, as then files will only be readable if they have stat data.Note that stat data is made automatically from the file/.superblock.txt if found on the server./.superblock.txt can be generated or updated with theincluded standalone script.

Options object:

Returns the size of a file or directory in bytes.

All the same functions as above, but instead of passing a callback they return a promise.

There are only two reasons I can think of that you would want to do this:

1. Thefs module is normally a singleton. LightningFS allows you to safely(ish) hotswap between various data sources by callinginit multiple times with different options. (It keeps track of file system operations in flight and waits until there's an idle moment to do the switch.)

2. LightningFS normalizes all the lovely variations of node'sfs arguments:

• fs.writeFile('filename.txt', 'Hello', cb)
• fs.writeFile('filename.txt', 'Hello', 'utf8', cb)
• fs.writeFile('filename.txt', 'Hello', { encoding: 'utf8' }, cb)
• fs.promises.writeFile('filename.txt', 'Hello')
• fs.promises.writeFile('filename.txt', 'Hello', 'utf8')
• fs.promises.writeFile('filename.txt', 'Hello', { encoding: 'utf8' })

And it normalizes filepaths. And will convert plainStatLike objects intoStat objects with methods likeisFile,isDirectory, etc.

If that fits your needs, then you can provide abackend option and LightningFS will use that. Implement as few/many methods as you need for your application to work.

Note: If you use a custom backend, you are responsible for managing multi-threaded access - there are no magic mutexes included by default.

Note: throwing an error with the correct.code property for any given situation is often important for utilities likemkdirp andrimraf to work.

typeEncodingOpts={encoding?:'utf8';}typeStatLike={type:'file'|'dir'|'symlink';mode:number;size:number;ino:number|string|BigInt;mtimeMs:number;ctimeMs?:number;}interfaceIBackend{// highly recommended - usually necessary for apps to workreadFile(filepath:string,opts:EncodingOpts):Awaited<Uint8Array|string>;// throws ENOENTwriteFile(filepath:string,data:Uint8Array|string,opts:EncodingOpts):void;// throws ENOENTunlink(filepath:string,opts:any):void;// throws ENOENTreaddir(filepath:string,opts:any):Awaited<string[]>;// throws ENOENT, ENOTDIRmkdir(filepath:string,opts:any):void;// throws ENOENT, EEXISTrmdir(filepath:string,opts:any):void;// throws ENOENT, ENOTDIR, ENOTEMPTY// recommended - often necessary for apps to workstat(filepath:string,opts:any):Awaited<StatLike>;// throws ENOENTlstat(filepath:string,opts:any):Awaited<StatLike>;// throws ENOENT// suggested - used occasionally by appsrename(oldFilepath:string,newFilepath:string):void;// throws ENOENTreadlink(filepath:string,opts:any):Awaited<string>;// throws ENOENTsymlink(target:string,filepath:string):void;// throws ENOENT// bonus - not part of the standard `fs` modulebackFile(filepath:string,opts:any):void;du(filepath:string):Awaited<number>;// lifecycle - useful if your backend needs setup and teardowninit?(name:string,opts:any):Awaited<void>;// passes initialization optionsactivate?():Awaited<void>;// called before fs operations are starteddeactivate?():Awaited<void>;// called after fs has been idle for a whiledestroy?():Awaited<void>;// called before hotswapping backends}interfaceIDB{saveSuperblock(superblock):void;loadSuperblock():Awaited<Buffer>;readFile(inode):Awaited<Buffer>;writeFile(inode,data):Awaited<void>;unlink(inode):Awaited<void>;wipe():void;close():void;}

MIT