java-watch

a java file watcher that works across platforms and supports recursion, single file watches, and tries to make sure no events are missed.

Features

Currently working features in java-watch:

• Regular directory watches
• Recursive watches, even if platform doesn't support it natively.
• Recursive watches also work inside directories created after the watch started
• Even in case of overflow you will get notifications of new directories (and it's recursive files), modification events will however not be simulated
• Single file watches
• Multiple watches for the same directory are merged to avoid overloading the kernel
• Events are process on a worker pool, which you can customize.

Future features:

• Avoid poll based watcher in macOS/OSX that only detects changes every 2 seconds (see #4)
• Support file watches natively in linux
• Monitor only specific events (such as only CREATES)

Usage

Import dependency in pom.xml:

<dependency>
  <groupId>engineering.swat</groupId>
  <artifactId>java-watch</artifactId>
  <version>${java-watch-version}</version>
</dependency>

Start using java-watch:

var directory = Path.of("tmp", "test-dir");
var watcherSetup = Watcher.watch(directory, WatchScope.PATH_AND_CHILDREN)
  .withExecutor(Executors.newCachedThreadPool()) // optionally configure a custom thread pool
  .onEvent(watchEvent -> {
    System.err.println(watchEvent);
  });

try(var active = watcherSetup.start()) {
  System.out.println("Monitoring files, press any key to stop");
  System.in.read();
}
// after active.close(), the watch is stopped and
// no new events will be scheduled on the threadpool

Related work

Before starting this library, we wanted to use existing libraries, but they all lacked proper support for recursive file watches or lacked configurability. This library now has a growing collection of tests and a small API that should allow for future improvements without breaking compatibility.

The following section describes the related work research on the libraries and underlying limitations.

After reading the documentation of the following discussion on file system watches:

• Paul Millr's nodejs chokidar
• Enrico Maria Crisostomo's c++ fswatch/libfswatch
• work by rjeczalik in the go notify package
• Greg Methvin's java library directory-watcher
• Pathikrit Bhowmick's scala library better-files
• Java's documentation on a recursive file watcher
• openjdk's PR 10140/8293067 to try and use FSEventStream on macOS
• nodejs watch function caveats
• Overview in 2012 by regedit

We can come to the following conclusion: file system watches are hard and have platform specific limitations. In summary:

OS	API	file	directory	recursive directory	overflow	network	notes
Windows	ReadDirectoryChangesW	❌	✅	✅	generic error marker	some, error in case not supported	hard to correctly setup (there are multiple ways to get updates), can be chatty with it's events. can lock the directory it's monitoring.
Linux	inotify	✅	✅	❌	generic error marker	only local changes, no error	note that the new fanotify supports recursive watches, but only at mount points, not for arbitrary directories.
macOS & BSD	kqueue	?	✅	❌	can quickly run out of file descriptors	?	implementing recursive directory watches this way will quickly run out of file descriptors
macOS	FSEvents	✅	✅	✅	generic error marker	?	Some report it works great, but openjdk stopped doing this direction of the implementation as it consistently failed a test with a lot of IO operations and register and unregisters of watches. Reporting that the API would just stop reporting any events

To avoid licensing conflicts we have not read the source code of any of these libraries/frameworks. The related work study is based purely on public documentation and discussions.