File Watcher Internals

High Level

We have 2 different implementations for file watching file and folder paths:

• recursive: ParcelWatcher via parcel-watcher
• non-recursive: NodeJSWatcherLibrary via fs.watch

Event Correlation vs. Non-Correlation

Traditionally file events are not correlated: that means, any request to IFileService.watch() will contribute to the global IFileService.onDidFilesChange event reaching a lot of consumers, including extensions.

To make file watching more efficient, event correlation was added: IFileService.createWatcher() is a new method that returns an emitter for events specific to the watch request. None of the events will end up on the global IFileService.onDidFilesChange event which helps to reduce compute need. Correlated file watcher have a unique identifier (number) to be able to distinguish them from others.

Today, correlated file watching is not yet being used in our product. We plan to use it for extensions that have complex file watching requests such as TypeScript.

End to End Flow

IFileService.watch() / IFileService.createWatcher()

Requests for watching are deduplicated if the request is an identical match. That means, if each property of the watch request (path, recursive, includes, excludes, correlation identifier) is the same.

This is done to avoid duplicate identical requests for file watching that are easy to identify.

The request to file watch is passed onto the IFileSystemProvider that matches the scheme of the resource path to watch.

DiskFileSystemProvider.watch()

Our DiskFileSystemProvider deals with all file resource schemes and runs from the Electron main process. Since file watching is compute intense, we leverage UtilityProcess to host the file watchers in a separate process.

We look if the request to watch is recursive or not and hand it off to ParcelWatcher or NodeJSWatcherLibrary

ParcelWatcher / NodeJSWatcherLibrary

We do support correlated and uncorrelated watch requests. A request is considered to be equal in correlation if the correlation identifier matches or if both requests are uncorrelated.

We do some deduplication of watch requests to avoid watching the same path twice:

• requests for same path and same correlation are ignored (last one wins)
• recursive requests for overlapping path and same correlation are ignored (shortest path wins)
• requests for watching files try to reuse an existing recursive watcher

If a watched path either does not exist in the beginning or gets deleted after being watched, the watcher maybe suspended and resumed when the path comes back, based on these rules:

• correlated watch requests support suspend/resume unconditionally (this was added to support TypeScript)
• uncorrelated recursive watch requests try to resume watching by installing a listener on the parent path which can fail if the parent is deleted as well [1]
• uncorrelated non-recursive watch requests are ignored if the path does not exist in the beginning or remain failed if the path is deleted after watching has begun [1]

Suspend/resume is implemented with two strategies:

• if the path is already watched by a recursive watcher, reuse that watcher
• otherwise install a polling watcher on the path (fs.watchFile)

[1] this is arguably an issue that we need to revisit at one point: a watcher that stopped because its path got deleted is in a failed state until disposed, which may never happen. Given this failed state is not communicated to the outside, another request for watching might wrongly be deduplicated with this failing instance. A better approach would be to support suspend/resume always or detect failed watchers and dispose them.

vscode.workspace.createFileSystemWatcher

Extensions are able to create file watchers via the vscode.workspace.createFileSystemWatcher() API. The behaviour of the file watcher depends on the options that are passed in.

Today, the stable API will only create uncorrelated file watchers and extensions cannot specify custom exclude rules. A new proposed API is offered with support for custom exclude rules that will create correlated file watchers (https://github.com/microsoft/vscode/issues/169724).

If the pattern to watch is a string (and not RelativePattern), we limit events to paths that are inside the workspace only. Any event for a path outside the workspace is ignored. Under the hood, we do not install any file watcher because we do workspace watching by default.

For when RelativePattern is used, patterns that include ** or / are considered recursive watch requests if the path is a folder.

Correlated watch requests are pretty much handed off to the file service without further massaging, but uncorrelated recursive requests are massaged to reduce the impact on everyone that listens to IFileService.onDidFilesChange. These get their exclude rules configured based on the files.watcherExclude setting. As such, an extension that uses uncorrelated recursive file watching is at the mercy of how files.watcherExclude is configured. For that reason, the new proposed API was added.

Raw File Watcher Internals

node.js / parcel watcher library

• requests for non existing paths are ignored unless correlated
• requests for same path and same correlation (including undefined) are ignored (last one wins)
• recursive requests for overlapping path and same correlation are ignored (shortest path wins)
• non-recursive requests for files try to reuse an existing recursive watcher
• requests for a path that gets deleted later maybe rewatched
  • correlated requests get rewatched either by reusing an existing recursive watcher or via fs.watchFile
  • uncorrelated recursive requests get rewatched by fs.watch on the parent path if exists

DiskFileSystemProvider

• every request to watch is passed through and not deduplicated in any way

IFileService

• requests that are identical (resource and options) are deduplicated
• uncorrelated watch requests emit globally via onDidFilesChange
• correlated watch requests emit only to the one that requested the watch

createFileSystemWatcher

• correlates if new proposed API is used that allows to pass in excludes
• if pattern is a string we assume workspace watch mode
  • any "out of workspace" events are ignored
  • no request to watch is sent to the file service assuming the workspace is already watched
• patterns with a ** or / are treated as recursive watch requests if the path is a folder, otherwise non-recursive
• correlated watch requests
  • get to match on events from same correlation identifier
• uncorrelated watch requests
  • get to match on events from all other uncorrelated watchers
  • exclude rules are automatically added from files.watcherExclude setting to recursive watch requests
  • include rules will be computed for non-recursive watchers that are inside the workspace to match on configured exclude rules as a way to prevent duplicate events from non-recursive and recursive watchers inside the workspace
    • if no exclude rules are configured, the non-recursive watcher is ignored