Directory Notifications to find changes
Pending Changes is now faster than ever because it doesn’t need to traverse the workspace anymore. We have implemented a new mechanism based on Windows Directory Notifications to detect workspace changes faster than ever.
It is available only on Windows but we’ll eventually implement it for Linux and Mac (based on their corresponding notification mechanisms).
What does it mean for you? Well, as soon as you install 220.127.116.118 or 18.104.22.1684 (or higher) Pending Changes will be faster. You’ll clearly notice the speed up with really large workspaces (in number of files) and with slow disks. The slower the disk is, the clearer the speed up will be.
How Pending Changes works? (without directory notifications)
Whenever you click on “refresh” on Pending Changes Plastic triggers a search to find the files that have been modified on your workspace.
The diagram below describes the process in detail:
- The process checks the “pending changes options” first: if only checkouts are requested, then there’s nothing to look for, just print the list. That’s why working with checkouts makes sense for huge workspaces (>400k files).
- If the options to find changed files on disk are set, the directory walk will start.
- For each directory starting on the root of the workspace Plastic will try to find changed files. It will compare the timestamp on disk with the stored timestamp on the wktree file: the plastic.wktree file (inside .plastic) stores the metadata of the file. It know “how it was” after the last update or checkin. So if the timestamp and size doesn’t match, the file was changed. If timestamp doesn’t match and size does, Plastic hashes the file. It is slower but it makes sure the file is different. Alternatively there’s an option to force Plastic to always find changes based on file contents (ignoring the timestamp) which is definitely slower but required on some scenarios.
- At the end of the disk walk, Plastic has a list of all the modified files on your workspace.
The diagram doesn’t show the last step (if the option is set): find “moved and renamed files” by matching potential added and deleted files.
What I want to explain with the diagram is that there is at least one IO operation by directory. If you keep pressing “refresh” on the Pending Changes view, chances are the list will be filled quickly: after the first traversal the workspace will be loaded in the file system cache so the next reads will be blazing fast. But if your disk is not very fast, or your computer is performing a lot of IO, or using a lot of RAM, chances are that your workspace won’t be entirely loaded in the file system cache, and then walking it will take longer.
You probably noticed it when after some coding you go back to Plastic, click refresh, and it takes longer than usual. This is exactly what we wanted to improve with this feature.
How Pending Changes with Directory Notifications works?
It is rather simple: we use Windows directory notifications to listen to events on the workspace directory. Each time a file is written, deleted, added, moved or renamed inside the workspace, Plastic gets a notification.
So, while we perform an initial directory traversal the first time the Pending Changes view loads, no other full directory walk will be needed later, greatly speeding up the operation.
What we do is the following: after the first traversal we keep a tree with the metadata of what is on disk, and we invalidate parts of it (on a directory basis) each time a change happens inside it. This way Pending Changes only has to reload parts of the tree instead of walking the workspace entirely. It saves precious time while still being a robust solution.
One of the issues with Directory Notifications on Windows is that it can’t really notify file or directory moves, so you have to match pairs of added/deleted. Instead of trying to pair the notifications we just invalidate parts of the tree and let the regular Pending Changes code to do the rest.
So, there’s still room for improvement but our initial tests probed that the extra complexity of doing a more precise event tracking didn’t pay off compared to just invalidating parts of the tree.
Chances are you are already using it :-)
If you’re using 22.214.171.1248 or higher or 126.96.36.1994 or higher, you’re already enjoying the Directory Notifications powered Pending Changes view.