You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Following up from #149 - it's really a bit of rabbit hole if you look at fsnotify/fsnotify#635 and the numerous other related issues.
In short, the comment for the fsnotify.Write event describes the issue:
// The pathname was written to; this does *not* mean the write has finished,
// and a write can be followed by more writes.
So what I've observed when using os.WriteFile on Linux systems + K8s PVCs is that initially fswatcher processes a fsnotify.Write event where the file was rewritten / has size of 0 bytes. It then receives a subsequent fsnotify.Write and the file is finished being written with a non-zero size. The problem is both the 0 bytes and the non-zero bytes events change the tracked hash, causing two unnecessary onChange events in the case where the file did not actually change, and one extra onChange event when the file did change. I've had to add a defense in my use of this utility to ignore onChange if the size of the file is 0 for now (which still isn't great because of the second onChange). For example from my logs:
These features are still in main of the lib / unreleased, but once they are we can detect if such events are supported and change what events we listen for as a result - make Linux optimized for detecting file changes which is the main OS targetted for production use cases of this library since it is used primarily in containerized applications.
This issue to track the fsnotify upstream for the more efficient event types and fix this once they are released. If too much time passes, we can consider maintaining an fsnotify fork.
The text was updated successfully, but these errors were encountered:
onelapahead
changed the title
fswatcher Does Dedup Writes or Detect "CloseWrites"
fswatcher Should Dedup Writes or Detect "CloseWrites"
Oct 31, 2024
Following up from #149 - it's really a bit of rabbit hole if you look at fsnotify/fsnotify#635 and the numerous other related issues.
In short, the comment for the
fsnotify.Write
event describes the issue:So what I've observed when using
os.WriteFile
on Linux systems + K8s PVCs is that initiallyfswatcher
processes afsnotify.Write
event where the file was rewritten / has size of 0 bytes. It then receives a subsequentfsnotify.Write
and the file is finished being written with a non-zero size. The problem is both the 0 bytes and the non-zero bytes events change the tracked hash, causing two unnecessaryonChange
events in the case where the file did not actually change, and one extraonChange
event when the file did change. I've had to add a defense in my use of this utility to ignoreonChange
if the size of the file is 0 for now (which still isn't great because of the secondonChange
). For example from my logs:There's examples of how we could go to the trouble deduping the writes: https://github.com/fsnotify/fsnotify/blob/main/cmd/fsnotify/dedup.go
But that feels a bit convoluted if we can avoid it (sorry Windows).
Instead it looks like the maintainer is actively working on support non-portable events for Linux systems like
CLOSE_WRITE
: https://github.com/fsnotify/fsnotify/blob/main/fsnotify.go#L211-L219These features are still in main of the lib / unreleased, but once they are we can detect if such events are supported and change what events we listen for as a result - make Linux optimized for detecting file changes which is the main OS targetted for production use cases of this library since it is used primarily in containerized applications.
This issue to track the
fsnotify
upstream for the more efficient event types and fix this once they are released. If too much time passes, we can consider maintaining anfsnotify
fork.The text was updated successfully, but these errors were encountered: