Is there a way to avoid races involving `tusd` deleting files from storage?

[timabbott]

The Zulip implementation of tusd involves using a pre-finish hook to create some data structures in our database (used to manage quota, track the messages referencing the file, etc.), verify that image files are thumbnailable, and kick off a worker job to begin thumbnailing images.

The problem I'm seeing is that the tusd protocol allows the client to delete the upload via a DELETE request at any time, including after an upload has completed, and tusd implements that by first removing the file from the backing storage and then notifying the application via the post-terminate hook. A specific case where this occurred involved nginx deciding the client had gone away when tusd was trying to write the response in the finish stage, serving a 499, which appears to have caused uppy to ask to delete/terminate the upload? At least, that's what we can piece together from logs. The end result is that our worker job for thumbnailing images sees a corrupt state where the application database thinks the file exists, but it's been deleted from the backing store by tusd.

I don't see a good way to implement this aspect of the protocol in a way that is safe from races where the application database is in a corrupt state. Options that we've considered include:

• Enabling the --disable-terminate option on the tusd command line. This is not ideal, in that I think this will prevent all terminations, including of partially uploaded files? I'm also not sure after some source-diving whether clients like uppy have a way to be configured to support that setting.
• Having the post-terminate hook clean up our database structures associated with the upload after tusd has deleted the file. The problem is that this happens after the file is removed from the backing storage, so there is a fundamental race where the database thinks the file exists but it is missing when other application code is trying to process it (for thumbnailing or whatever). This results in file-not-found exceptions that normally would be very scary being a routine/normal thing.
• Deferring visible side effects, like queuing the file to be processed for thumbnailing, to the post-finish hook. This might limit our exposure to the race, but it doesn't offer a way to solve the problem that tusd might at any time delete already-uploaded files from the backing storage before adjusting our data structures.

I see the following options for how tusd could support a race-free application implementation:

• Recommended: Add a pre-terminate hook that can be used by the application to either reject the deletion request (say, if the user doesn't have permission to delete the file for whatever reason) or indicate to tusd that as part of its processing, it has deleted the file from the backing store, and tusd does not need to do that itself. This would allow us to make that hook use our existing application-layer code to delete uploaded files safely (using 2-phase commit or other standard techniques to avoid invariant violations).
• Add an option to disallow tusd deleting/terminated an upload after the pre-finish hook has completed successfully.

But maybe I'm misunderstanding something about the protocol -- is there a way that we can close this race?

[Acconut]

That's a valid point! Access to the uploaded files during the pre-finish hook is protected through tusd's internal locking system. The pre-finish hook is part of the PATCH request handling and the lock is only released once the request handling is completed. Deleting uploads requires acquiring the corresponding locks, which avoids prevents races between pre-finish and termination requestss.

However, post-finish hooks or worker jobs that were started by pre-finish are not protected through these locks, of course. This leads to the situation you observe where a user can delete an upload while your application is trying to process it. Uppy does send a termination request in a few situations. For example, when the user cancels the upload, Uppy terminates the tus upload, as far as I know. You might want to double check with the Uppy project to see if that's correct and whether it can be configured. How a 499 response from Nginx can trigger this though, I don't know.

With all of that in mind, we still need a solution. A pre-terminate hook for giving the user control over file termination is a useful tool, but also defers responsibility of preventing data races to the user, which is not ideal. A more general solution might be better to help in situations where the application is accessing the uploaded files and needs to prevent concurrent access from tusd. I can think of three typical situations where this happens:

1. Application processes uploaded file and tusd is instructed to terminate the upload (what you are mentioning here)
2. Application modifies or deleted uploaded file and tusd attempts to read the upload for a concatenation operation
3. Application wants to cleanup unfinished upload, but user wants to upload at the same time

With a pre-terminate hook (and custom logic) or an option to disable termination of completed uploads, you can prevent 1), but not the others.

I wonder if it's sensible to allow application to interact with tusd's locks. For example, before an application access an uploaded file, it asks tusd to acquire the lock/lease for the corresponding upload resource. This prevents concurrent access from tusd, so no data races are possible. Once the application is done processing, the lock/lease can be released again. This might be a bit more involved, but should cover all cases. What do you think?

[timabbott]

I think a pre-terminate hook would be safe enough if tusd guaranteed that it was holding its own locks, and accepted a parameter in the hook response for whether the server deleted the file itself or not... if one could rely on the application to not try to interact with unfinished uploads (which seems likely?), those second two problems seem like they'd mostly not be important.

I guess the structural question is "who owns the uploaded file?" I feel like there's a reasonable mode where tusd does and the application isn't allowed to mess with the file, or needs to grab tusd locks to do it. As an application author, though, I don't want to have to think about tusd's locking scheme; I'd prefer to abstract that away.

It seems like a lot of applications will want a model where once tusd has uploaded a file, the upload process is done, and the application has full ownership over the file. (Zulip is OK with clients deleting previously uploaded files as long as our data structures aren't corrupted in so doing, but one can certainly imagine applications that aren't!).

So conceptually, that'd be a mode where there is a handoff once the upload is complete from tusd owning the file to the application owning it, and after that handoff, tusd is unable to mutate/delete the file after upload completion, and needs to do something to verify whether the file is still there and unmodified before using it.

I've not had a chance to think this through fully, but that's my initial reaction.