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
I would love the ability to specify some sort of hook to run on downloaded/requested assets. Could happen synchronously or asynchronously for my purposes (maybe even preferable asynchronously). I'd see it as dispatching an action to the RBE scheduler, but probably lots of ways to skin that cat.
I'd use it to unpack and directly upload all constituent files into the CAS. This way clients aren't downloading large assets only to unzip and upload them right back.
I'd also use it to log assets I'm using and maybe automatically mirror them into persistent object storage that I control (a la mirror.tensorflow.org).
Another use, less personally interesting, would be to enforce policy (whitelisting sources), etc.
The text was updated successfully, but these errors were encountered:
Unpacking zips at the asset server level is definitely something that makes a lot of sense. I'm not sure how the clients currently behave and whether or not they fetch the zip regardless.
We definitely discussed this at one point as a bonus towards implementing the remote executor fetcher. Although I think it would likely be more standardised build actions such as unpacking zips rather than user customizable actions as it would be tricky to support these in situations without a remote executor configured.
This related issue is a little old but definitely has overlap #10
I would love the ability to specify some sort of hook to run on downloaded/requested assets. Could happen synchronously or asynchronously for my purposes (maybe even preferable asynchronously). I'd see it as dispatching an action to the RBE scheduler, but probably lots of ways to skin that cat.
I'd use it to unpack and directly upload all constituent files into the CAS. This way clients aren't downloading large assets only to unzip and upload them right back.
I'd also use it to log assets I'm using and maybe automatically mirror them into persistent object storage that I control (a la mirror.tensorflow.org).
Another use, less personally interesting, would be to enforce policy (whitelisting sources), etc.
The text was updated successfully, but these errors were encountered: