Prevent concurrent model uploads from overwriting versions
This commit is contained in:
parent
002a4aaaba
commit
df83ebc834
1 changed files with 261 additions and 53 deletions
|
|
@ -70,6 +70,13 @@ Often, especially when working with data-heavy applications, large files can pro
|
|||
|
||||
## From the command-line
|
||||
|
||||
Versions are allocated when an upload commits. Writers sharing a cache directory
|
||||
serialize their uploads with a file lock. S3 and GridFS also reject conflicting
|
||||
version commits from other hosts with `FileExistsError`; retry the upload to
|
||||
allocate a fresh version. S3-compatible services must support conditional writes
|
||||
(`If-None-Match`), including multipart completion. GridFS writers need permission
|
||||
to create the unique index used for GreatAI's versioned filenames.
|
||||
|
||||
Remote cache filenames now include a storage-specific namespace. Existing local
|
||||
model stores remain readable, while remote models cached by older releases are
|
||||
downloaded again into the new namespace. Configure the same bucket or database
|
||||
|
|
|
|||
Loading…
Reference in a new issue