2df6b670c6
There is probably a better place for this documentation but we'd like to move this elsewhere than a github issue. We can move this to a more appropriate location with the documentation effort. Signed-off-by: Stephen J Day <stephen.day@docker.com>
2.2 KiB
Architecture
TODO(stevvooe): Discuss the architecture of the registry, internally and externally, in a few different deployment scenarios.
Design
Eventual Consistency
NOTE: This section belongs somewhere, perhaps in a design document. We are leaving this here so the information is not lost.
Running the registry on eventually consistent backends has been part of the design from the beginning. This section covers some of the approaches to dealing with this reality.
There are a few classes of issues that we need to worry about when implementing something on top of the storage drivers:
- Read-After-Write consistency (see this article on s3).
- Write-Write Conflicts.
In reality, the registry must worry about these kinds of errors when doing the following:
- Accepting data into a temporary upload file may not have latest data block yet (read-after-write).
- Moving uploaded data into its blob location (write-write race).
- Modifying the "current" manifest for given tag (write-write race).
- A whole slew of operations around deletes (read-after-write, delete-write races, garbage collection, etc.).
The backend path layout employs a few techniques to avoid these problems:
- Large writes are done to private upload directories. This alleviates most of the corruption potential under multiple writers by avoiding multiple writers.
- Constraints in storage driver implementations, such as support for writing after the end of a file to extend it.
- Digest verification to avoid data corruption.
- Manifest files are stored by digest and cannot change.
- All other non-content files (links, hashes, etc.) are written as an atomic unit. Anything that requires additions and deletions is broken out into separate "files". Last writer still wins.
Unfortunately, one must play this game when trying to build something like this on top of eventually consistent storage systems. If we run into serious problems, we can wrap the storagedrivers in a shared consistency layer but that would increase complexity and hinder registry cluster performance.