Document approach to eventual consistency

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>

doc/architecture.md

@@ -2,3 +2,49 @@
 
 **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:
+
+1. Read-After-Write consistency (see this [article on
+   s3](http://shlomoswidler.com/2009/12/read-after-write-consistency-in-amazon.html)).
+2. [Write-Write Conflicts](http://en.wikipedia.org/wiki/Write%E2%80%93write_conflict).
+
+In reality, the registry must worry about these kinds of errors when doing the
+following:
+
+1. Accepting data into a temporary upload file may not have latest data block
+   yet (read-after-write).
+2. Moving uploaded data into its blob location (write-write race).
+3. Modifying the "current" manifest for given tag (write-write race).
+4. 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:
+
+1. Large writes are done to private upload directories. This alleviates most
+   of the corruption potential under multiple writers by avoiding multiple
+   writers.
+2. Constraints in storage driver implementations, such as support for writing
+   after the end of a file to extend it.
+3. Digest verification to avoid data corruption.
+4. Manifest files are stored by digest and cannot change.
+5. 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.