Include the design doc in repo

* Also add some new diagrams in the design doc
* Fix some issues in ADR

Signed-off-by: Jussi Kukkonen <[email protected]>

Author: Jussi Kukkonen

docs/adr/0010-repository-library-design.md

@@ -18,7 +18,7 @@ some issues (as do many other implementations):
 * The "library like" parts of the implementation make decisions that look like
   application decisions. As an example, repository_tool loads _every_ metadata
   file in the repository: this is fine for CLI that operates on a small
-  repository but is unlikely to be a good choice for PyPI.
+  repository but is unlikely to be a good choice for a large scale server.
 
 
 ## Decision Drivers
@@ -28,12 +28,12 @@ some issues (as do many other implementations):
 * Metadata API makes modifying metadata far easier than legacy code base: this
   makes significantly different designs possible
 * Not providing a "repository library" (and leaving implementers on their own)
-  may be a short term solution because of the previous point, but it does seem
-  like the project would benefit from some shared repository code and shared
-  repository design
+  may be a short term solution because of the previous point, but to make
+  adoption easier and to help adopters create safe implementations the project
+  would benefit from some shared repository code and a shared repository design 
 * Maintainability of new library code must be a top concern
 * Allowing a wide range of repository implementations (from CLI tools to
-  minimal in-memory implementations to large scale applications like Warehouse)
+  minimal in-memory implementations to large scale application servers)
   would be good: unfortunately these can have wildly differing requirements
 
 
@@ -64,9 +64,7 @@ forward: decision could be re-evaluated in a few months if not in weeks.
 ### No repository packages
 
 Metadata API makes editing the repository content vastly simpler. There are
-already repository implementations built with it (RepositorySimulator in
-python-tuf tests is an in-memory implementation, while
-repository-editor-for-tuf is an external CLI tool) so clearly a repository
+already repository implementations built with it[^1] so clearly a repository
 library is not an absolute requirement.
 
 Not providing repository packages in python-tuf does mean that external
@@ -76,9 +74,10 @@ different designs and hopefully find good ones in the end.
 
 That said, there are some tricky parts of repository maintenance (e.g.
 initialization, snapshot update, hashed bin management) that would benefit from
-having a canonical implementation. Likewise, a well designed library could make
-some repeated actions (e.g. version bumps, expiry updates, signing) much easier
-to manage.
+having a canonical implementation, both for easier adoption of python-tuf and
+as a reference for other implementations. Likewise, a well designed library
+could make some repeated actions (e.g. version bumps, expiry updates, signing)
+much easier to manage.
 
 ### repository_tool -like API
 
@@ -97,8 +96,7 @@ being a substantial amount of code that is only a good fit for one application.
 
 python-tuf could define a tiny repository API that
 * provides carefully selected core functionality (like core snapshot update)
-  but...
-* does not implement all repository actions itself, instead i makes it easy
+* does not implement all repository actions itself, instead it makes it easy
   for the application code to do them
 * leaves application details to specific implementations (examples of decisions
   a library should not always decide: "are targets stored with the repo?",
@@ -107,7 +105,7 @@ python-tuf could define a tiny repository API that
   date?", "which targets versions should be part of new snapshot?")
 
 python-tuf could also provide one or more implementations of this abstraction
-as examples -- this could include a repo.py- or repository_tool-like
+as examples -- this could include a _repo.py_- or _repository_tool_-like
 implementation.
 
 This could be a compromise that allows:
@@ -123,6 +121,16 @@ The approach does have some downsides:
 * A prototype has been implemented (see Links below) but the concept is still
   unproven
 
+More details in [Design document](../repository-library-design.md).
+
 ## Links
-[Design document for minimal repository abstraction](https://docs.google.com/document/d/1YY83J4ihztsi1Qv0dJ22EcqND8dT80AGTduwgh0trpY)
-[Prototype implementation of minimal repository abstraction](https://github.com/vmware-labs/repository-editor-for-tuf/)
+* [Design document for minimal repository abstraction](../repository-library-design.md)
+* [Prototype implementation of minimal repository abstraction](https://github.com/vmware-labs/repository-editor-for-tuf/)
+
+
+[^1]:
+    [RepositorySimulator](https://github.com/theupdateframework/python-tuf/blob/develop/tests/repository_simulator.py)
+    in python-tuf tests is an in-memory implementation, while
+    [repository-editor-for-tuf](https://github.com/vmware-labs/repository-editor-for-tuf)
+    is an external Command line repository maintenance tool.
+

docs/repository-library-design-ownership.jpg

@@ -0,0 +1,194 @@
+# Python-tuf repository API proposal: _minimal repository abstraction_
+
+This is an attachment to ADR 10: _Repository library design built on top of
+Metadata API_, and documents the design proposal in Dec 2020.
+
+## Design principles
+
+Primary goals of this repository library design are
+1. Support full range of repository implementations: from command line
+   “repository editing” tools to production repositories like PyPI
+2. Provide canonical solutions for the difficult repository problems but avoid
+   making implementation decisions
+3. Keep python-tuf maintenance burden in mind: less is more
+
+Why does this design look so different from both legacy python-tuf code and
+other implementations?
+* Most existing implementations are focused on a specific use case (typically a
+  command line application): this is a valid design choice but severely limits
+  goal #1
+* The problem space contains many application decisions. Many implementations
+  solve this by creating functions with 15 arguments: this design tries to find
+  another way (#2)
+* The Metadata API makes modifying individual pieces of metadata simpler. This,
+  combined with good repository API design, should enable more variance in
+  where things are implemented: The repository library does not have to
+  implement every little detail as we can safely let specific implementations
+  handle things, see goal #3
+* This variance means we can start by implementing a minimal design: as
+  experience from implementations is collected, we can then move implementation
+  details into the library (goals #2, #3)
+
+## Design
+
+![Design: Application and library components](repository-library-design-ownership.jpg)
+
+The design expects a fully functional repository application to contain code at
+three levels:
+* Repository library (abstract classes that are part of python-tuf)
+  * The Repository abstract class provides an ergonomic metadata editing API
+    for all code levels to use. It also implements some core edit actions like
+    snapshot update
+  * A small amount of related functionality is also provided (private key
+    management API, maybe repository validation)
+  * is a very small library: possibly a few hundred lines of code
+* Concrete Repository implementation (typically part of application code,
+  implements interfaces provided by the repository API in python-tuf)
+  * Contains the “application level” decisions that the Repository abstraction
+    requires to operate: examples of application decisions include
+    * _when should “targets” metadata next expire when it is edited?_
+    * _What is the current “targets” metadata version? Where do we load it 
+      from?_
+    * _Where to store current “targets” after editing? Should the previous
+      version be deleted from storage?_
+* Actual application
+  * Uses the Repository API to do the repository actions it needs to do
+
+For context here’s a trivial example showing what “ergonomic editing” means --
+this key-adding code could be in the application or in the python-tuf library:
+
+```python
+with repository.edit(“targets”) as targets:
+    # adds a key for role1 (as an example, arbitrary edits are allowed)
+    targets.add_key(“role1”, key)
+```
+
+This code loads current targets metadata for editing, adds the key to a role,
+and handles version and expiry bumps before persisting the new targets version.
+The reason for the context manager style is that it manages two things
+simultaneously:
+* Hides the complexity of loading and persisting metadata, and updating expiry
+  and versions from the editing code (by putting it in the repository
+  implementation – which may still be provided by the application)
+* Still allows completely arbitrary edits on the metadata in question: now the
+  library does not need to anticipate what application wants to do and on the
+  other hand library can still provide e.g. snapshot functionality without
+  knowing about the application decisions mentioned in previous point.
+
+Other designs do not seem to manage both of these.
+
+![Design: How components are used](repository-library-design-usage.jpg)
+
+The core idea here is that because editing is ergonomic enough, when new
+functionality (like “developer uploads new targets”) is added, _it can be added
+at any level_: the application might add a `handle_new_target_files()` method
+that adds a bunch of targets into the metadata, but one of the previous layers
+could offer that as a helper function as well: code in both cases would look
+similar as it would use the common editing interface.
+
+There are a few additional items worth mentioning:
+* Private key management: the Repository API should come with a “keyring
+  abstraction” -- a way for the application to provide roles’ private keys for
+  the Repository to use. Some implementations could be provided as well.
+* Validating repository state: the design is very much focused on enabling
+  efficient editing of individual metadata. Implementations are also likely to
+  be interested in validating (after some edits) that the repository is correct
+  according to client workflow and that it contains the expected changes. The
+  Repository API should provide some validation, but we should recognise that
+  validation may be implementation specific.
+* Improved metadata editing: There are a small number of improvements that
+  could be made to metadata editing. These do not necessarily need to be part
+  of the repository API: they could be part of Metadata API as well
+
+It would make sense for python-tuf to ship with at least one concrete
+Repository implementation: possibly a repo.py look alike. This implementation
+should not be part of the library but an example.
+
+## Details
+
+This section includes links to a Proof of Concept implementation in
+[repository-editor-for-tuf](https://github.com/vmware-labs/repository-editor-for-tuf/):
+it should not be seen as the exact proposed API but a prototype of the ideas.
+
+The ideas in this document map to POC components like this:
+
+| Concept | repository-editor-for-tuf implementation |
+|-|-|
+| Repository API | [librepo/repo.py](https://github.com/vmware-labs/repository-editor-for-tuf/blob/main/tufrepo/librepo/repo.py), [librepo/keys.py](https://github.com/vmware-labs/repository-editor-for-tuf/blob/main/tufrepo/librepo/repo.py)  |
+| Example of repository implementation | [git_repo.py](https://github.com/vmware-labs/repository-editor-for-tuf/blob/main/tufrepo/git_repo.py) |
+|Application code | [cli.py (command line app)](https://github.com/vmware-labs/repository-editor-for-tuf/blob/main/tufrepo/cli.py), [keys_impl.py (keyring implementation)](https://github.com/vmware-labs/repository-editor-for-tuf/blob/main/tufrepo/keys_impl.py) |
+| Repository validation | [verifier.py (very rough, not intended for python-tuf)](https://github.com/vmware-labs/repository-editor-for-tuf/blob/main/tufrepo/verifier.py)
+| Improved Metadata editing | [helpers.py](https://github.com/vmware-labs/repository-editor-for-tuf/blob/main/tufrepo/helpers.py) 
+
+
+### Repository API
+
+Repository itself is a minimal abstract class: The value of this class is in
+defining the abstract method signatures (most importantly `_load`, `_save()`,
+`edit()`) that enable ergonomic metadata editing. The Repository class in this
+proposal includes concrete implementations only for the following:
+* `sign()` -- signing without editing metadata payload
+* `snapshot()` -- updates snapshot and timestamp metadata based on given input.
+  Note that a concrete Repository implementation could provide an easier to use
+  snapshot that does not require input (see example in git_repo.py)
+
+More concrete implementations (see cli.py for examples) could be added to
+Repository itself but none seem essential at this point.
+
+The API requires a “Keyring” abstraction that the repository code can use to
+lookup a set of signers for a specific role. Specific implementations of
+Keyring could include a file-based keyring for testing, env-var keyring for CI
+use, etc. Some implementations should be provided in the python-tuf code base
+and more could be implemented in applications.
+
+_Prototype status: Prototype Repository and Keyring abstractions exist in
+librepo/repo.py._
+
+### Example of Repository implementation
+
+The design decisions that the included example `GitRepository` makes are not
+important but provide an example of what is possible:
+* Metadata versions are stored in files in git, with filenames that allow
+  serving the metadata directory as is over HTTP
+* Version bumps are made based on git status (so edits in staging area only
+  bump version once)
+* “Current version” when loading metadata is decided based on filenames on disk
+* Files are removed once they are no longer part of the snapshot (to keep
+  directory uncluttered)
+* Expiry times are decided based on an application specific metadata field
+* Private keys can be stored in a file or in environment variables (for CI use)
+
+Note that GitRepository implementation is significantly larger than the
+Repository interface -- but all of the complexity in GitRepository is really
+related to the design decisions made there.
+
+_Prototype status: The GitRepository example exists in git_repo.py._
+
+### Validating repository state
+
+This is mostly undesigned but something built on top of TrustedMetadataSet
+(currently ngclient component) might work as a way to easily check specific
+aspects like:
+* Is top-level metadata valid according to client workflow
+* Is a role included in the snapshot and the delegation tree
+
+It’s likely that different implementations will have different needs though: a
+command line app for small repos might want to validate loading all metadata
+into memory, but a server application hosting tens of thousands of pieces of
+metadata is unlikely to do so.
+
+_Prototype status: A very rough implementation exists in verifier.py : this is
+unlikely to be very useful_
+
+### Improved metadata editing
+
+Currently the identified improvement areas are:
+* Metadata initialization: this could potentially be improved by adding
+  default argument values to Metadata API constructors 
+* Modifying and looking up data about roles in delegating metadata
+  (root/targets): they do similar things but root and targets do not have
+  identical API. This may be a very specific use case and not interesting
+  for some applications
+
+_Prototype status: Some potential improvements have been collected in
+helpers.py_