We were using a breadth-first search which doesn't work as-is for
topological ordering in all cases. Instead of using Kahn's algorithm
which is a little complex, we switch to a depth-first search.
Instead of returning histories from `radicle-cob`, we return an
evaluated object which we build during initial graph traversal.
We do this so that branches can be correctly pruned when operations are
invalid at the application level. This way, new operations are not
building on top of invalid ones contained in the history.
To achieve this, we introduce a new `Evaluate` trait that is implemented
by all COBs, and we make change graph evaluation fallible.
Doing this means that we get `apply` errors returned for free when an
invalid transaction is applied, and no longer need to check for action
validity in multiple places.
We also implement a new `Dag::prune` method to avoid having to copy
graph nodes during traversal.
Instead of having two types, we simply have one, called `Entry`, which
we rename from `Change`.
There was no real benefit to having two almost identical types, but lots
of added complexity. This patch simplifies the crate by removing one of
the types.
> I'm going to be making a set of breaking changes to COBs in order to
stabilize the data formats. This is hopefully a one-time change that
bundles various breaking changes.
All COBs have been reworked: issue, patch, id.
The changes included are:
* Revise the assign and tag actions to take a single list of
assignees/tags to set, instead of an "add" and a "remove" list. This
makes API usage simpler when editing issues, and simplifies the apply
function
* Rename "tags" to "labels", and the tag action to label. This is
because tag is confusing in the context of git, as it could mean a git
tag. Using label removes that confusion.
* Use DIDs instead of PublicKeys for assignees -- this is more
future-proof
* Modify the manifest file format in the COB tree. Mainly, remove the
`history_type` key which is redundant, and use camelCase for keys
* Flatten the `Thread` actions into the parent action type
* Ensure that operations on redacted objects do not fail, since
redactions could have happened concurrently
* Use a consistent naming scheme for actions, using `.` as separator
* Consolidate comment types and remove `CodeComment`, by adding an
optional `location` field to `Comment`
* Add many placeholder actions that are not yet implemented
To preserve backwards compatibility, a `legacy` module is created with
the old `apply` function. When loading the manifest, we check whether it
is a legacy COB or a "stable" COB, and in the legacy case, use the
legacy code to materialize the state and then convert the object into
the stable type. Eventually, we'll delete the legacy code.
It turns out that the CRDT formed by the union of Git DAGs
is enough to guarantee everything we need for COBs.
This changes the following things:
* COB operations no longer need to be commutative
* COB histories are traversed in the same deterministic order on all
replicas
* It's now possible to implement RSMs on top of COBs, eg. scripting
* Lamport clocks have been removed
* `radicle-crdt` is no longer a dependency of `radicle`
* COBs are no longer instances of `Semilattice`
* The `Ops` type was removed in favor of having `Op` contain multiple
actions
We test to make sure that in the event of concurrent operations,
traversal order is deterministic.
This required some changes to the testing infrastructure. We're also
able to simplify the `Actor` type that now has redundant functionality.
Simplify graph traversal and evaluation by building in some of the
functionality into `radicle-dag`, namely the pruning fold.
We avoid building vectors of graph nodes this way, and simply iterate
over the graph in one go.
By introducing a small limitation: only allowing entries in the change
graph to be addressable, instead of individual operations; we
drastically simplify the CRDT implementation.
There are four advantages:
1. Op ids are just regular SHA-1s
2. There's no need for relative IDs, ops never refer to other ops within the
same commit
3. There's no need for a nonce, since commits can't collide, and neither can op IDs
4. `OpId` can just be an alias of `EntryId`
The disadvantage of course, is that we have to be mindful of how we
create op transactions, since each transaction creates an addressable
unit. For example, we must not include multiple patch revisions in the
same transaction.
By using `0` as the clock value for an empty state (without operations),
we simplify transaction code. Previously, there was no clock value for
empty states, which meant that we had to sometimes use an `Option`.
We're using petgraph to encode an operation based CRDT (each graph node
is an operation) as a DAG, and one of the important things to ensure
is that the final state is not influenced by the order of concurrent
operations. In this case a concurrent operation would mean two potential
graph traversal orders, as there's no direct edge between the two
concurrent ops. This is essential a "partial order" of operations.
To be able to test this, we'd want some kind of control over how
neighbors are iterated over I guess, when there's more than one possible
sort order. Say we have a graph like this:
┌────────b◄─────┐
▼ │
a d
▲ │
└────────c◄─────┘
There are two possible topological traverse orders: [a, b, c, d] and [a,
c, b, d]. Having a way to go through these different orders would
be super handy.
One option would be to allow random order traversal. This would allow us
to test all orders by running the test enough times to likely test all
permutations.
Since petgraph doesn't support this, we implement our own simple DAG in
`radicle-dag`, which implements random-order topological orders.
For now, we don't make explicit use of these improvements, we simply
replace the underlying graph with our own.
We weren't accounting for batches ops in an entry.
This change should account for it by adding N-1
to the clock value of an entry, where N is the number
of entries.
Signed-off-by: Alexis Sellier <alexis@radicle.xyz>
This is a minimal port of the collaborative objects code from
radicle-link[0].
A `CollaborativeObject` is made up of a `History` of `Change`s that
are expected to be stored in a `git` backend.
The library exposes a CRU series of functions `create`, `get`, `list`,
and `update`.
This implementation differs, to previous one, in that it does not
assume any kind of identity system -- its only assumption is that
identities have an identifier and can be addressed in the `git`
database. This means that it does not require an `author` to be passed
in and it only uses the `resource` (previously `authorizing_identity`)
to store its content-address in the `Change`.
It also foregoes any caching to simplify the port and not make any
early optimisation assumptions.
[0]: https://github.com/radicle-dev/radicle-link/tree/master/cob
Signed-off-by: Fintan Halpenny <fintan.halpenny@gmail.com>
cloudhead
Alexis Sellier
Alexis Sellier
Fintan Halpenny