Delta: Structural differencing in Rust

The delta crate defines the trait Delta, along with a derive macro for auto-generating instances of this trait for most data types. Primarily the purpose of this trait is to offer a method, delta, by which two values of any type supporting that trait can yield a summary of the differences between them.

Note that unlike other crates that do data differencing (primarily between scalars and collections), delta has been written primarily with testing in mind. That is, the purpose of generating such change descriptions is to enable writing tests that assert the set of expected changes after some operation between an initial state and the resulting state. This goal also means that some types, like HashMap, must be differenced after ordering the keys first, so that the set of changes produced can be made deterministic and thus expressible as a test expectation.

To these ends, the function delta::assert_changes is also provided, taking two values of the same type along with an expected "change description" as returned by foo.delta(&bar). This function uses the pretty_assertions crate under the hood so that minute differences within deep structures can be easily seen in the failure output.

Quickstart

If you want to get started quickly with the delta crate to enhance unit testing, do the following:

1. Add the delta crate as a dependency, enabling features = ["derive"].
2. Derive the delta::Delta trait on as many structs and enums as needed.
3. Structure your unit tests to follow these three phases: a. Create the initial state or dataset you intend to test and make a copy of it. b. Apply your operations and changes to this state. c. Use delta::assert_changes between the initial state and the resulting state to assert that whatever happened is exactly what you expected to happen.

The main benefit of this approach over the usual method of "probing" the resulting state -- to ensure it changed as you expected it to-- is that it asserts against the exhaustive set of changes to ensure that no unintended side-effects occurred beyond what you expected to happen. In this way, it is both a positive and a negative test: checking for what you expect to see as well as what you don't expect to see.

The Delta trait

The Delta trait has two associated types and two methods, one pair corresponding to value descriptions and the other to value changes:

pub trait Delta {
    type Desc: PartialEq + Debug;
    fn describe(&self) -> Self::Desc;

    type Change: PartialEq + Debug;
    fn delta(&self, other: &Self) -> Changed<Self::Change>;
}

Descriptions: the Desc associated type

Value descriptions (the Desc associated type) are needed because value hierarchies can involve many types. Perhaps some of these types implement PartialEq and Debug, but not all. To work around this limitation, the Delta derive macro creates a "mirror" of your data structure with all the same constructors ands field, but using the Desc associated type for each of its contained types.

#[derive(Delta)]
struct Foo {
  bar: Bar,
  baz: Baz
}

This generates a description that mirrors the original type, but using type descriptions rather than the types themselves:

struct FooDesc {
  bar: <Bar as Delta>::Desc,
  baz: <Baz as Delta>::Desc
}

You may also choose an alternate description type, such as a reduced form of a value or some other type entirely. For example, complex structures could describe themselves by the set of changes they represent from a Default value. This is so common, that it's supported via a compare_default macro attribute provided by delta:

#[derive(Delta)]
#[compare_default]
struct Foo { /* ...lots of fields... */ }

impl Default for Foo { /* ... */ }

For scalars, the Desc type is the same as the type it's describing, and these are called "self-describing".

There are other macro attributes provided for customizing things even further, which are covered below, beginning at the section on Structures.

Changes: the Change associated type

When two values of a type differ, this difference gets represented using the associated type Change. Such values are produced by the delta method, which actually returns Changed<Change> since the result may be either Changed::Unchanged or Changed::Changed(_changes_).[^option]

[^option] Changed is just a different flavor of the Option type, created to make changesets clearer than just seeing Some in various places.

The primary purpose of a Change value is to compare it to a set of changes you expected to see, so design choices have been made to optimize for clarity and printing rather than, say, the ability to transform one value into another by applying a changeset. This is entirely possible give a dataset and a change description, but no work has been done to achieve this goal.

How changes are represented can differ greatly between scalars, collections, structs and enums, so more detail is given below in the section discussing each of these types.

Scalars

Delta traits have been implemented for all of the basic scalar types. These are self-describing, and use a Change structure named after the type that holds the previous and changed values. For example, the following assertions hold:

assert_changes(&100, &100, Changed::Unchanged);
assert_changes(&100, &200, Changed::Changed(I32Change(100, 200)));
assert_changes(&true, &false, Changed::Changed(BoolChange(true, false)));
assert_changes(
    &"foo",
    &"bar",
    Changed::Changed(StringChange("foo".to_string(), "bar".to_string())),
);

Vec and Set Collections

The set collections for which Delta has been implemented are: Vec, HashSet, and BTreeSet.

The Vec uses Vec<VecChange> to report all of the indices at which changes happened. Note that it cannot detect insertions in the middle, and so will likely report every item as changed from there until the end of the vector, at which point it will report an added member.

HashSet and BTreeSet types both report changes the same way, using the SetChange type. Note that in order for HashSet change results to be deterministic, the values in a HashSet must support the Ord trait so they can be sorted prior to comparison. Sets cannot tell when specific members have change, and so only report changes in terms of SetChange::Added and SetChange::Removed.

Here are a few examples, taken from the delta_test test suite:

// Vectors
assert_changes(
    &vec![1 as i32, 2],
    &vec![1 as i32, 2, 3],
    Changed::Changed(vec![VecChange::Added(2, 3)]),
);
assert_changes(
    &vec![1 as i32, 3],
    &vec![1 as i32, 2, 3],
    Changed::Changed(vec![
        VecChange::Changed(1, I32Change(3, 2)),
        VecChange::Added(2, 3),
    ]),
);
assert_changes(
    &vec![1 as i32, 2, 3],
    &vec![1 as i32, 3],
    Changed::Changed(vec![
        VecChange::Changed(1, I32Change(2, 3)),
        VecChange::Removed(2, 3),
    ]),
);
assert_changes(
    &vec![1 as i32, 2, 3],
    &vec![1 as i32, 4, 3],
    Changed::Changed(vec![VecChange::Changed(1, I32Change(2, 4))]),
);

// Sets
assert_changes(
    &HashSet::from(vec![1 as i32, 2].into_iter().collect()),
    &HashSet::from(vec![1 as i32, 2, 3].into_iter().collect()),
    Changed::Changed(vec![SetChange::Added(3)]),
);
assert_changes(
    &HashSet::from(vec![1 as i32, 3].into_iter().collect()),
    &HashSet::from(vec![1 as i32, 2, 3].into_iter().collect()),
    Changed::Changed(vec![SetChange::Added(2)]),
);
assert_changes(
    &HashSet::from(vec![1 as i32, 2, 3].into_iter().collect()),
    &HashSet::from(vec![1 as i32, 3].into_iter().collect()),
    Changed::Changed(vec![SetChange::Removed(2)]),
);
assert_changes(
    &HashSet::from(vec![1 as i32, 2, 3].into_iter().collect()),
    &HashSet::from(vec![1 as i32, 4, 3].into_iter().collect()),
    Changed::Changed(vec![SetChange::Added(4), SetChange::Removed(2)]),
);

Note that if the first VecChange::Change above had used an index of 1 instead of 0, the resulting failure would look something like this:

running 1 test
test test_delta_bar ... FAILED

failures:

---- test_delta_bar stdout ----
thread 'test_delta_bar' panicked at 'assertion failed: `(left == right)`

Diff < left / right > :
 Changed(
     [
         Change(
<            1,
>            0,
             I32Change(
                 100,
                 200,
             ),
         ),
     ],
 )

', /Users/johnw/src/delta/delta/src/lib.rs:19:5
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace


failures:
    test_delta_bar

Map Collections

jww (2021-11-01): TODO

Structures

Differencing arbitrary structures was the original motive for creating delta. This is made feasible using a Delta derive macro that auto-generates code needed for such comparisons. The purpose of this section is to explain how this macro works, and the various attribute macros that can be used to guide the process. If all else fails, manual trait implementations are always an alternative.

For the purpose of the following sub-sections, we consider the following structure:

#[derive(Delta)]
struct Foo {
  bar: Bar,
  baz: Baz,
  #[delta_ignore]
  baz: Box<dyn FnOnce(u32)>
}

The first attribute macro you'll notice that can be applied to individual fields is #[delta_ignore], which must be used if the type in question cannot be compared for differences.

TODO: jww (2021-11-01): Allow the Desc and Change suffixes to both be changed.

TODO: jww (2021-11-01): For each multi-field variant in an enum, generate a helper Change struct and set that variant's type for the enum's Change to be Vec<Change>.

TODO: jww (2021-11-01): Provide an attribute macro #[delta_wrap] that defines a wrapping type that can be used for comparison. When the field is encountered during delta, construct a temporary value using the wrapper and then call delta on that.

TODO: jww (2021-11-01): Provide an attribute macro #[delta_view(function)] for defining synthetic properties that receive &self as an argument and return a type implementing Delta that can be differenced.

Deriving Delta for structs: the Desc type

By default, deriving Delta for a structure will create a "mirror" of that structure, with all the same fields, but replacing every type T with <T as Delta>::Desc:

struct FooDesc {
  bar: <Bar as Delta>::Desc,
  baz: <Baz as Delta>::Desc
}

This process can be influenced using several attribute macros.

compare_default

When the #[compare_default] attribute macro is used, the Desc type is defined to be the same as the Change type, with the describe method being implemented as a comparison against the value of Default::default():

type Desc = Self::Change;

fn describe(&self) -> Self::Desc {
    Foo::default().delta(self).unwrap_or_default()
}

type Change = Vec<FooChange>;

Note that changes for structures are always a vector, since this allows changes to be reported separately for each field. More on this in the following section.

no_description

If you want no description at all for a type, since you only care about how it has changed and never want to report a description of the value in any other context, then you can use #[no_description]. This sets the Desc type to be unit, and the describe method accordingly:

type Desc = ();

fn describe(&self) -> Self::Desc {
    ()
}

It is assumed that when this is appropriate, such values will never appear in any change output, so consider a different approach if you see lots of units turning up.

describe_type and describe_body

You can have more control over description by specifying exactly the text that should appear for the Desc type and the body of the describe function. Basically, for the following definition:

#[derive(Delta)]
#[describe_type(T)]
#[describe_body(B)]
struct Foo {
  bar: Bar,
  baz: Baz
}

The following is generated:

type Desc = T;

fn describe(&self) -> Self::Desc {
    B
}

This also means that the expression argument passed to describe_body may reference the self parameter. Here is a real-world use case:

#[cfg_attr(feature = "delta",
           derive(delta::Delta),
           describe_type(String),
           describe_body(self.to_string()))]

This same approach could be used to represent large blobs of data by their checksum hash, for example, or large data structures that you don't need to ever display by their Merkle root hash.

Deriving Delta for structs: the Change type

By default for structs, deriving Delta creates an enum with variants for each field in the struct, and it represents changes using a vector of such values. This means that for the following definition:

#[derive(Delta)]
struct Foo {
  bar: Bar,
  baz: Baz
}

The Change type is defined to be Vec<FooChange>, with FooChange as follows:

#[derive(PartialEq, Debug)]
enum FooChange {
    Bar(<Bar as Delta>::Change),
    Baz(<Baz as Delta>::Change),
}

impl Delta for Foo {
    type Desc = FooDesc;
    type Change = Vec<FooChange>;
}

Here is an abbreviated example of how this looks when asserting changes:

assert_changes(
    &initial_foo, &later_foo,
    Changed::Changed(vec![
        FooChange::Bar(...),
        FooChange::Baz(...),
    ]));

If the field hasn't been changed it won't appear in the vector, and each field appears at most once. The reason for taking this approach is that structures with many, many fields can be represented by a small change set if most of the other fields were left untouched.

Special case: Unit structs

If a struct has no fields it can never change, and so only a unitary Desc type is generated.

Special case: Singleton structs

If a struct has only one field, there is no reason to specify changes using a vector, since either the struct is unchanged or just that one field has changed. For this reason, singleton structs optimize away the vector and use type Change = [type]Change in their Delta derivation, rather than type Change = Vec<[type]Change> as for multi-field structs.

delta_public and delta_private

By default, the auto-generated Desc and Change types have the same visibility as their parent. This may not be appropriate, however, if you want to keep the original data type private but allow exporting of descriptions and change sets. To support this -- and the converse -- you can use #[delta_public] and #[delta_private] to be explicit about the visibility of these generated types.

Enumerations

Enumerations are handled quite differently from structures, for the main reason that while a struct is always a product of fields, an enum can be more than a sum of variants -- but also a sum of products.

To unpack that a bit: By a product of fields, it is meant that a struct is a simple grouping of typed fields, where the same fields are available for every value of such a structure.

Meanwhile, an enum is a sum, or choice, among variants, but some of these variants can themselves contain groups of fields, as though there were an unnamed structure embedded in the variant. Consider the following enum, which will be used for all the following examples:

#[derive(Delta)]
enum MyEnum {
    One(bool),
    Two { two: Vec<bool>, two_more: Baz },
    Three,
}

Here we see variant that has a variant with no fields (Three), one with unnamed fields (One), and one with named fields like a usual structure (Two). The problem, though, is that these embedded structures are never represented as independent types, so we can't define Delta for them and just compute the differences between the enum arguments. Nor can we just create a copy of the field type with a real name and generate Delta for it, because not every value is copyable or clonable, and it gets very tricky to auto-generate a new hierarchy built out fields with reference types all the way down...

Instead, the following gets generated, which can end up being a bit verbose, but captures the full nature of any differences:

enum MyEnumChange {
    BothOne(<bool as Delta>::Change),
    BothTwo {
        two: Changed<<Vec<bool> as Delta>::Change>,
        two_more: Changed<Baz as Delta>::Change
    },
    BothThree,
    Different(<MyEnum as Delta>::Desc, <MyEnum as Delta>::Desc),
}

Note that variants with singleton fields do not use Change, since that information is already reflected when the variant is reported as having changed at all using, for example, BothOne. In the case of BothTwo, each of the field types is wrapped in Changed because it's possible that either one or both of the fields may changed.

Special case: Empty enums

If a enum has no variants it cannot be constructed, so both the Desc or Change types are omitted and it is always reported as unchanged.

Unions

Unions cannot derive Delta instances at the present time.