diff --git a/adr/001_design_principles.md b/adr/001_design_principles.md index f7c4de5..43dd41f 100644 --- a/adr/001_design_principles.md +++ b/adr/001_design_principles.md @@ -30,13 +30,15 @@ Do not equate them. ## Consequences 1. The repository should support both design principles. - - The `README.md` and `doc.go` should reflect these principles. - - The contributing guide and pull request template should require these principles. + - [ ] Update the `README.md` and `doc.go` to reflect these principles. + - [ ] Update the contributing guide and pull request template to require these principles are met. 2. The repository should contain a living document, describing the interface differences between `go-cuckoo` and `map`. - - Its construction should uncover any current incongruencies in the interfaces. - - I should prioritize limiting any incongruencies. - - The document should be visible from the `README.md`. -3. An analysis of the familiarity of `go-cuckoo`'s interface should be made. - - Unlike the analysis of congruency, this should be a one time document. - Familiarity is implicit to users, and does not need to be referenced. - But, any rationale should be documented in commit messages, or future ADRs. + I should prioritize limiting any incongruencies. + - [ ] Produce the first draft to uncover any current incongruencies. + - [ ] Link the document to the `README.md`. +3. Analyze the familiarity of `go-cuckoo`'s current interface. + Unlike the analysis of congruency, this should be a one time document. + Familiarity is implicit to users, and does not need to be referenced. + But, any rationale should be documented in commit messages, or future ADRs. + - [ ] Produce the analysis document. + - [ ] Resolve any issues found. diff --git a/adr/001_interface_design.md b/adr/001_interface_design.md new file mode 100644 index 0000000..28a2bf5 --- /dev/null +++ b/adr/001_interface_design.md @@ -0,0 +1,643 @@ +# Designing an Idiomatic API Interface + +- [Designing an Idiomatic API Interface](#designing-an-idiomatic-api-interface) + - [Current State](#current-state) + - [Interface of the Built-in Map](#interface-of-the-built-in-map) + - [Interface of `go-cuckoo`](#interface-of-go-cuckoo) + - [Determining Congruency](#determining-congruency) + - [Determining Familiarity](#determining-familiarity) + - [Target State](#target-state) + - [Solving Congruency](#solving-congruency) + +We (the maintainers) built `go-cuckoo`'s API interface without design intent. +Up until now, we paid more attention implementing the underlying functionality of the cuckoo hashing. + +With the fundamentals of the algorithm built, we should revisit the interface. +It should align closer to the following principles: + +- **Congruency** + A `go-cuckoo` table should have the same core functionality as Go's built-in map. + +- **Familiarity** + A `go-cuckoo` table should behave similarly to Go's standard map, so users will intuitively know how to use it. + In effect, its users will carry less cognitive load. + +## Current State + +### Interface of the Built-in Map + +Listed below is every interface provided by Go to the built-in map object. +Also included, are the functions from the package `maps` in the standard library. + +
+Interfaces + +| # | Built-in Interface | Description | +| --- | ---------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- | +| 1 | `m := make(map[K]V)` | Returns an empty map using the built-in `make()` function. | +| 2 | `m := make(map[K]V, hint)` | Returns an empty map using `make()`, with a capacity 'hint'. This hint is how many items the map expects to hold, _not_ a measure of how large it is. | +| 3 | `m := map[K]V{...}` | Returns a map, which may be filled with entries in the ellipsis (optional). | +| 4 | `var m map[K]V` | Defines an empty _variable_ that holds a map. This differs from #1 because `m` is uninitialized (nil) here. | +| 5 | `m[k] := v` | Assigns the value of `k` to `v`. | +| 6 | `v := m[k]` | Returns the value of `k` if it exists. Otherwise, `v` is uninitialized. | +| 7 | `v, ok := m[k]` | Similar to #6, except `ok` is equal to whether `v` is initialized. This is comma-ok notation. | +| 8 | `for k, v := range m` | Iterates over every key-value pair in `m`. The order is random. | +| 9 | `delete(m, k)` | Unassigns the value `k`. Returns no value. | +| 10 | `clear(m)` | Unassigns all keys in `m`. Returns no value. | +| 11 | `n := len(m)` | Returns the number of entries in `m`. If nil, `m` returns 0. | +| 12 | `m2 := maps.Clone(m)` | Returns a copy of `m`. | +| 13 | `maps.Copy(dst, src)` | Assigns every entry of `src` in `dst`. | +| 14 | `ok := maps.Equal(m1, m2)` | Returns true iff `m1` and `m2` the same entries. | +| 15 | `ok := maps.EqualFunc(m1, m2, fn)` | Like #14, but with a custom comparator for non-comparable values. | +| 16 | `maps.DeleteFunc(m, fn)` | Removes every entry in `m` which satisfies `fn`. Returns no value. | +| 17 | `it2 := maps.All(m)` | Returns an 2D iterator over every key-value pair. | +| 18 | `it := maps.Keys(m)` | Returns an iterator over every key. | +| 19 | `it := maps.Values(m)` | Returns an iterator over every value. There can be duplicates. | +| 20 | `m := maps.Collect(seq)` | Returns a map, with every entry defined in a 2D iterator over key-value pairs. | +| 21 | `maps.Insert(m, seq)` | Assigns to `m` all key-value pairs in 2D iterator `seq`. Returns no value. | + +
+ +### Interface of `go-cuckoo` + +On the other hand, here is the current contract for `go-cuckoo`. + +
+Interfaces + +| # | `go-cuckoo` Interface | Description | +| --- | -------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------- | +| 1 | `m := New(opts...)` | Creates a table using the default hash and equal function. The options configure its behavior. Confined to comparable keys. | +| 2 | `m := NewBy(keyFunc, opts...)` | Like #1, but allows any key type. A `keyFunc` is used to derive a comparable key. | +| 3 | `m := NewCustom(hashA, hashB, equalFunc, opts...)` | Like #1, but allows control over the hashes used to allow any key type. An `equalFunc` determines key equality. | +| 4 | `seq := m.Entries()` | Returns an unordered 2D iterator of all key-value pairs in the table. | +| 5 | `v := m.Find(k)` | Removes the value for `k`. Returns true if `k` existed. | +| 6 | `v, ok := m.Get(k)` | Returns the value for `k` in the table. Also, returns true if the `k` exists, otherwise false. When false, `v` is undefined. | +| 7 | `ok := m.Has(k)` | Returns true if `k` is in the table. | +| 8 | `err := m.Put(k, v)` | Sets value `v` for key `k`. Otherwise, returns error. | +| 9 | `n := m.Size()` | Returns the number of items in `m`. | +| 10 | `str := m.String()` | Returns `m` as a string in the format "table[k1:v1 k2:v2 ...]". | +| 11 | `cap := m.TotalCapacity()` | Returns how many slots `m` has allocated. | +| 12 | `ok := m.Drop(k)` | Removes `k` from the table. Returns whether the key had existed. | + +
+ +### Determining Congruency + +So, how does the core functionality compare? +Listed below is an analysis of every interface in Go's standard map. +Each is compared against what `go-cuckoo` offers, and categorized into the following groups: + +- ✅ Covered: an analog exists. +- ⚠️ Partial: workaround available. +- ❌ Gap: no analog yet; addressed in [Target State](#solving-congruency). + +Specifically, here we are checking for functionality. +Is there functionality that this offers which `go-cuckoo` does not? +We are checking accessibility, but not discoverability. +The latter will be considered later. + +
+✅ m := make(map[K]V) + +The analog is `m := New()`. + +
+ +
+⚠️ m := make(map[K]V, hint) + +This has no simple analog. + +It is close to `m := New(Capacity(hint))`, but it assigns starting capacity, not expected size. +For the built-in map, these are two separate things. + +- Capacity is an internal measure, used to optimize space/speed. + It is hidden from the user because it depends on the underlying implementation, which may change. +- Expected size requires the map must hold a number of items before resizing. + This is tangeable and agnostic to implementation, hence why it is given to the user. + +In short, this interface defines expected size, but `Capacity()` defines capacity. + +
+ +
+❌ m := map[K]V{...} + +This has no simple analog, the closest being: + +```go +m := New[K, V]() +for k, v := range startingEntries { + m.Put(k, v) +} +``` + +It is idiomatic, but far less ergonomic. + +
+ +
+✅ var m map[K]V + +The analog is `var m Table[K, V]`. + +
+ +
+✅ m[k] := v + +The analog is `err := m.Put(k, v)`. + +
+ +
+✅ v := m[k] + +The analog is `v := m.Find(k)`. + +
+ +
+✅ v, ok := m[k] + +The analog is `v, ok := m.Get(k)`. + +
+ +
+✅ for k, v := range m + +The analog is `for k, v := range m.Entries()`. + +
+ +
+✅ delete(m, k) + +The analog is `ok := m.Drop(k)`. + +
+ +
+❌ clear(m) + +There is no analog. + +The easiest may to do this is to delete all items individually: + +```go +for k := range m.Entries() { + m.Drop(k) +} +``` + +
+ +
+✅ n := len(m) + +The analog is `n := m.Size()`. + +
+ +
+❌ m2 := maps.Clone(m) + +There is no analog. + +The easiest way to do this currently is to make a new map, and manually add the items. + +```go +m2 := cuckoo.Table[K, V]() + +for k, v := range m.Entries() { + m2.Put(k, v) +} +``` + +This gets complicated by the various options available to the user. +Furthermore, any custom `EqualFunc`, `keyFunc` or `Hash` is not transferred. + +
+ +
+❌ maps.Copy(dst, src) + +There is no analog. + +The simplest way to do this is with a for-loop. + +```go +for k, v := range src.Entries() { + dst.Put(k, v) +} +``` + +
+ +
+❌ ok := maps.Equal(m1, m2) + +There is no analog. + +Users have to manually check the key-value pairs to determine equality. + +
+ +
+❌ ok := maps.EqualFunc(m1, m2, fn) + +There is no analog. + +Users have to manually check the key-value pairs to determine equality. + +
+ +
+❌ maps.DeleteFunc(m, fn) + +There is no analog. + +Users have to manually delete keys. + +
+ +
+✅ it2 := maps.All(m) + +The analog is `it2 := m.Entries()`. + +
+ +
+⚠️ it := maps.Keys(m) + +There is no simple analog. + +A close neighbor is `it2 := m.Entries()`. +Users can use this in a for-loop, and pick out just the keys: + +```go +for k := range m.Entries() { + // ... +} +``` + +
+ +
+⚠️ it := maps.Values(m) + +There is no simple analog. + +A close neighbor is `it2 := m.Entries()`. +Users can use this in a for-loop, and pick out just the values: + +```go +for _, v := range m.Entries() { + // ... +} +``` + +
+ +
+❌ m := maps.Collect(seq) + +There is no analog. + +
+ +
+❌ maps.Insert(m, seq) + +There is no analog. + +
+ +### Determining Familiarity + +We can categorize all existing table functionality by each interface's familiarity: + +- ✅ Idiomatic: is clear, intuitive, and easily understood. +- ❌ Non-idiomatic: is misleading; addressed in [Target State](#solving-congruency). + +
+✅ m := New(opts...) + +Criteria: + +Noun/adjective form — Go constructors use NewX where X is a noun or adjective, not a verb or past participle (NewReaderSize, not NewSized) +Names what the user provides — the suffix should hint at the distinguishing parameter (NewBufferString tells you it takes a string) +Progression is readable — the three names together should imply simple → intermediate → advanced +Not misleadingly generic — NewWith or NewConfig could mean anything + +
+ +
+m := NewBy(keyFunc, opts...) + +- Use `NewKeyed()`. + +
+ +
+m := NewCustom(hashA, hashB, equalFunc, opts...) + +- Use `NewHashed()`. + +
+ +
+❌ seq := m.Entries() + +- Call it `All()`. + +
+ +
+v := m.Find(k) + +- Call it `m.Lookup()`. The name `m.Find` implies a search algorithm. + +
+ +
+v, ok := m.Get(k) + +
+ +
+ok := m.Has(k) + +
+ +
+❌ err := m.Put(k, v) + +- Call it `Set()`. +- No built-in library consensus, but 3rd party packages prefer `Set()`. + +
+ +
+❌ n := m.Size() + +- Call it `Len()`. +- Size is a Java idiom. + +
+ +
+✅ str := m.String() + +
+ +
+❌ cap := m.TotalCapacity() + +- Remove. This is an implementation detail. + +
+ +
+❌ ok := m.Drop(k) + +- Call it `Delete()`. + +
+ +## Target State + +### Solving Congruency + +We should make the following changes to accomodate for congruency: + +
+ok := maps.EqualFunc(m1, m2, fn) + +We should implement a new function: + +```go +func EqualFunc[K, V1, V2 any](t1 *Table[K, V1], t2 *Table[K, V2], eq func(V1, V2) bool) bool +``` + +This function is free, and not bound as a receiver function. +(It is called `cuckoo.Equal(t1, t2)`, not `t1.Equals(t2)`.) +The latter implies `t1` has authority, when in fact neither do. + +We define equality as: + +1. Neither table has a key the other doesn't. +2. Each key has the same value in each table. + Parameter `eq` determines this equality. + +Custom `EqualFunc`'s complicate this, as they modulate key identity in tables. +If two tables may differ on whether two keys are different, this function might break. +So, we must assume that: + +- Both tables have `EqualFunc`'s which 'agree' on the identity of the keys present in the tables. + Agreement is defined as: if two keys are distinct in one table, they are distinct in the other. + +The name `EqualFunc` is already taken by `EqualFunc[K, V]`: an alias for `func(a, b K) bool`. +Inlining `EqualFunc[K, V]` would solve this problem. +We will move the documentation attached to it to `DefaultEqualFunc`. + +
+ +
+ok := maps.Equal(m1, m2) + +We should implement a new function, to conform with the standard library: + +```go +func Equal[K any, V comparable](t1, t2 *Table[K, V]) bool +``` + +It uses the same equality check as in `EqualFunc`. +Once again, the function is free because it is symmetric. + +
+ +
+maps.Insert(m, seq) + +We should implement a new receiver for the table: + +```go +func (t *Table[K, V]) Insert(seq iter.Seq2[K, V]) error +``` + +A receiver fits better even though `maps.Insert` is a free function, because copying it is asymmetric. +Map `dst` receives entries from map `src`. +It's only free because Go's standard map is built into the language, and so cannot have receivers. + +In terms of naming, `t.Extend` is more accurate, and has precedent in [Python](docs.python.org/3/tutorial/datastructures.html#more-on-lists) and [Rust](https://doc.rust-lang.org/std/iter/trait.Extend.html). +When [adding iterator function](https://github.com/golang/go/issues/61900) to the `maps` package, the Go team chose to frame it as 'sources' and 'sinks'. +With this model, `maps.Insert` made more sense than `maps.Extend`. +Ultimately, `t.Insert()` is a better choice to be consistent with `maps`. + +
+ +
+maps.Copy(dst, src) + +We should implement a new receiver for the table: + +```go +func (t *Table[K, V]) Copy(src *Table[K, V]) error +``` + +It's functionality should match that of `t.Insert()`. + +A receiver fits better even though `maps.Copy` is a free function, 'copying' it is asymmetric: `dst` is writen into by `src`. +It is only free because Go's standard map is built into the language, and so cannot have receivers. + +The name `t.Merge()` might be more accurate, but it does work because: + +- `t.Copy()` matches Go's built-in `copy()`, and `io.Copy()`. The Go team used [the same logic](https://github.com/golang/go/discussions/47330#discussioncomment-1167799) to name `maps.Copy()`. + In this case, `t.Merge()` would be an outlier. +- `t.Merge()` implies some sort of conflict-resolution, when there is not. + It simply overwrites the values. + +
+ +
+maps.DeleteFunc(m, fn) + +We should implement a new receiver for the table: + +```go +func (t *Table[K, V]) DeleteFunc(del func(K, V) bool) +``` + +It would have the same functionality as `maps.DeleteFunc`. + +A free function could work here, but `t` has clear authority over `del`. +Other than being consistent with the `maps` package, `t.DeleteFunc` follows the Go convention of appending `Func` to higher-order equivalents of functions. +This trumps names like `t.DeleteIf`, which lend more to [Java](https://docs.oracle.com/javase/8/docs/api/java/util/ArrayList.html#removeIf-java.util.function.Predicate-) or [C++](https://en.cppreference.com/cpp/algorithm/remove). +The word `Delete` is also convention, tying back to the built-in `delete()`. + +
+ +
+m := maps.Collect(seq) + +We should implement a new constructor. + +```go +func Collect[K comparable, V any](seq iter.Seq2[K, V]) (*Table[K, V], error) +``` + +It would create a `New()` table, and insert all entries in `seq`. + +This reveicer only supports the standard table constructor, with comparable keys. +It is tempting to add `CollectBy` or `CollectCustom` to support all table types, but doing so would pollute the public interface. + +It would be just one more line to initialize the table and then call `t.Insert` directly: + +```go +t := // ... +err := t.Insert(seq) +``` + +
+ +
+m := map[K]V{...} + +We should make a new constructor, because entries are generic. +So, creating an option with inialized entries doesn't work. + +With the previous additions, users have a few options. +If they want to use a `New()` table, `t.Collect` matches well: + +```go +t, err := cuckoo.Collect(func(yield func(K, V) bool) { + yield(key1, val1) + yield(key2, val2) +}) +``` + +For `NewCustom()` or `NewBy()` tables, users can call `t.Insert` after initialization: + +```go +t := // ... +err := t.Insert(func(yield func(K, V) bool) { + yield(key1, val1) + yield(key2, val2) +}) +``` + +It is one more line. +But, the alternative is polluting the public interface with corresponding `*WithEntries` constuctors. + +
+ +
+m := make(map[K]V, hint) + +We should add a new option: + +```go +func ExpectedSize(n int) Option +``` + +When fed to a table, it will allocate enough space to hold `n` entries without a resize. + +
+ +
+clear(m) + +We should implement a new receiver: + +```go +func (t *Table[K, V]) Clear() +``` + +It will remove all entries from the table. + +
+ +
+m2 := maps.Clone(m) + +We should implement a matching function: + +```go +func (t *Table[K, V]) Clone() *Table[K, V] +``` + +Also, it will copy the hash, equality function, and options used in the table. + +
+ +
+it := maps.Keys(m) + +We should implement a matching function: + +```go +func (t *Table[K, V]) Keys() iter.Seq[K] +``` + +It is tempting to just have `All()`, but it returns a `Seq2`, not a `Seq`. +There is no iterator adaptor between `Seq` and `Seq2`, and will not be for the foreseeable future. +This function, while it feels superfluous, is required. + +
+ +
+it := maps.Values(m) + +We should implement a matching function: + +```go +func (t *Table[K, V]) Values() iter.Seq[V] +``` + +For the same reason we need `Keys()`, we also need `Values()`. + +