|
| 1 | +# Immutable Collections |
| 2 | + |
| 3 | +* **Type**: Standard/kotlinx Library API proposal |
| 4 | +* **Author**: Ilya Gorbunov |
| 5 | +* **Status**: Submitted |
| 6 | +* **Prototype**: In progress |
| 7 | +* **Discussion**: In the [issues](https://github.com/Kotlin/kotlinx.collections.immutable/issues) of this repository |
| 8 | + |
| 9 | +## Summary |
| 10 | + |
| 11 | +Provide interfaces to represent truly immutable collections and maps. |
| 12 | + |
| 13 | +Provide implementation of those interfaces. |
| 14 | + |
| 15 | +## Similar API review |
| 16 | + |
| 17 | +* [.NET immutable collections](https://msdn.microsoft.com/en-us/library/mt452182.aspx) |
| 18 | +* [Scala immutable collections](http://docs.scala-lang.org/overviews/collections/concrete-immutable-collection-classes) |
| 19 | + |
| 20 | +## Description |
| 21 | + |
| 22 | +There are two families of collection interfaces in the Standard Library: |
| 23 | + |
| 24 | +* read-only interfaces, which only allow to read elements from a collection; |
| 25 | +* mutable interfaces, which additionally allows to modify elements in a collection. |
| 26 | + |
| 27 | +Kotlin mostly doesn't provide its own collection implementations. |
| 28 | +Its collection interfaces are mapped to the standard JDK collection interfaces |
| 29 | +and implemented by the JDK collection implementations, such as `ArrayList`, `HashSet`, `HashMap` and other |
| 30 | +classes from `java.util` package. |
| 31 | + |
| 32 | +It is proposed to provide immutable collection interfaces which extend read-only interfaces |
| 33 | +and specify by the contract the real immutability of their implementors. |
| 34 | + |
| 35 | +### Modification operations |
| 36 | + |
| 37 | +Immutable collections could have same modification operations as their mutable counterparts, such as `add(element)`, |
| 38 | +`remove(element)`, `put(key, value)` etc., but unlike mutating operations they do not mutate the receiver, but instead |
| 39 | +return new immutable collection instance with the modification applied. |
| 40 | + |
| 41 | +In case if no modification is made during the operation, for example, adding an existing element to a set or clearing |
| 42 | +already empty collection, it should return the same instance of the immutable collection. |
| 43 | + |
| 44 | +### Builders |
| 45 | + |
| 46 | +Often there is a need to perform several subsequent modification operations on an immutable collection. It can be done by |
| 47 | +- chaining operations together: |
| 48 | + ``` |
| 49 | + collection = collection.add(element).add(anotherElement) |
| 50 | + ``` |
| 51 | +- applying operations one by one: |
| 52 | + ``` |
| 53 | + collection += element |
| 54 | + collection += anotherElement |
| 55 | + ``` |
| 56 | +- using collection builders. |
| 57 | +
|
| 58 | +An immutable collection builder is a wrapper around an immutable collection, which exposes its modification operations |
| 59 | +within the mutable collection interface. For example, `ImmutableSet.Builder<T>` extends `MutableSet<T>`. |
| 60 | +
|
| 61 | +An immutable collection could be transformed to a builder with `builder()` method, and a builder could be |
| 62 | +transformed back to the immutable collection with its `build()` method. |
| 63 | +
|
| 64 | +``` |
| 65 | +val builder = immutableList.builder() |
| 66 | +builder.removeAll { it.value > treshold } |
| 67 | +if (builder.isEmpty()) { |
| 68 | + builder.add(defaultValue) |
| 69 | +} |
| 70 | +return builder.build() |
| 71 | +``` |
| 72 | +
|
| 73 | +In case if each operation invoked on a builder causes no modifications, |
| 74 | +the collection returned by `build()` method is the same collection the builder was obtained from. |
| 75 | +
|
| 76 | +Builders have two advantages: |
| 77 | +- Allow to utilize existing API taking mutable collections. |
| 78 | +- Builders could keep their internal data structures mutable and finalize them only when `build()` method is called. |
| 79 | +
|
| 80 | +### Variance |
| 81 | +
|
| 82 | +It's convenient for immutable collections to have the same variance as the read-only collections they extend. |
| 83 | +However their modification operations could have covariant type appear in parameter (contravariant) position, |
| 84 | +therefore that parameter type must be annotated with `@UnsafeVariance` annotation, and special care must be given when |
| 85 | +implementing such methods. |
| 86 | +
|
| 87 | +### Interfaces and implementations |
| 88 | +
|
| 89 | +#### ImmutableCollection |
| 90 | +
|
| 91 | +- Extends read-only Collection. |
| 92 | +- Provides builder as a mutable collection. |
| 93 | +
|
| 94 | +#### ImmutableList |
| 95 | +
|
| 96 | +- Extends read-only List and immutable Collection. |
| 97 | +- Provides builder as a mutable list. |
| 98 | +- Has the single implementation, returned by `immutableListOf` function. |
| 99 | +
|
| 100 | +#### ImmutableSet |
| 101 | +
|
| 102 | +- Extends read-only Set and immutable Collection. |
| 103 | +- Provides builder as a mutable set. |
| 104 | +- Has two implementations (`immutableSetOf`, `immutableHashSetOf`) |
| 105 | +differing by whether the insertion order of elements is preserved. |
| 106 | +
|
| 107 | +#### ImmutableMap |
| 108 | +
|
| 109 | +- Extends read-only Map. |
| 110 | +- Provides builder as a mutable map. |
| 111 | +- Has two implementations (`immutableMapOf`, `immutableHashMapOf`) |
| 112 | +differing by whether the insertion order of keys is preserved. |
| 113 | +
|
| 114 | +### Extension functions |
| 115 | +
|
| 116 | +#### toImmutableList/Set/Map |
| 117 | +Converts read-only or mutable collection to immutable one. |
| 118 | +If the receiver is already immutable and has the required type, returns itself. |
| 119 | +
|
| 120 | + fun Iterable<T>.toImmutableList(): ImmutableList<T> |
| 121 | + fun Iterable<T>.toImmutableSet(): ImmutableSet<T> |
| 122 | +
|
| 123 | +#### `+` and `-` operators |
| 124 | +
|
| 125 | +`plus` and `minus` operators on immutable collections can exploit their immutability |
| 126 | +and delegate the implementation to the collections themselves. |
| 127 | +
|
| 128 | +#### mutate |
| 129 | +
|
| 130 | +A quite common pattern with builders arises: get a builder, apply some mutating operations on it, |
| 131 | +transform it back to an immutable collection: |
| 132 | +
|
| 133 | + collection.builder().apply { some_actions_on(this) }.build() |
| 134 | +
|
| 135 | +This pattern could be extracted to an extension function. |
| 136 | +
|
| 137 | +``` |
| 138 | +ImmutableList.mutate(action: (MutableList) -> Unit) = |
| 139 | + builder().apply { action(this) }.build() |
| 140 | +``` |
| 141 | +
|
| 142 | +## Use cases |
| 143 | +
|
| 144 | +* Avoiding defensive copying of collection passed as a parameter |
| 145 | + when you need an unchangeable snapshot and you're unsure whether the collection could be changed later. |
| 146 | +
|
| 147 | + ``` |
| 148 | + class Query(val parameters: ImmutableList<Parameter>) |
| 149 | +
|
| 150 | + // or |
| 151 | +
|
| 152 | + class Query(parameters: List<Parameter>) { |
| 153 | + // creates an immutable list or does nothing if it's already immutable |
| 154 | + val parameters = parameters.toImmutable() |
| 155 | + } |
| 156 | + ``` |
| 157 | +
|
| 158 | +* Sharing a single immutable collection between multiple threads. Since the collection |
| 159 | +cannot be changed neither by producer, nor by any consumer, you'll never get a `ConcurrentModificationException` |
| 160 | +during iteration, or any other symptoms of corrupted state. |
| 161 | +
|
| 162 | +* Having multiple snapshots of a collection sharing their element storage. |
| 163 | +
|
| 164 | +## Alternatives |
| 165 | +
|
| 166 | +* Create a mutable collection, but expose it only as its read-only interface. |
| 167 | + This pattern is mostly employed in the standard library collection operations. |
| 168 | +
|
| 169 | + * Pro: mutable collection implementations are generally more memory efficient, |
| 170 | + require less allocations during and after their populating. |
| 171 | + * Con: collection can be cast back to mutable (e.g. involving platform types), |
| 172 | + and then unexpectedly mutated. |
| 173 | + * Con: different snapshots of a collection do not share any element storage. |
| 174 | + * Con: defensive copying of the resulting collection is still required later |
| 175 | + in a chain of operations. |
| 176 | +
|
| 177 | +
|
| 178 | +## Dependencies |
| 179 | +
|
| 180 | +What are the dependencies of the proposed API: |
| 181 | +
|
| 182 | +* interfaces depend only on a subset of Kotlin Standard Library available on all supported platforms. |
| 183 | +* current collection implementations are based on [PCollections](http://pcollections.org) library, |
| 184 | + which is being shaded into the resulting artifact. This results in dependency on JVM 1.6 and upper. |
| 185 | +
|
| 186 | +## Placement |
| 187 | +
|
| 188 | +* Package: `kotlin.collections.immutable` |
| 189 | +* At first the API is to be provided in a separate module `kotlin-collections-immutable`, |
| 190 | + but later it may be merged into `kotlin-stdlib`. |
| 191 | +
|
| 192 | +## Reference implementation |
| 193 | +
|
| 194 | +Reference implementation is given in this repository, see the |
| 195 | +[source](https://github.com/ilya-g/kotlinx.collections.immutable/tree/master/kotlinx-collections-immutable/src/main/kotlin/kotlinx/collections/immutable). |
| 196 | +
|
| 197 | +## Known issues |
| 198 | +
|
| 199 | +1. There is no implementation of immutable ordered map. |
| 200 | +2. Immutable collections currently aren't `Serializable`. |
| 201 | +
|
| 202 | +## Unresolved questions |
| 203 | +
|
| 204 | +1. Should we expose immutable collection implementations to the public API? Current experience shows |
| 205 | +that it's not mandatory. |
| 206 | +2. Immutable map provides `keys`, `entries` sets and `values` collection. |
| 207 | +Should these collections be also immutable, or just read-only is enough? |
| 208 | +Note that their implementations are not persistent and most of their |
| 209 | +modification operations require to make a copy of the collection. |
| 210 | +3. `Map - key` operation: we do not support such operation for read-only maps, and for mutable maps we |
| 211 | +can do `MutableMap.keys -= key`. What would be the analogous operation for `ImmutableMap`? |
| 212 | +4. `mutate` extension: should the action take `MutableList` as a receiver |
| 213 | +or as a parameter (`this` or `it` inside lambda)? |
| 214 | +5. `immutableMapOf<K,V>().builder()` requires explicit specification of K and V type arguments |
| 215 | +(until the common type inference system is implemented) and is quite lengthy. |
| 216 | +Should we provide some shortcut to infer types from expected map type? |
| 217 | +
|
| 218 | +
|
| 219 | +## Future advancements |
| 220 | +
|
| 221 | +* Provide other immutable collection interfaces such as `ImmutableStack` and `ImmutableQueue` |
| 222 | +and their implementations. |
| 223 | +* Improve standard library collection operations to exploit benefits of immutability of the source collections. |
| 224 | +
|
0 commit comments