proposal: iter: new package for iterators · Issue #61897 · golang/go · GitHub

We propose to add a new package iter that defines helpful types for iterating over sequences. We expect these types will be used by other APIs to signal that they return iterable functions.

This is one of a collection of proposals updating the standard library for the new 'range over function' feature (#61405). It would only be accepted if that proposal is accepted.

See also:

Note regarding push vs pull iterator types: The vast majority of the time, push iterators are more convenient to implement and to use, because setup and teardown can be done around the yield calls rather than having to implement those as separate operations and then expose them to the caller. Direct use (including with a range loop) of the push iterator requires giving up storing any data in control flow, so individual clients may occasionally want a pull iterator instead. Any such code can trivially call Pull and defer stop.

I’m unaware of any significant evidence in favor of a parallel set of pull-based APIs: instead, iterators can be defined in push form and preserved in that form by any general combination functions and then only converted to pull form as needed at call sites, once all adapters and other transformations have been applied. This avoids the need for any APIs that make cleanup (stop functions) explicit, other than Pull. Adapters that need to convert to pull form inside an iterator function can defer stop and hide that conversion from callers. See the implementation of the adapters in #TODO for examples.

Note regarding standard library: There a few important reasons to convert the standard library:

• Ship a complete solution. We should not release a package we don’t use in obvious places where it should be used. This applies especially to new interfaces.
• Put functionality in the right places. Ian’s earlier draft included FromSlice and FromMap, but these are more appropriately slices.All and maps.All.
• Find problems or rough edges in the iter package itself that we want to find before its release. I have already changed a few definitions from what I started with as a result of working through the standard library changes.
• Set an example (hopefully a good one) for others to follow.

Note that os.ReadDir and filepath.Glob do not get iterator treatment, since the sorted results imply they must collect the full slice before returning any elements of the sequence. filepath.SplitList could add an iterator form, but PATH lists are short enough that it doesn’t seem worth adding new API. A few other packages, like bufio, archive/tar, and database/sql might benefit from iterators as well, but they are not used as much, so they seem okay to leave out from the first round of changes.

The iter package would be:

/*
Package iter provides basic definitions and operations related to
iterators over sequences.

Iterators

An iterator is a function that passes successive elements of a
sequence to a callback function, conventionally named yield, stopping
either when the sequence is finished or when yield breaks the sequence
by returning false. This package defines [Seq] and [Seq2]
(pronounced like seek - the first syllable of sequence)
as shorthands for iterators that pass 1 or 2 values per sequence element
to yield:

type (
	Seq[V any]     func(yield func(V) bool) bool
	Seq2[K, V any] func(yield func(K, V) bool) bool
)

Seq2 represents a sequence of paired values, conventionally key-value,
index-value, or value-error pairs.

Yield returns true when the iterator should continue with the next
element in the sequence, false if it should stop. The iterator returns
true if it finished the sequence, false if it stopped early at yield's
request. The iterator function's result is used when composing
iterators, such as in [Concat]:

func Concat[V any](seqs ...Seq[V]) Seq[V] {
	return func(yield func(V) bool) bool {
		for _, seq := range seqs {
			if !seq(yield) {
				return false
			}
		}
		return true
	}
}

Iterator functions are most often called by a range loop, as in:

func PrintAll[V any](seq iter.Seq[V]) {
	for _, v := range seq {
		fmt.Println(v)
	}
}

Naming

Iterator functions and methods are named for the sequence being walked:

// All returns an iterator over elements in s.
func (s *Set[V]) All() iter.Seq[V]

The iterator method on a collection type is conventionally named All,
as in the second example, because it iterates a sequence of all the
values in the collection.

When there are multiple possible iteration orders, the method name may
indicate that order:

// All iterates through the list from head to tail.
func (l *List[V]) All() iter.Seq[V]

// Backward iterates backward through the list from tail to head.
func (l *List[V]) Backward() iter.Seq[V]

If an iterator requires additional configuration, the constructor function
can take additional configuration arguments:

// Bytes iterates through the indexes and bytes in the string s.
func Bytes(s string) iter.Seq2[int, byte]

// Split iterates through the (possibly-empty) substrings of s
// separated by sep.
func Split(s, sep string) iter.Seq[string]

Single-Use Iterators

Most iterators provide the ability to walk an entire sequence:
when called, the iterator does any setup necessary to start the
sequence, then calls yield on successive elements of the sequence,
and then cleans up before returning. Calling the iterator again
walks the sequence again.

Some iterators break that convention, providing the ability to walk a
sequence only once. These “single-use iterators” typically report values
from a data stream that cannot be rewound to start over.
Calling the iterator again after stopping early may continue the
stream, but calling it again after the sequence is finished will yield
no values at all, immediately returning true. Doc comments for
functions or methods that return single-use iterators should document
this fact:

// Lines iterates through lines read from r.
// It returns a single-use iterator.
func (r *Reader) Lines() iter.Seq[string]

Errors

If iteration can fail, it is conventional to iterate value, error pairs:

// Lines iterates through the lines of the named file.
// Each line in the sequence is paired with a nil error.
// If an error is encountered, the final element of the
// sequence is an empty string paired with the error.
func Lines(file string) iter.Seq2[string, error]

Pulling Values

Functions and methods that are iterators or accept or return iterators
should use the standard yield-based function signature, to ensure
compatibility with range loops and with other iterator adapters.
The standard iterators can be thought of as “push iterator”, which
push values to the yield function.

Sometimes a range loop is not the most natural way to consume values
of the sequence. In this case, [Pull] converts a standard push iterator
to a “pull iterator”, which can be called to pull one value at a time
from the sequence. [Pull] starts an iterator and returns a pair
of functions next and stop, which return the next value from the iterator
and stop it, respectively.

For example:

// Pairs returns an iterator over successive pairs of values from seq.
func Pairs[V any](seq iter.Seq[V]) iter.Seq2[V, V] {
	return func(yield func(V, V) bool) bool {
		next, stop := iter.Pull(it)
		defer stop()
		v1, ok1 := next()
		v2, ok2 := next()
		for ok1 || ok2 {
			if !yield(v1, v2) {
				return false
			}
		}
		return true
	}
}

Clients must call stop if they do not read the sequence to completion,
so that the iterator function can be allowed to finish. As shown in
the example, the conventional way to ensure this is to use defer.

Other Packages

Many packages in the standard library provide iterator-based APIs. Here are some notable examples.

TODO FILL THIS IN AS OTHER PACKAGES ARE UPDATED

Mutation

Iterators only provide the values of the sequence, not any direct way
to modify it. If an iterator wishes to provide a mechanism for modifying
a sequence during iteration, the usual method is to define a position type
with the extra operations and then provide an iterator over positions.

For example, a tree implementation might provide:

// Positions iterates through positions in the sequence.
func (t *Tree[V]) Positions() iter.Seq[*Pos]

// A Pos represents a position in the sequence.
// It is only valid during the yield call it is passed to.
type Pos[V any] struct { ... }

// Pos returns the value at the cursor.
func (p *Pos[V]) Value() V

// Delete deletes the value at this point in the iteration.
func (p *Pos[V]) Delete()

// Set changes the value v at the cursor.
func (p *Pos[V]) Set(v V)

And then a client could delete boring values from the tree using:

for p := range t.Positions() {
	if boring(p.Value()) {
		p.Delete()
	}
}

*/
package iter

// Seq is an iterator over sequences of individual values.
// See the [iter] package documentation for details.
type Seq[V any] func(yield func(V) bool) bool

// Seq2 is an iterator over pairs of values, conventionally
// key-value or value-error pairs.
// See the [iter] package documentation for details.
type Seq2[K, V any] func(yield func(K, V) bool) bool

Note: If and when generic type aliases are implemented (#46477), we might also want to add type Yield[V any] = func(V bool) and type Yield2[K, V any] = func(K, V) bool. That way, code writing a function signature to implement Seq or Seq2 can write the argument as yield iter.Yield[V].

// Pull starts the iterator in its own coroutine, returning accessor functions.
// Calling next returns the next value from the sequence. When there is
// such a value v, next returns v, true. When the sequence is over, next
// returns zero, false. Stop ends the iteration, allowing the iterator function
// to return. Callers that do not iterate to the end must call stop to let the
// function return. It is safe to call stop after next has returned false,
// and it is safe to call stop multiple times. Typically callers should defer stop().
func Pull[V any](seq Seq[V]) (next func() (V, bool), stop func())

// Pull2 starts the iterator in its own coroutine, returning accessor functions.
// Calling next returns the next key-value pair from the sequence. When there is
// such a pair k, v, next returns k, v, true. When the sequence is over, next
// returns zero, zero, false. Stop ends the iteration, allowing the iterator function
// to return. Callers that do not iterate to the end must call stop to let the
// function return. It is safe to call stop after next has returned false,
// and it is safe to call stop multiple times. Typically callers should defer stop().
func Pull2[K, V any](seq Seq[K, V]) (next func() (K, V, bool), stop func())