I agree that we need a better way to have bytes.Buffer interact with []byte, but I believe it is wrong to target Put-like operations where you need to know the size of the buffer up-front †. Doing so would greatly limit the ways this API can be used. Put-like operation are much rarer compared to append-like APIs (e.g.,
strconv.AppendXXX, fmt.AppendXXX, binary.AppendXXX, utf8.AppendRune, hash.Sum, etc.)

In #47527, we added bufio.Writer.AvailableBuffer so that it could be used together with append-like APIs.
I believe it would be better to provide a method like:

// AvailableBuffer returns an empty buffer with non-zero capacity.
// This buffer is intended to be appended to and
// passed to an immediately succeeding Write call.
func (b *Buffer) AvailableBuffer() []byte {
    if cap(b.buf) - len(b.buf) {
        b.grow(1)
    }
    return b.buf[len(b.buf):]
}

Thus, usage would look like:

var bb *bytes.Buffer = ...
b := bb.AvailableBuffer()
b = binary.AppendUvarint(b, ...)
bb.Write(b)

Note that this operation is allocation-free and copy-free in the case where the Buffer has enough capacity. The Write operation uses copy under the hood, which is a noop if the source and destination are the same.

Aside: If we want to be consistent with bufio.Writer, we could also add:

func (b *Buffer) Available() int {
    return cap(b.buf) - len(b.buf)
}

† Also, most Put-like operations have had an append-like equivalent added in recent years. Examples:

• utf8.AppendRune instead of utf8.EncodeRune
• binary.AppendUvarint instead of binary.PutUvarint
• binary.LittleEndian.AppendUint64 instead of binary.LittleEndian.PutUint64

If other Put-like operations lack an append-like equivalent, we should propose that one be added.

I've often wished for a way to get at the underlying buffer so I support this proposal. It seems like this has fallen through the cracks maybe, since there has been no update since July 6th?

@dsnet could you explain why you'd have AvailableBuffer guarantee non-zero capacity? Seems like that is different from bufio.Writer.AvailableBuffer.

I can't remember why it guarantees non-zero capacity. It isn't strictly necessary.

This proposal is in the list of incoming proposals, of which there are many. It hasn't fallen through the cracks.

I remember now. If the usage pattern is:

var bb *bytes.Buffer = ...
b := bb.AvailableBuffer()
b = binary.AppendUvarint(b, ...)
bb.Write(b)

we would want the allocation to occur within AvailableBuffer so that the internal buffer of Buffer is grown first. If the allocation occurs in binary.AppendUvarint, then it will allocate again when we call bb.Write.

It's a small optimization to reduce the amount of allocations when an empty Buffer is first used.

Ian: I didn't mean this as a criticism, sorry if it came across that way. It's not clear from the outside whether a proposal has been evaluated to go into active or not.

To add something to the discussion, I've been playing around with AvailableBuffer and friends and came up with the following two functions:

func readBuffer(r io.Reader, n int) ([]byte, error) {
	switch r := r.(type) {
	case *bufio.Reader:
		buf, err := r.Peek(n)
		if errors.Is(err, bufio.ErrBufferFull) {
			// We've requested a buffer than is larger than the reader. Fall
			// back to allocating a new slice.
			break
		}
		if err != nil {
			return nil, err
		}

		// Advance the internal state of r by reading into the already peeked
		// buffer. There is no additional copy, since dst == src.
		_, err = r.Read(buf)
		return buf, err

	case *bytes.Buffer:
		buf := r.Next(n)
		if len(buf) == 0 {
			return nil, io.EOF
		} else if len(buf) < n {
			return nil, io.ErrUnexpectedEOF
		}

		return buf, nil
	}

	buf := make([]byte, n)
	_, err := io.ReadFull(r, buf)
	return buf, err
}

func writeBuffer(w io.Writer, n int) []byte {
	switch w := w.(type) {
	case *bufio.Writer:
		if w.Available() < n {
			break
		}

		return w.AvailableBuffer()

	// XXX: Need bytes.Buffer.AvailableBuffer here.
	}

	return make([]byte, 0, n)
}

With these two functions it's possible to wrap append-style encoding APIs into functions that take io.Reader / io.Writer which don't allocate in the common case. One example where I would like to see this is binary.Read and binary.Write.

Some more observations:

• Having both bytes.Buffer.AvilableBuffer and bufio.Writer.AvailableBuffer with the same semantics would allow switching on interface{AvailableBuffer() []byte; io.Writer } in writeBuffer and would allow users to benefit from the same optimization with custom readers.
• readBuffer is more complicated since Peek() and Next() have distinct semantics. Using Peek + Read this way is counterintuitive, since one has to know that the copy in Read is elided for this sequence to make any sense. OTOH Next() requires synthesising io.EOF / io.UnexpectedEOF in the caller.

P.S.: Should we split AvailableBuffer into a new, distinct proposal?

@lmb:

• In writeBuffer, it is probably more efficient to call w.Grow if the available buffer is too small.
• In readBuffer, it is generally more common to use r.Peek with r.Discard.

@lmb I didn't take it as criticism, I was just explaining the current status.

• In writeBuffer, it is probably more efficient to call w.Grow if the available buffer is too small.

Something like this (if and when bytes.Buffer supports AvailableBuffer())?

	case *bytes.Buffer:
		w.Grow(n)
		return w.AvailableBuffer()

Change https://go.dev/cl/440135 mentions this issue: encoding/binary: avoid allocations for common Readers / Writers

Given that we added https://pkg.go.dev/bufio#Writer.AvailableBuffer in Go 1.18, it sounds like we should use the same pattern in bytes.Buffer to allow the same idioms?

This proposal has been added to the active column of the proposals project
and will now be reviewed at the weekly proposal review meetings.
— rsc for the proposal review group

It sounds like people are happy with AvailableBuffer() []byte like in bufio.Writer.
Do we also need to add Available() int for alignment between bufio.Writer, or is that not important?
If there's not a strong argument for including Available() int, we can leave it out to start.

I have a use for Available to compute whether I should manually grow the buffer to reduce the probability of allocations later on with append.

Something like:

// Always ensure at least 25% capacity remains before the next batch of append operations.
if b.Available() < b.Len()/4 {
    b.Grow(b.Available()+1)
}

That said, this is no different than:

if cap(b.AvailableBuffer()) < b.Len()/4 {
    b.Grow(cap(b.AvailableBuffer())+1)
}

but isn't as readable.

Change https://go.dev/cl/469558 mentions this issue: encoding/json: use append-like operations for encoding

Have all concerns about this proposal been addressed?

This is more of an aside and not a specific concern: I find it surprising that bytes.Buffer has been fine without this method for so long (and it seems to me like a frequently used struct) and so I’m curious if people have an opinion about why this is necessary now/what changed.

This feature was technically always possible where:

b := bb.AvailableBuffer()

is semantically identical to:

b := bb.Bytes()[bb.Len():]

but doing so depended upon undocumented guarantees of Buffer.Bytes.

Years ago, I tried modifying the documentation on Bytes to guarantee that this was safe. I never followed through on that CL, but have relied on this property for performance in several applications over the years. It's perhaps the only performant way to use append-like APIs with the Buffer type.

I’m curious if people have an opinion about why this is necessary now/what changed.

Pure anecdata / speculation, I think that people who found performance with Buffer to be lacking switched to []byte. That has the downside of bifurcating APIs into "takes a byte slice" and "takes an io.Writer". With AvailableBuffer on the two common stdlib buffers it we can write wrappers that bridge Reader / Writer to []byte efficiently for the common case.

Based on the discussion above, this proposal seems like a likely accept.
— rsc for the proposal review group

No change in consensus, so accepted. 🎉
This issue now tracks the work of implementing the proposal.
— rsc for the proposal review group

Change https://go.dev/cl/474635 mentions this issue: bytes: add Buffer.Available and Buffer.AvailableBuffer

Change https://go.dev/cl/499617 mentions this issue: doc/go1.21: document new bytes.Buffer methods