Announcing chunked-bytes: a non-contiguous byte output buffer

In network programming, there is often a need to serialize data structures into a wire protocol representation and send the resulting sequence of bytes to an output object, such as a socket. For a variety of reasons, developers of protocol implementations prefer to serialize the data into an intermediate buffer in memory rather than deal with output objects directly in either synchronous or asynchronous form, or both. When a contiguous buffer like Vec is used for this, reallocations to fit a larger length of serialized data may adversely impact performance, while evaluating the required length to pre-allocate beforehand may be cumbersome or difficult in the context. The single contiguous buffer also forms a boundary for write requests, creating a need for copy-back schemes to avoid inefficiently fragmented writes of tail data.

Another important use case is passing network data through. If some of the data is received into Bytes handles, it should be possible to inject the data into the output stream without extra copying.

Enter ChunkedBytes, containers that can be used to coalesce data added as byte slices via the BufMut interface, as well as possible Bytes input, into a sequence of chunks suitable for implementations of the Write::write_vectored method. This design aims to deliver good performance regardless of the size of buffered data and with no need to pre-allocate sufficient capacity for it.

Release 0.1.0 of chunked-bytes is now available at crates.io.

4 Likes

A rework of tokio_util::codec::Encoder could benefit, in the general case, from using ChunkedBytes as the buffer instead of BytesMut.