Partial reads and write coalescing

For a positive-length read on a stream socket, a positive return value reports the bytes obtained, possibly fewer than requested. A zero result means orderly end of the incoming byte stream, after already buffered bytes have been consumed; a zero-length request is a separate case and cannot establish EOF. With nonblocking I/O, EAGAIN or EWOULDBLOCK means no progress is currently possible. Preserve state and resume after readiness instead of treating this temporary condition as a disconnected peer.

Define a framing rule independently: perhaps a fixed-width length header followed by that many payload bytes. Keep an accumulator, parse only complete headers, validate the advertised size before allocating, and dispatch only complete frames. Retain surplus bytes for the next record. EOF between frames can be normal; EOF halfway through a declared frame is truncation. Test every possible split of a small sample frame, including a split inside its header, rather than testing only convenient network packet boundaries.

A successful low-level write can also transfer fewer bytes than requested. Retain the unsent suffix and advance its cursor only by the returned count. Restarting from the beginning duplicates the accepted prefix; discarding the whole buffer loses the suffix. A library may instead copy the complete input into its own queue and expose flow control through another API. Read that contract before applying syscall-level retry logic to it. Local acceptance still does not prove that the remote application processed the record.

Serialize writers or maintain an explicitly ordered output queue so concurrent tasks cannot interleave fragments of application frames. Keep each buffer alive until the relevant API releases it. A readiness notification is permission to try, not a guarantee that the whole suffix will fit. Handle permanent failures through the connection state machine and resolve pending operations consistently. If application retries can repeat a completed command after a disconnect, use a protocol-level identifier and documented deduplication policy rather than assuming byte offsets solve distributed uncertainty.

Combining small queued writes can reduce per-call overhead. Scatter/gather I/O can present several buffers without first concatenating them, but a partial completion still needs a cursor spanning buffer boundaries. Choose both a maximum batch size and a maximum waiting time: a size-only policy can strand a small final message indefinitely. Flush promptly when a request depends on its response, when the timer expires, or when closing according to the protocol. Batch memory must remain bounded under a slow reader.

Use the queue lab to examine the pressure behind that bound. Press Reset, set arrival rate to 4 and service rate to 3, predict growth from four initial items, then press Play. For a separate draining comparison, press Reset again, explicitly set arrival rate to 2 and service rate to 3, predict the slope, then press Play. Slider edits pause but retain elapsed time, so they are not a live change applied to an earlier backlog. The model uses equal abstract work units; it does not simulate TCP segments, partial writes, batching timers, or kernel buffers. Measure syscall count, queued bytes, and end-to-end latency in your implementation before claiming that a larger batch is better.