[shreyanshjain7174]

Step 2a of #52.

StdIo::write_at did seek + write_all. That's 2 syscalls per request, about 160 per group of 80. On unix it now calls FileExt::write_all_at (pwrite). read_at uses pread. Non-unix keeps the seek path.

No semantic change. Placement and partial-failure handling are untouched. Nothing else in the crate reads or moves the file cursor.

Coalescing a group into 1 write (step 2b) is not in this PR. It changes partial-failure acks and needs its own tests. It's deferred, because halving syscalls here gains only about 1%.

Measured

Azure L8s v3, main 8a66c83 against this PR f98b082. Systems ran interleaved in shuffled order in one VM session, with Redis as a control, 5 reps. Ratios are medians of per-rep throughput ratios to main.

Disk	Writers	Turns/s main → PR	Ratio	p99 µs main → PR
Local NVMe	16	48,587 → 48,886	1.006	508 → 510
	64	64,135 → 64,791	1.011 (1.004-1.013)	1,782 → 1,774
	128	66,550 → 67,178	1.010 (1.008-1.014)	3,563 → 3,606
Premium SSD v2	16	9,588 → 9,629	1.009 (0.987-1.034)	2,590 → 2,522
	64	31,480 → 32,505	1.020 (0.892-1.547)	3,610 → 3,334
	128	44,515 → 44,964	1.015 (1.002-1.019)	5,124 → 4,956

perf stat on the writer thread at 128 writers and 640k appends: 641,771 write + 641,771 lseek become 641,771 pwrite64.

A small but free gain, with no regression on either disk. The Premium SSD v2 64-writer runs were noisy for every system. Full data is in #56 and on #52.

Validation

• CI passes on Linux and macOS: fmt, clippy -D warnings, debug and release tests, both sims, fault-inject and Python client tests.
• This branch includes current main (perf(tape): move reply wakeups off the writer thread #53), so CI covers the combination.

[shreyanshjain7174]

A/B (details on #52): +0.6% / +1.1% / +1.0% at 16/64/128 writers. perf confirms 642k write+lseek become 642k pwrite64. Small but free. Marking ready.

[Copilot]

🔵 Needs a closer look

The draft’s required Azure A/B measurements are still pending, so throughput gains and latency regressions need validation before approval.

No open issues 🎉

What changed in this PR

This PR reduces syscall overhead in StdIo as part of #52’s writer-throughput work, without changing write placement or partial-failure handling.

Changes:

• Uses Unix positional I/O for complete writes and single reads, avoiding a separate seek.
• Preserves locking and the seek-based fallback on non-Unix platforms.
• Leaves write coalescing out of scope; performance measurements remain pending.

File	Description
crates/​zerodb-tape/​src/​io.rs	Adds Unix FileExt read/write paths and conditionally imports fallback I/O traits.

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

The localized change preserves explicit-offset access, full-write behavior, and locking, with no unresolved review findings.

No open issues 🎉

_{Review effort: Balanced}