# Local webmail sending: paired NAS retest

Each pass began with the same 50 isolated inboxes and one password-verification slot. Five clients sent 2,500 local messages over fresh HTTPS connections (50 distinct senders per inbox), then logged out, performed five bursts of five simultaneous logins and made 100 authenticated reads. Every pass was audited against the stored MIME sender and recipient. Tests ran in baseline, optimized, optimized, baseline order. No message was sent to the internet in this local throughput test.

| Measured per pass | Baseline 1 | Optimized 1 | Optimized 2 | Baseline 2 |
|---|---:|---:|---:|---:|
| Time for 2,500 local sends (s) | 50.21 | 47.27 | 48.20 | 48.48 |
| Individual send, p95 (ms) | 153.80 | 123.50 | 123.80 | 124.94 |
| Send-phase process CPU (seconds) | 27.47 | 26.55 | 26.68 | 27.49 |
| Send-phase peak sampled RSS (MiB) | 10.23 | 7.06 | 7.02 | 7.10 |
| Five-login peak sampled RSS (MiB) | 22.43 | 19.14 | 18.96 | 19.21 |
| Five-login mean process CPU (% of one core) | 98.11 | 96.57 | 96.13 | 96.35 |
| Login + first inbox, p95 (ms) | 1097.26 | 1105.03 | 1100.03 | 1101.89 |

In the second pair, the client measured the mean time to establish TCP/TLS, including address resolution, and the remaining time through the durable HTTP response. Baseline: 27.39 69.04 ms, respectively. Optimized: 26.3 69.5 ms. Client timings include contention; they are not measurements of disk latency alone.

The baseline wrote 2,500 Worker hand-off records even though none of these messages came from a Worker. Local webmail generates a new ID per submission and never replays a Worker key. Skipping this table for local composition and reusing prepared SQLite statements reduced the database from 516,096 to 356,352 bytes for the same 2,500 messages. Inbound retry protection, file and directory syncs, and the full SQLite commit remain in place. These two changes were tested together.

The earlier 7% slowdown came from comparing one run per release. The exact historical cause cannot be isolated from those samples. Compare matched runs and CPU seconds as well as wall time; other NAS activity and fresh TLS connections contribute to variation. RAM is process RSS sampled at 250 ms during sending and 50 ms during login.