Measure Ithaca engine-to-MSSQL traffic for separate database host sizing

Read-only production measurement on ithaca.project-rk.com to assess moving MSSQL to a separate same-room host over 2.5 Gbps Ethernet. Identify engine/database paths; measure bidirectional traffic and short-window peaks with low-overhead counters or header-only capture, avoiding credentials and payload storage. No service restart, deployment or network configuration changes. Acceptance: report duration, observed average/peak rates and link utilization, client attribution where available, and limitations including latency, startup bursts and unobserved peak load. Verified 2026-11-01 21:11:40–21:16:40 UTC: 300 seconds, 1200 quarter-second samples of engines-db network namespace eth0 counters; ss TCP counters every ~5 seconds for conservative client attribution. DB 172.18.0.9 on kamal bridge, all engines colocated. Interface RX 752225382 bytes / average20.059 Mbps, TX3604489003 / average96.120 Mbps; quarter-second max156.206/482.830 Mbps RX/TX, quarter-second p9564.311/281.037 Mbps. Nonoverlapping sequential four-sample (one-second) output max60.154/346.548 Mbps. IMPORTANT: separate elapsed-time-bucket one-second summary has boundary jitter and is NOT used; use sequential four-sample output above. Socket deltas lower bounds due24 new/10 closed sockets; four PFEs together TCP RX713497247/TX3460543189 bytes; reporting workflow RX11586551/TX115813646, portal RX3218526/TX3538818. ~96.5% of attributed outbound TCP bytes to four PFEs. No interface drops during sample; no payload capture, configuration change, service restart or deployment. All engine containers remain healthy. Bandwidth suitable for dedicated full-duplex2.5Gbps at observed load (~3.84% average TX/19.31% quarter-second max); exclude Ethernet wire overhead, sub-quarter-second microbursts, busy-hour/startup loads and unknown added SQL round-trip latency. Recommend longer busy-hour/restart-window measurement and actual RTT/application performance validation before migration.