In June 2026, the ERP server at a hardware and sheet-metal factory in Foshan raised a disk warning. Following the usual routine, the administrator who also handles IT reached for the USB backup drive. It would not mount. The last restore anyone had tested was in 2023.
"Stop for a day and more than 60 people on the floor wait for material," the plant manager told us later. "Customer delivery dates slip with it. We cannot carry that risk any longer."
On site: three numbers
| Measure | What we found | Note |
|---|---|---|
| Size | 120 staff, 8 CAD seats | Hardware tooling and sheet-metal parts |
| ERP server | 2013 tower | 4-core CPU, 16 GB RAM, 2 x 2 TB SATA RAID1, single PSU |
| Month-end close | About 4 hours | SQL Server and file sharing on the same box |
CAD was a separate bill. A 500 MB mold assembly took 90 s to open, and one rotation stalled for two seconds. Designers had a habit of opening a drawing and going to make tea.
Diagnosis: a server on borrowed time, CAD short on VRAM
The server. Random I/O on those mechanical drives measured in the low hundreds of IOPS. SQL Server's tempdb and the ERP report queries fought over the same spindles. One power supply meant a single voltage dip could stop the line. The backup drive sat in the same room, had never been verified, and had no copy anywhere else.
CAD. The eight seats ran i5-8400 / 16 GB / Quadro P1000 4 GB. A P1000 carries 4 GB of VRAM, and a 500 MB assembly with textures and real-time rendering runs past that ceiling. 16 GB of system memory was also too little to keep SolidWorks and the ERP client open at the same time.
Design: keep the business running first, then go faster
| Part | Approach | Budget | Why |
|---|---|---|---|
| ERP server | 2U dual-socket, 128 GB, SSD RAID1 plus 16 TB RAID10, dual PSU | RMB 46,000 | Virtualization splits ERP and file server into two VMs |
| Backup | NAS and rotating offline drives | RMB 13,000 | Daily incremental, weekly full, quarterly drill; RPO 24 h, RTO 4 h |
| CAD workstations | i7-14700 / 32 GB / 1 TB NVMe / RTX A2000 12 GB | 8 x RMB 10,500 | All eight replaced; old machines kept as spares |
| Implementation | Weekend migration, data validation, training | RMB 9,000 | No production downtime |
The backup design is deliberately boring. A 4-bay NAS with four 8 TB drives in RAID5 (about 24 TB usable) sits on its own network segment, and USB drives are rotated off site every week. Daily jobs are incremental, weekly jobs are full, and once a quarter someone restores a real database and signs the sheet. A 24-port 1 GbE switch with VLANs keeps office traffic away from production, so a large file copy in the front office does not compete with a goods receipt posting in the workshop.
Configuration
| Device | Specification | Qty |
|---|---|---|
| ERP server | 2 x Xeon Silver 4310 (12C/24T), 128 GB DDR4 ECC, 2 x 480 GB SSD RAID1, 4 x 8 TB SATA RAID10, 800 W 1+1, 2U | 1 |
| Backup NAS | 4 bays, 4 x 8 TB, RAID5 (about 24 TB usable), separate network segment | 1 |
| CAD workstation | i7-14700 (20C/28T), 32 GB DDR5, 1 TB NVMe, RTX A2000 12 GB, 500 W | 8 |
| Switch | 24-port 1 GbE, VLANs separating office and production traffic | 1 |
Rollout: Friday evening to Monday morning
The cutover had to fit inside a weekend, because the shop floor starts again on Monday morning. That constraint shaped the plan more than any hardware choice.
- Friday 18:00. ERP stopped. Full backup of database and attachments, then a restore check before anything else moved.
- Friday 20:00 to Saturday 02:00. New server racked, virtualization installed, ERP and file data migrated.
- Saturday. Eight CAD workstations swapped in, SolidWorks and print drivers installed, large-assembly tests run.
- Sunday. Full ERP regression: order entry, material issue, goods receipt, reconciliation. Then a backup restore drill, not a screenshot of a success message.
- Monday 07:30. The shop floor logged in. The old server stayed on site for 30 days as a fallback.
The riskiest hour was Friday night, when the old array was still the only copy of the data. That is why the full backup was verified before the migration started, and why the old server was left intact instead of being wiped and sold.
Results
Month-end close dropped from about 4 hours to 45 minutes. The same 500 MB assembly opens in 18 s, and rotating it no longer stalls. Designers no longer open a drawing and go make tea.
The bigger change is that maintenance became something the factory can schedule. Dual power supplies, RAID10, a daily backup and a quarterly drill gave the administrator a server that can be managed rather than one that has to be gambled on. Replacing a failed drive is now a task for a Tuesday afternoon instead of an emergency on the last day of the month.
The bill: where RMB 152,000 went
| Item | Amount |
|---|---|
| ERP server | RMB 46,000 |
| Backup NAS and offline drives | RMB 13,000 |
| CAD workstations, 8 x RMB 10,500 | RMB 84,000 |
| Implementation and training | RMB 9,000 |
| Total | RMB 152,000 |
The eight workstations, not the server, took the largest share of the budget. The backup line was the smallest hardware line on the invoice and the one that would have kept the June scare from becoming a shutdown.
Five lessons
- RAID is not backup. RAID survives a dead drive. Backup survives a deletion, ransomware or a fire. They are different jobs and belong on different systems.
- A backup you have never restored is only a hope. Put the drill on the calendar and keep the record, because the person who runs it may not be the person who set it up.
- Dual power supplies are the minimum, not an upgrade. Factory power moves more than office power, and one outage teaches the lesson for good.
- For CAD, check VRAM before CPU. A large assembly that stutters is often short of memory and VRAM, and sometimes limited by single-thread speed and disk.
- Keep the old machines. Two retired CAD boxes stayed as spares, so a failed new workstation does not idle a designer.
Note: at the client's request, the organization's name, some department names and some timestamps have been generalized. Configurations and prices reflect the market as of October 2026, and an actual project depends on a site survey.