I moved my first MES to the cloud in 2016. Failed. Not the tech — the approach. We lifted a SQL Server instance into an EC2 box and called it cloud. Three months later the plant manager called me at 2 AM because the VPN tunnel dropped during a shift change and batch records stopped writing. That is when I learned that manufacturing IT is not regular IT. The patterns that work for your average enterprise app do not translate cleanly to a plant floor where downtime is measured in thousands of dollars per minute.
The pattern I see work most often is not lift-and-shift. It is selective re-platforming. Take the reporting layer, the data historian, the dashboards that managers actually use — pull those into the cloud first. Leave the real-time control systems where they are. I had a client in Ohio running a Wonderware historian on-prem that nobody wanted to touch. We replicated that data to a time-series service in Azure, built Power BI dashboards on top, and the plant team saw value within six weeks. No control loops were touched. The OT engineers did not care because we did not make them care.
Another pattern that comes up is edge-plus-cloud. You put a small compute node on the floor — something like an Azure Stack Edge or a ruggedized Dell box — and let it handle local aggregation and buffering. The cloud side does the heavy analytics. This saved one of our clients in Michigan when their internet circuit went down for nine hours during a fiber cut. The edge node kept collecting, buffered locally, and synced up when the connection came back. Without that pattern, they would have lost nine hours of batch genealogy. In food and beverage, that is a recall waiting to happen.
Do not try to migrate everything at once. I tried that at a plant in Georgia — tried to move ERP, MES, WMS, and the label printing system in one six-month project. The label printers broke. Nobody tested the Zebra printers against the new print server driver. Production stopped for four hours on a Tuesday. Now I break migrations into phases of eight to ten weeks max, and every phase has a rollback plan that someone actually tested, not just wrote down.
One thing nobody talks about enough is DNS. On a factory floor, half the devices talk to hardcoded IP addresses. HMI panels, PLCs, scale indicators — they do not use hostnames. When you move the server to the cloud, those connections break unless you re-IP everything or put in a NAT layer that maps the old addresses. I now budget two weeks minimum for DNS and IP mapping on every manufacturing migration. It is never as simple as the network diagram suggests.
The pattern that matters most is people. Get the plant controls engineer involved on day one. Not in a steering committee. In the architecture session. They know which systems cannot tolerate latency and which ones the operators will manually work around if the screen takes an extra two seconds. That knowledge is worth more than any cloud reference architecture diagram you pull from Microsoft or AWS.
The cloud is good for manufacturing. It really is. But the migration has to respect the floor, not just the data center. Otherwise you end up like I did in 2016 — on the phone at 2 AM, explaining to a very angry plant manager why the batch records stopped writing.