Modbus RTU is old, simple and everywhere, which is exactly why it is worth getting right. Most of the problems we see on site are physical layer problems wearing a protocol costume.
Termination and biasing
RS-485 is a transmission line. Terminate both ends of the bus with a resistor matched to the cable impedance, and terminate only the ends, not every node. Failure to bias the idle state is the other common fault: without bias resistors holding the line in a defined state, noise on an idle bus is read as start bits and your first character disappears.
Topology discipline
Daisy chain the bus. Star topologies and long stubs work on the bench and fail in a cabinet full of contactors. Keep stubs short, keep the pair twisted all the way to the terminal, and give the shield a single ground reference so it does not become a loop.
Baud rate is a trade, not a target
Higher baud rates shorten the polling cycle but reduce noise margin over long cable runs. If you are seeing intermittent CRC errors, dropping the rate is a legitimate fix rather than a defeat. A reliable poll at 9600 beats a flaky one at 115200.
Register maps are rarely as documented
Expect off-by-one addressing between documentation and implementation, since some vendors document one-based register numbers and others use zero-based addresses. Expect word order for 32-bit values to differ between devices on the same bus. Verify against a known value on the equipment display before trusting a scaling factor.
Timeouts and retries
Set your inter-frame timeout from the actual baud rate rather than a fixed constant, and implement a bounded retry with backoff. A device that is slow to answer during its own internal calculation should not knock your poller into a loop of retries that starves every other node on the bus.
Isolation
Isolate the transceiver. Ground potential between a control cabinet and a remote pump house is not a theoretical concern, and an isolated interface converts an expensive board failure into a non-event.