A slave recently began to report syntax errors in replicated queries. This happend about once a day and was completly random. As the master ran fine and the connection between the two hosts has no errors, the error must be somewhere located on the slave. I began testing the server including the discs without result. Later I accidently did something that made a service on the slave crash. In the error message I could see that the kernel discovered some kind of memory corruption in the process and terminated it.

After running memtest for a couple of houres on the system some memory errors were found. The server was tested before deployment and was running without errors for about two years so memory was the last thing I thought about. But now I can imagine why enterprise servers use ECC memory.