Optimizing Interrupt Handling Performance for Memory Failures
research.fb.com
research.fb.com
Their setup with all the different services and auto-remediation sounds extremely complicated and is probably massive overkill for anyone not at FB scale. On the other hand that makes a compelling case for running your infra on the cloud, as you would have to deal with these issues yourself on-prem.
Many people would bin the hardware with just one memory checksum failure. If they start happening a million times a day, I would bin the hardware, and I'm pretty stingy when it comes to replacing usable stuff...
But I agree with the overall sentiment. If I'm getting more than one bit flip a week, it's time to look at hardware remedies.
it might even allow for alternative, less stable architectures/hardware, trading off for other things (like size or speed)