Now, computers are built with subsystems from a half dozen different companies, none of which can be bothered to fully document their product's behavior. Product life cycles are so fast it isn't worth working out what a device actually does, because by the time you do you either won't be able to buy it anymore, or someone's new iteration will be 20% cheaper/faster/bigger and you will get killed in the market if you don't switch to it.
␄
PS: We were also happy to have a whole MIP of Vax equivalent (or identical) processing power, and I was slightly reknowned for my ability to code cleverly enough to get 42 disk IO/second out of our main disk drive, so I'd rather not go back, think you very much. I'll get along as best I can with the miraculous and inexpensive rubbish we build our systems from today.
So, if the drives didn't lie about a flush that problem is solved. But if it's on disk at time t, that's no guarantee that you can read it back at time t+1. The drives can physically fail. These "days of old" are before my time, but I really doubt they had magic disks that never physically degraded.
As someone mentioned, this is contrast to say how MongoDB was shipping not too long ago. They had turned of any acknowledgement for writes. So doing a db write was more like a throw over the fence and pray operation. So it didn't even try to be durable, by design.
This is in stark contrast to something like MongoDB, which barely parses the request before reporting a success, and doesn't make any guarantees of when or even if it will ever save the data (though it usually does).
My understanding is that the kernel (for Windows and Linux values of "kernel") will never lie to you. They will accurately report what the storage drivers told it, and the lying occurs at that level.
Storage drivers seem to be a big source of corruption for MSSQL [1]:
> The most common cause of database corruption (more than 95% of all corruption cases) that we in PSS encounter turn out to be caused by a platform issue, which is a layer below the SQL Server. The most common individual cause is a 3rd party driver or firmware bug.
And as well
> My guess would be that hard-drives have explicit instruction(s)
They do, but the problem is that the storage drivers lie about what actually happened. E.g. Basically they implement write caching for the "flush the write cache" instruction.
[1] http://blogs.msdn.com/b/suhde/archive/2009/04/08/introductio...
But of course doing that is up to the user. Using write-back caching with long sync interval and disabling fsync. It's really nice, gives better than SSD performance with regular HDD. Until you shut it down uncleanly, then you're screwed. But of course any sane person would use this only for temporary or other really non-important data which can be regenerated or lost without problems in such situation.
I'm using such configuration with ERP,BI/ETL (Extract, Transform, Load) tasks. When I start the task, I anyway drop and recreate any tables required for the task. SO I don't really mind if data gets corrupted. That's just life. Doing safe commits would make task very slow.
Only good question is how to balance smartly, in application data / caching, database engine caching and file system caching. In cases where database runs on same server as the processing application.