4 ms·

It's not losing pending writes, it's the drive saying there are no pending writes, but losing them anyways. ie the drive is most likely lying

monocasa 5y ago
As I said in the recent Apple discussion, pretty much all drives are lying and have been for decades at this point. The good brands just spec out enough capacitance that you don't see the difference externally. https://news.ycombinator.com/item?id=30370551#30374585 https://news.ycombinator.com/item?id=30370551#30374585
- brigade 5y ago
  So if SSDs rely solely on capacitors for data integrity and lie about flushes, what do they do on a flush that takes any amount of time? Are they just taking a speed hit for funsies? Heck, from this test, the magnitude of the speed hit isn't even correlated with whether they lose writes...
  - monocasa 5y ago
    At one point it was different barriers on the different submission queues inside the drive. Not externally visible queues, but between internal implementation layers. It's been a few years since I've checked up on this and it was for the most part pre SSDs though.
  - supermatt 5y ago
    Probably implementing a barrier for ordering io…
  - dboreham 5y ago
    When you look at how long it takes to perform a block write on a flash device, you'll see that no SSD is going to honor flush semantics.
- WrtCdEvrydy 5y ago
  The more you look at speculative execution and these drive issues, the more you see that we're giving up a lot of what computing "safe" for just performance.
  - _greim_ 5y ago
    Brings to mind Goodhart's law: > When a measure becomes a target, it ceases to be a good measure. In this case, performance as a measure of value.
  - blibble 5y ago
    it's mongodb all over again gotta get dem sweet sweet benchmarks, to hell if we're persisting garbage
- consp 5y ago
  Considering the amount of 10uF Tantalum caps (30) on one of the bricked enterprise SSDs I opened I'm not surprised at all.
  - dboreham 5y ago
    300uF is not a supercap.
    - consp 5y ago
      Never said they were and you don't need several seconds of power buffer for this purpose. They might just as well serve a different purpose like I know it just an observation. Either explain why why you definitely need F range for this or don't comment. They are rated 16v so take that into account in your explanation.
  - pixl97 5y ago
    You should have cooling in your datacenter.
    - consp 5y ago
      All of your assumptions are wrong... (cause of failure, location, usage)
- fire 5y ago
  Any ides how one finds out whether a drive is actually capped properly to handle power off data loss?
dboreham 5y ago
Understood. Low-end drives have always done this because: performance.

monocasa 5y ago

As I said in the recent Apple discussion, pretty much all drives are lying and have been for decades at this point. The good brands just spec out enough capacitance that you don't see the difference externally. https://news.ycombinator.com/item?id=30370551#30374585 https://news.ycombinator.com/item?id=30370551#30374585

brigade 5y ago
So if SSDs rely solely on capacitors for data integrity and lie about flushes, what do they do on a flush that takes any amount of time? Are they just taking a speed hit for funsies? Heck, from this test, the magnitude of the speed hit isn't even correlated with whether they lose writes...
- monocasa 5y ago
  At one point it was different barriers on the different submission queues inside the drive. Not externally visible queues, but between internal implementation layers. It's been a few years since I've checked up on this and it was for the most part pre SSDs though.
- supermatt 5y ago
  Probably implementing a barrier for ordering io…
- dboreham 5y ago
  When you look at how long it takes to perform a block write on a flash device, you'll see that no SSD is going to honor flush semantics.
WrtCdEvrydy 5y ago
The more you look at speculative execution and these drive issues, the more you see that we're giving up a lot of what computing "safe" for just performance.
- _greim_ 5y ago
  Brings to mind Goodhart's law: > When a measure becomes a target, it ceases to be a good measure. In this case, performance as a measure of value.
- blibble 5y ago
  it's mongodb all over again gotta get dem sweet sweet benchmarks, to hell if we're persisting garbage
consp 5y ago
Considering the amount of 10uF Tantalum caps (30) on one of the bricked enterprise SSDs I opened I'm not surprised at all.
- dboreham 5y ago
  300uF is not a supercap.
  - consp 5y ago
    Never said they were and you don't need several seconds of power buffer for this purpose. They might just as well serve a different purpose like I know it just an observation. Either explain why why you definitely need F range for this or don't comment. They are rated 16v so take that into account in your explanation.
- pixl97 5y ago
  You should have cooling in your datacenter.
  - consp 5y ago
    All of your assumptions are wrong... (cause of failure, location, usage)
fire 5y ago
Any ides how one finds out whether a drive is actually capped properly to handle power off data loss?

dboreham 5y ago

Understood. Low-end drives have always done this because: performance.