Casey Muratori

@cmuratori @nothings @won3d Note: high level metadata is always changing on ZFS because of copy-on-write plus pointers to lower-level data.

@thatcks @nothings @won3d Yes and this worries me a bit, separate from the question I asked earlier about uberblocks in contiguous arrays.

@thatcks @nothings @won3d Specifically, because you're always rewriting the uberblocks to the same addresses, you have a "superpage" problem

@thatcks @nothings @won3d Ie., 8K may be the write size for an SSD, but the _clear_ size is actually like half a meg or something.

@thatcks @nothings @won3d So, I think this only maps well to an SSD if the firmware happens to do a full 8K page mapping.

@thatcks @nothings @won3d If they actually do mapping only at the half meg level, then you will thrash the fuck out of the drive, won't you?

@thatcks @nothings @won3d You'll be doing half-meg writes _times four_ to rewrite the uberblock arrays every time. Makes me very nervous.

@cmuratori @won3d As you go back further, ever more stuff will have been recycled. According to the link, ZFS won't even try recovering > 3.

@cmuratori @won3d Anyway, I have no idea why you're drilling down into this since (1) should have answered your question satisfactorily.

@nothings @won3d I'm not drilling down into it, I was correcting your erroneous statement about not needing the other 127 blocks :)

@cmuratori @won3d Dude, how many tweets ago did I say "you're right"?!? https://twitter.com/nothings/status/390956570609057792 … (7)

@nothings @won3d In the same tweet you said I was right about the early statement, you made _another_ erroneous statement :P

@cmuratori @won3d Dude, I said "that doesn't work ONCE [the relevant] FREE SPACE gets recycled". This is true & described in detail at link!

@nothings @won3d That bracket insert was not in the original, and that's what we spent the next tweets talking about.