3 ms·
I stand amazed that we still haven't gotten back to what AFS had back in 1992 or what VAX had back in 1986.
by bsder 11y ago
I stand amazed that we still haven't gotten back to what AFS had back in 1992 or what VAX had back in 1986.
- pjc50 11y ago.. which was what, exactly? Could you elaborate please?
- notacoward 11y agoI'm not the GP, but I think I can answer for AFS. To this day, nothing has come along that handles the global-scale filesystem use case as well as AFS did way back when. What we have instead are these: * Parallel filesystems that fall apart at more than a few ms latency. * Geo-distributed filesystems that are too slow/insecure/unstable to be useful for anything that really matters. * Ways to sync separate remote filesystems, but with no pretense of consistency. * Things that don't satisfy even basic filesystem semantics (e.g. object stores). * Things that are proprietary, and thus useless/irrelevant outside of their originating organizations (e.g. Google). That's why there are still some very large installations of AFS or its lineal descendants; I know of one that's 50K+ machines. I once had a dream of creating such a successor, but have largely abandoned it. It's not that the technical challenges aren't both interesting and solvable. It's more that the financial, logistical, and personnel aspects of getting such a difficult project completed and used by anyone would drown out all of the fun. I know a few others who have reached the same conclusion, sometimes up front and sometimes after losing years to the effort. The only way such a thing will ever get done is if some deep-pockets org needs it enough to fund and support its development, while also being fully committed to open source. Unfortunately, every org with such capability seems satisfied with one of the alternatives mentioned above.
- cbsmith 11y ago> It's more that the financial, logistical, and personnel aspects of getting such a difficult project completed and used by anyone would drown out all of the fun. Nailed it. I would argue that MapR is basically a company that did exactly this and figured out a way to monetize it by packaging it as part of a map/reduce system.
- bsder 11y agoFile versions--who needs a revision control system when the filesystem can keep track Extended attributes Fine-grained access controls Functionality even over high-latency links. Local caching and consistency control. Survival even through netsplits. Quorum and consistency management on the fly (add and remove computers from the cluster without killing yourself) I could go on much further, but it will just make me cry.
- cbsmith 11y ago> File versions--who needs a revision control system when the filesystem can keep track I think that is looking at it the wrong way... It's more, "who needs the complexity of a version control system?". Turns out, not everyone. > Extended attributes Umm... that has been there since forever. NFS is the only place where you struggle for EA. Again though... turns out it is not as important a feature for most people. > Fine-grained access controls The mother of all security audit nightmares. > Functionality even over high-latency links. NFSv4 and CIFS worked out perfectly functional solutions to this ages ago... and those are the filesystems that predated AFS. > Local caching and consistency control. Not surprisingly, same as above. > Survival even through netsplits. NFS 4.1 has this, not to mention most modern distributed filesystems, and while AFS has it, the problem can be complex enough one sometimes wonders if it really should be solved above the filesystem. > Quorum and consistency management on the fly (add and remove computers from the cluster without killing yourself) Well, NFS always had a pretty simple solution to this: stateless clients, and I can't think of a post-AFS filesystem where this wasn't handled in the local caching solution.
- bsder 11y agoNFSv4 did NOT predate either the VAX filesystem or AFS. I can tell you that as of 2000, Linux didn't even have a functional NFS lockd. FreeBSD's worked, but didn't handle partial file locking. NFSv4 was a pile of fail for a VERY long time. Maybe it's better now, but given that everybody still seems to use Samba (or sneakernet) for file sharing, I'm wouldn't bet on it.
- 11y ago