For ‘Google-scale storage’ you might want to substitute ‘Amazon-style storage’, because Amazon got the basic distribution and organization mechanisms right for general-purpose use (consistent hashing and peer-oriented rather than hierarchical interaction).

Flash SSDs won’t change storage centers significantly unless their per-GB prices drop far faster than predicted (to at most 10x the price of SATA storage, which is poised to drive out ‘enterprise-class’ disks unless the price gap between them narrows dramatically). Otherwise, flash SSD is just cheaper non-volatile cache - and given how far the bottom has dropped out of conventional RAM prices there’s enough room for both volatile (read) RAM cache and battery-backed (read/write) RAM cache/conventional SSD prices to drop that their performance advantages over flash SSD (especially in cases where the cache can be situated in the host close to the point of use rather than way out on the SAN somewhere) may keep flash SSD from getting much traction.

Another trend which may help keep flash SSD from gaining traction may be ‘write anywhere’ storage (which is often functionally similar to log-structured storage but more flexible in avoiding the latter’s drawbacks) - because dramatically improving a disk’s effective ‘random-write’ bandwidth removes SSD’s (at least potential) advantages in that area. NetApp’s 15-year virtual monopoly on this technology may be coming to a close even if they succeed in their current patent litigation: there’s a lot of room to work around those patents if one makes the effort, and a lot better approaches than ZFS’s.

End-to-end data integrity checks also aren’t new, but (just as with write-anywhere approaches) may become a lot more common for block-level storage as well as for file-level storage (where I still haven’t had the chance to look at ZFS’s code to see whether its implementation is as inefficient as some of its literature suggests it may be) - and in doing so will help alleviate concerns about the wisdom of substituting SATA for SAS/SCSI/FC disks.

‘Self-healing’ storage can leverage such additional integrity checks effectively to help make storage management a negative-growth profession despite continuing exponential capacity increases - if accompanied by the kind of automated configuration management that’s possible (both in terms of handling hardware changes transparently and eliminating knob-tweaking: ZFS barely scratches the surface in this area - think Drobo at the enterprise level).

Increasingly intelligent storage software running on ever-faster, ever-cheaper (and increasingly reliable/self-checking) suitably-redundant commodity hardware is what makes all the above possible: we just have to write it.

- bill

- ZFS and other end-to-end storage system designs. These designs
will ensure reliable, easy to manage storage as other trends
make storage less reliable and more complex. The flip side
of the coming death of traditional RAID 5, RAID 6, and possibly
SANs. These designs are enabled by …

- Massively multicore computing - hundreds of cores per computer
will transform storage system designs and requirements — as soon
as the software catches up. One dedicated core per disk channel?
Virtual machines managing storage pools, talking to other virtual
machines at memory bus speeds? How do we feed 100 simultaneously
executing threads each trying to work on part of the same file?

- Cloud storage - how high can we go before we run into WAN
bottlenecks? The (first order) economics and flexibility are
compelling; security issues can be solved; but no magic wand can
grow WAN speeds and drop costs to match the disk drive industry.

- Consumerization of storage - Enterprise storage market share will
continue shrinking as consumer storage needs explode. What will
enterprise storage look like with hundreds or thousands of cheap,
physically small, slow, unreliable disk drives? Will we need flash
drives as a buffer?

Should be interesting, wish I could be there!

– Rex