3 ms·
...except that HDFS has never been a distributed filesystem, despite the name. It's an object store with just enough hackery so that a particular set of apps wi
by notacoward 7y ago
...except that HDFS has never been a distributed filesystem, despite the name. It's an object store with just enough hackery so that a particular set of apps with very limited needs could be modified in a short time to use it instead of a filesystem. Arguably that's still a big deal, but nowhere near as impressive or generally useful as a distributed filesystem. To me it's a bit amazing that anyone involved in its design still has a job in this industry.
- vkazanov 7y ago...ok, then what are the alternatives? And why is it so bad? Something that can handle hundreds of terabytes on hundreds of machines and provides useful tools on top of the whole thing (Spark, Hive, etc)?
- notacoward 7y agoWell, rather topically, there's MapR. Depending on the desired scale and level of fidelity to real-filesystem semantics, one might also count Gluster, Ceph, and Lustre. Don't know if PVFS is still around. There are more proprietary offerings from IBM (GPFS/SpectrumScale), EMC, NetApp, etc. plus a plethora of startups. There are other object stores that don't make a pretense of being filesystems. Alternatives abound. BTW, "hundreds of terabytes on hundreds of machines" isn't interesting territory any more. Most people's needs are far smaller, so HDFS isn't much help. Those who need more generally need much more, so HDFS isn't much help again. Richer semantics are nice either way. Imagine thousands of machines with dozens of terabytes each and you might start to see the problems with HDFS's design (though you'll still be far short of the domain I work in).
- vkazanov 7y agoThe thing is that this is exactly my company's scale, and hdfs is perfectly fine here. :-) And I am curious about possible alternatives: open source, about 100-200 machines, with good support for analytics and SQlish systems.