Scaleout Metadata File Systems already store much of your data. What are they?

jim-dowling.medium.com

40 points·jamesblonde··3 comments

Plenty of wrong in here about Colossus, which makes me suspect the rest of it could be wrong, too. For instance, the article suggests that atomic rename in Colossus may be possible within a "volume", which is not a thing in Colossus but the author seems to have meant either tablet or node (tablet server). That is a misunderstanding of Bigtable, which not only fails to offer cross-tablet transactions, but also does not offer multi-row transactions even within a tablet.

All the stuff at the end of the article descends from this misunderstanding and is therefore tainted by it. And what's with the obsession over atomic rename? It's a crutch. Bigtable doesn't need it therefore Colossus doesn't provide it. This is clear from Denis Serenyi's 2017 keynote at PDSW-DISCS, which states both that Colossus semantics are derived from Bigtable's requirements and that "Directories unnecessary". It's not hierarchical, it's a big flat namespace where you can put in some slashes if it makes you feel better.

Atomic rename is a mechanism to provide compare-and-swap, upon which you can build higher level things like transaction commit in a database built on top of the distributed file system.

Atomic rename is useful, but one of the points of not having it is that it makes these giant distributed file systems much much easier.

Building DBs on top of these systems is done, but they always try to be shard aware to avoid cross shard interactions. Atomic rename across shards tends to suck. These interactions tend to be where lots of not fun problems with latency and errors happen.

Reply on news.ycombinator.com