HN Explorer

9 comments

[ 3.7 ms ] story [ 40.3 ms ] thread
vibhavs
Very interesting white paper. (I didn't realize that Clustix is a YCombinator company -- they're class of 2006.)
bayareaguy
Sounds a lot like the HyPra / ClustRa system.

http://books.google.com/books?id=KH_kfSsMu80C&dq=HypRa+C...

phliar
You mean the Clustra that was acquired by Sun and renamed HADB? I used it, and as far as I can remember, the HADB fault-tolerant story was that each node has a mirror that takes over if the first fails; and it's basically all in-memory. The Clustrix redundancy granularity is much finer, and everything is on disk.
bayareaguy
I never worked with one, but I believe the HyPra system included disks (as shown on page 148. http://books.google.com/books?id=KH_kfSsMu80C&pg=PA148 ). I recall it was one of the earlier systems to achieve high transaction rates by logging updates to memories of multiple replicas (described on page 181: http://books.google.com/books?id=KH_kfSsMu80C&pg=PA181 ).
phliar
Clustra/HADB did use disks (can't have all your data disappear if you turn off the cluster!) but a running system was in-mem in operation -- if your database was much bigger than the RAM available it was unusable. (I never used Hypra.)
phliar
For anyone else interested in Clustra: Sun used to have decent HADB docs and whitepapers on bigadmin, but I can't find them any more in the new Oracle world. The two Sun projects that used HADB were Glassfish (clustered fault-tolerant app server, used HADB for session persistence data) and Honeycomb (clustered fault-tolerant archival storage, used HADB for metadata). Examples of how these systems used Clustra are on the web; in both cases the database size was small.
SpikeGronim
What's interesting to me is that they're selling this as an appliance, and a high end one at that. They're using Infiniband networking and seven SSDs per appliance. That's some serious cost. Their overall software design as described in the white paper isn't that new or interesting IMO. Teradata and others have been doing very similar things for a long time.
sergei
As far as I know, no has a system like ours that can actually scale OLTP workloads. Warehousing/Analytics databases are very different beast. Can you point me at an OLTP system that I can expand without taking any loss of availability? How about supporting online schema changes?

Infiniband ports cost less than 10GE right now. There's really nothing that ties our architecture to IB. It's all about price/performance. When a better/cheaper technology becomes more available, we'll switch to that.