I actually made some tests on my own (using generate_series) and did not find any disk space or performance issues yet. I've also found this paper from 2012 about "Sensor Data Storage Performance: SQL or NoSQL, Physical or Virtual" [2] which confirms my observations. Now, you have to know that there are about 100 attributes for the machines/tables - not only 40 - so I initially thought, it's easier to setup the schema using bit(50) and float8[50].
Did you try to use a (single) hstore or jsonb column instead where the attribute name is the key? You'd lose some type safety, but those data types can be compressed, so that might be worth the trade off Thomas
