Changes between Initial Version and Version 1 of db/Correlations


Ignore:
Timestamp:
11/29/2010 08:01:48 PM (9 years ago)
Author:
jbecla
Comment:

--

Legend:

Unmodified
Added
Removed
Modified
  • db/Correlations

    v1 v1  
     1= Correlations =  
     2 
     3[wiki:LSSTDatabase LSST Database] 
     4 
     5 
     6In addition to the well known correlations, such as analysis of objects close to other objects, we will likely need to support a more general usecase: correlating any two attributes. Example: within some selected sample of objects (based say on color redshift and magnitude ranges), is there a correlation between two attributes (for example shape and surface brightness). It is expected no more than ~1-2 billion objects will need to be correlated per query, pairs of attributes at a time [per Tony]. 
     7 
     8First thoughts: 
     9 * Allow such correlations for Object table only (not for Sources, not for !ForcedSources) 
     10 * Maintain a dedicated server used only for such correlations, with lots of very fast storage (ideally RAM, possibly SSD). Lots = say 64 GB, which is fairly modest given we are talking about 2018 - 2028 time scale 
     11 * How to: 
     12   * use shared scan to select a sample that needs to be correlated. Extract only the columns used for the correlation, plus objectId, that is ~16 bytes of data per row, ~32GB of data for 2 billion row sample, plus indexes, should comfortably fit on ~64 GB. 
     13   * use the entire machine and all available fast storage, allow only one correlation at any given time 
     14   * build indexes on the columns to be correlated 
     15   * run the correlation in memory