On Thu, 2007-02-22 at 12:47 +0000, Heikki Linnakangas wrote: > One question that I'm sure someone will ask is do we need this if we > have bitmap indexes? Both aim at having a smaller index, after all. > The use cases are quite different; GIT is effective whenever you have > a table that's reasonably well-clustered. Unlike the bitmap indexam, > GIT's effectiveness doesn't depend on the number of distinct values, > in particular it works well with unique indexes. GIT is comparable to > clustered indexes in other DBMSs (in fact we might want to call GIT > that in the end).
A few thoughts: Whether its proof, or merely strong interest, we need to see where the cross-over point is between bitmap indexes (BMI) and GIT. I'm assuming there is one, but others may need convincing that there is a clear win in common use cases. Graphs, numbers etc.. On your TODO list, it would be good to differentiate between ideas and must-complete items. My suggested list would be: Must complete - cost model for optimizer - user documentation - performance tests Good to complete - duplicate key flag (should increase range of overlap with BMI) - range re-check optimization (based upon your results) - better index AM - HOT, so that GIT can provide longer term benefits Others ... I seem to recall you mentioning there was a way to record that the heap tuples are stored in monotonic order and so the sort can be avoided when you return the tuples. Is that still true? Also "the patch roughly doubles the amount of WAL generated by WAL" not sure what that means -- Simon Riggs EnterpriseDB http://www.enterprisedb.com ---------------------------(end of broadcast)--------------------------- TIP 9: In versions below 8.0, the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match