Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
Description
In OAK-6735 we have improved cost estimation for Lucene indexes, however the following case is still not working as expected: a very common property is indexes (many nodes have that property), and each value of that property is more or less unique. In this case, currently the cost estimation is the total number of documents that contain that property. Assuming the condition "property is not null" this is correct, however for the common case "property = x" the estimated cost is far too high.
A known workaround is to set the "costPerEntry" for the given index to a low value, for example 0.2. However this isn't a good solution, as it affects all properties and queries.
It would be good to be able to set the selectivity per property, for example by specifying the number of distinct values, or (better yet) the average number of entries for a given key (1 for unique values, 2 meaning for each distinct values there are two documents on average).
That value can be set manually (cost override), and it can be set automatically, e.g. when building the index, or updated from time to time during the index update, using a cardinality
estimation algorithm. That doesn't have to be accurate; we could use an rough approximation such as hyperbitbit.