Victor Giannakouris - Salalidis <victorasgs@gmail.com> writes:
> For some research purposes, I am trying to modify the existing statistics
> of some tables in the catalogs in order to change the execution plan,
> experiment with the EXPLAIN call etc.
> Concretely, what I'd like to do is to create a "fake" table with a schema
> of my choice (that's the easy part) and then modify the
> statistics(particularly, the number of tuples and the number of pages).
> Firstly, I create an empty table (CREATE TABLE newTable(....)) and then I
> update the pg_class table as well (UPDATE pg_class SET relpages = #pages
> WHERE relname='newTable').
> The problem is that, even if I set the reltuples and relpages of my choice,
> when I run the EXPLAIN clause for a query in which the 'newTable' is
> involved in (e.g. EXPLAIN SELECT * FROM newTable), I get the same cost and
> row estimation.
You can't really do it like that, because the planner always looks at
the true relation size (RelationGetNumberOfBlocks()). It uses
reltuples/relpages as an estimate of tuple density, not as hard numbers.
The reason for this is to cope with any table growth that may have
occurred since the last VACUUM/ANALYZE.
You could modify the code in plancat.c to change that, or you could
plug into the get_relation_info_hook to tweak the constructed
RelOptInfo before anything is done with it.
regards, tom lane