Re: Understanding EXPLAIN

Robert Lichtenberger <r(dot)lichtenberger(at)synedra(dot)com> writes:
> I am trying to fully understand, how costs for queries are computed.
> Taking the following example: ...
> Index Scan using test_pkey on test (cost=0.00..8.27 rows=1 width=3)
> Index Cond: ((name)::text = '4'::text)

> The value I want to understand is 8.27. From reading the book
> "PostgreSQL 9.0 High Performance" I know, that we have one index page
> read (random page read, cost=4.0) and one database row read (random page
> read, cost=4.0) which comes up to a total of 8.0. But where are the
> missing 0.27 from?

I think you're neglecting CPU costs. We're going to charge at least one
cpu_operator_cost, one cpu_index_tuple_cost, and one cpu_tuple_cost on
top of the I/O costs. Now that still only adds 0.0175 at the default
settings, but there are various other second-order contributions. For
the whole story, there's no substitute for taking a look at the source
code; see cost_index here:
http://git.postgresql.org/gitweb/?p=postgresql.git;a=blob;f=src/backend/optimizer/path/costsize.c;h=885d8558c319fd283df351c2c8e062a449b72d3c;hb=HEAD#l208
which largely depends on btcostestimate and genericcostestimate here:
http://git.postgresql.org/gitweb/?p=postgresql.git;a=blob;f=src/backend/utils/adt/selfuncs.c;h=6d78068476e520f7dd2da6c0c8d48d93e0649768;hb=HEAD#l6003

In a quick look through that, I think the largest second-order component
is this charge in genericcostestimate:

6214 * We also add a CPU-cost component to represent the general costs of
6215 * starting an indexscan, such as analysis of btree index keys and initial
6216 * tree descent. This is estimated at 100x cpu_operator_cost, which is a
6217 * bit arbitrary but seems the right order of magnitude. (As noted above,
6218 * we don't charge any I/O for touching upper tree levels, but charging
6219 * nothing at all has been found too optimistic.)
...
6226 *indexTotalCost += num_sa_scans * 100.0 * cpu_operator_cost;

which accounts for 0.25 cost units at the default cpu_operator_cost setting.

regards, tom lane