How far are projections pushed down the execution tree?
Consider a table and a query referring to only a subset of the columns
in that table. How early in the query evaluation is the projection
carried out?
Are the columns to be selected filtered out as early as in the very
access method that reads the table rows from the buffer, or are the
projection handled later, after the whole row has been fetched by the
access method?
Does it depend on the complexity of the query, how far down the three
that the projection is handled out?
Thanks!
tmp wrote:
Consider a table and a query referring to only a subset of the columns
in that table. How early in the query evaluation is the projection
carried out?Are the columns to be selected filtered out as early as in the very
access method that reads the table rows from the buffer, or are the
projection handled later, after the whole row has been fetched by the
access method?Does it depend on the complexity of the query, how far down the three
that the projection is handled out?
It depends on the kind of plan chosen. In general the planner delays
projection and prefers to do it as high as possible when it can pass a
lower-level tuple to an upper node as is, but if it needs to e.g sort,
projections are done before the sort to avoid carrying unnecessary data
through the sort.
EXPLAIN VERBOSE (in 8.4 upwards) shows the columns that are being output
from node, you can use that to examine the behavior.
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com