Use fuzzy path cost tiebreaking rule in our oldest supported branches.
We've been doing it that way since 9.2, cf commit 33e99153e93b9acc,
but some recently-added regression test cases are making a few buildfarm
members fail (ie choose the "wrong" plan) in 9.0 and 9.1 due to
platform-specific roundoff differences in cost calculations. To fix,
back-port the patch that made add_path treat cost difference ratios of
less than 1e-10 as equal.
Branch
------
REL9_0_STABLE
Details
-------
http://git.postgresql.org/pg/commitdiff/9d6af7367e7b079c021c285f03128bad531b667e
Modified Files
--------------
src/backend/optimizer/util/pathnode.c | 50 +++++++++++++++++++--------------
1 file changed, 29 insertions(+), 21 deletions(-)