Introduction
The PostgreSQL regression tests are a comprehensive set of tests for the
SQL implementation embedded in PostgreSQL. They test standard SQL
operations as well as the extended capabilities of PostgreSQL.
The regression tests were originally developed by Jolly Chen and Andrew Yu,
and were extensively revised/repackaged by Marc Fournier and Thomas Lockhart.
From PostgreSQL v6.1 onward the regression tests are current for every
official release.
Some properly installed and fully functional PostgreSQL installations
can fail some of these regression tests due to artifacts of floating point
representation and time zone support. The current tests are evaluated
using a simple "diff" algorithm, and are sensitive to small system
differences. For apparently failed tests, examining the differences
may reveal that the differences are not significant.
Preparation
To prepare for regression testing, do "make all" in the regression test
directory. This compiles a 'C' program with PostgreSQL extension functions
into a shared library. Localized SQL scripts and output-comparison
files are also created for the tests that need them. The localization
replaces macros in the source files with absolute pathnames and user names.
Normally, the regression tests should be run as the postgres user since
the 'src/test/regress' directory and sub-directories are owned by the
postgres user. If you run the regression test as another user the
'src/test/regress' directory tree must be writeable to that user.
It was formerly necessary to run the postmaster with system time zone
set to PST, but this is no longer required. You can run the regression
tests under your normal postmaster configuration. The test script will
set the PGTZ environment variable to ensure that timezone-dependent tests
produce the expected results.
Directory Layout
input/ .... .source files that are converted using 'make all' into
some of the .sql files in the 'sql' subdirectory
output/ ... .source files that are converted using 'make all' into
.out files in the 'expected' subdirectory
sql/ ...... .sql files used to perform the regression tests
expected/ . .out files that represent what we *expect* the results to
look like
results/ .. .out files that contain what the results *actually* look
like. Also used as temporary storage for table copy testing.
Running the regression test
If you have previously run the regression test for a different Postgres
release, make sure you have up-to-date comparison files by doing
make clean all
The regression test is invoked with the command:
make runtest
or you can do
make runcheck
which invokes a parallel form of the regress tests, and does not
need an already-installed postmaster. Instead, runcheck creates
a temporary installation under the regress directory.
Comparing expected/actual output
The results are in files in the ./results directory. These results
can be compared with results in the ./expected directory using 'diff'.
(The test script now does this for you, and leaves the differences
in ./regression.diffs.)
The files might not compare exactly. The following paragraphs attempt
to explain the differences.
Once the output files have been verified for a particular platform,
it is possible to provide new platform-specific comparison files,
so that future test runs won't report bogus "failures". See
'Platform-specific comparison files', below.
Error message differences
Some of the regression tests involve intentional invalid input values.
Error messages can come from either the Postgres code or from the host
platform system routines. In the latter case, the messages may vary
between platforms, but should reflect similar information. These
differences in messages will result in a "failed" regression test which
can be validated by inspection.
DATE/TIME differences
Most of the date and time results are dependent on timezone environment.
The reference files are generated for timezone PST8PDT (Berkeley,
California) and there will be apparent failures if the tests are not
run with that timezone setting. The regression test driver sets
environment variable PGTZ to PST8PDT to ensure proper results.
There appear to be some systems which do not accept the recommended syntax
for explicitly setting the local time zone rules; you may need to use
a different PGTZ setting on such machines.
Some systems using older timezone libraries fail to apply daylight-savings
corrections to pre-1970 dates, causing pre-1970 PDT times to be displayed
in PST instead. This will result in localized differences in the test
results.
FLOATING POINT differences
Some of the tests involve computing 64-bit (FLOAT8) numbers from table
columns. Differences in results involving mathematical functions of
FLOAT8 columns have been observed. These differences occur where
different operating systems are used on the same platform ie:
BSDI and SOLARIS on Intel/86, and where the same operating system is
used used on different platforms, ie: SOLARIS on SPARC and Intel/86.
Human eyeball comparison is needed to determine the real significance
of these differences which are usually 10 places to the right of
the decimal point.
Some systems signal errors from pow() and exp() differently from
the mechanism expected by the current Postgres code.
POLYGON differences
Several of the tests involve operations on geographic data about the
Oakland/Berkley CA street map. The map data is expressed as polygons
whose vertices are represented as pairs of FLOAT8 numbers (decimal
latitude and longitude). Initially, some tables are created and
loaded with geographic data, then some views are created which join
two tables using the polygon intersection operator (##), then a select
is done on the view.
When comparing the results from different platforms, differences occur
in the 2nd or 3rd place to the right of the decimal point. The SQL
statements where these problems occur are the following:
QUERY: SELECT * from street;
QUERY: SELECT * from iexit;
Random differences
There is at least one test case in random.out which is intended to produce
random results. This causes random to fail the regression testing.
Typing "diff results/random.out expected/random.out" should produce only
one or a few lines of differences for this reason, but other floating
point differences on dissimilar architectures might cause many more
differences. See the release notes below.
The 'expected' files
The ./expected/*.out files were adapted from the original monolithic
'expected.input' file provided by Jolly Chen et al. Newer versions of these
files generated on various development machines have been substituted after
careful (?) inspection. Many of the development machines are running a
Unix OS variant (FreeBSD, Linux, etc) on Ix86 hardware.
Platform-specific comparison files
Since some of the tests inherently produce platform-specific results,
we have provided a way to supply platform-specific result comparison
files. Frequently, the same variation applies to multiple platforms;
rather than supplying a separate comparison file for every platform,
there is a mapping file that defines which comparison file to use.
So, to eliminate bogus test "failures" for a particular platform,
you must choose or make a variant result file, and then add a line
to the mapping file, which is "resultmap".
Each line in the mapping file is of the form
testname/platformnamepattern=comparisonfilename
The test name is just the name of the particular regression test module.
The platform name pattern is a pattern in the style of expr(1) (that is,
a regular expression with an implicit ^ anchor at the start). It is matched
against the platform name as printed by config.guess. The comparison
file name is the name of the substitute result comparison file.
For example: the int2 regress test includes a deliberate entry of a value
that is too large to fit in int2. The specific error message that is
produced is platform-dependent; our reference platform emits
ERROR: pg_atoi: error reading "100000": Numerical result out of range
but a fair number of other Unix platforms emit
ERROR: pg_atoi: error reading "100000": Result too large
Therefore, we provide a variant comparison file, int2-too-large.out,
that includes this spelling of the error message. To silence the
bogus "failure" message on HPPA platforms, resultmap includes
int2/hppa=int2-too-large
which will trigger on any machine for which config.guess's output
begins with 'hppa'. Other lines in resultmap select the variant
comparison file for other platforms where it's appropriate.
Current release notes (Thomas.Lockhart@jpl.nasa.gov)
The regression tests have been adapted and extensively modified for the
v6.1 release of PostgreSQL.
Three new data types (datetime, timespan, and circle) have been added to
the native set of PostgreSQL types. Points, boxes, paths, and polygons
have had their output formats made consistant across the data types.
The polygon output in misc.out has only been spot-checked for correctness
relative to the original regression output.
PostgreSQL v6.1 introduces a new, alternate optimizer which uses "genetic"
algorithms. These algorithms introduce a random behavior in the ordering
of query results when the query contains multiple qualifiers or multiple
tables (giving the optimizer a choice on order of evaluation). Several
regression tests have been modified to explicitly order the results, and
hence are insensitive to optimizer choices. A few regression tests are
for data types which are inherently unordered (e.g. points and time
intervals) and tests involving those types are explicitly bracketed with
"set geqo to 'off'" and "reset geqo".
The interpretation of array specifiers (the curly braces around atomic
values) appears to have changed sometime after the original regression
tests were generated. The current ./expected/*.out files reflect this
new interpretation, which may not be correct!
The float8 regression test fails on at least some platforms. This is due
to differences in implementations of pow() and exp() and the signaling
mechanisms used for overflow and underflow conditions.
The "random" results in the random test should cause the "random" test
to be "failed", since the regression tests are evaluated using a simple
diff. However, "random" does not seem to produce random results on my
test machine (Linux/gcc/i686).
Sample timing results
Timing under Linux 2.0.27 seems to have a roughly 5% variation from run
to run, presumably due to the timing vagaries of multitasking systems.
Time System
06:12 Pentium Pro 180, 32MB, Linux 2.0.30, gcc 2.7.2 -O2 -m486
12:06 P-100, 48MB, Linux 2.0.29, gcc
39:58 Sparc IPC 32MB, Solaris 2.5, gcc 2.7.2.1 -O -g
424f0edcb8