[MisterRaindrop]

What does this PR do?

Fixes six gpinitsystem defects that the gpMgmt Behave suite (#2170) exposes. Each is a line or two that Cloudberry lost relative to Greenplum. They are kept together because they share gpinitsystem / gp_bash_functions.sh and the same feature file; the commit message gives each one's full reasoning.

1. Mirror port after pg_basebackup. CREATE_QES_MIRROR started the mirror before fixing the port in the postgresql.conf that pg_basebackup copied from its primary, so on a single host the mirror bound its primary's port, or collided with it. The port is now rewritten before the start. This change also writes explicit 127.0.0.1/32 and ::1/128 trust entries into the coordinator's pg_hba.conf, because a hostname entry alone breaks when the client picks ::1.
2. FORCE_FTS_PROBE never asked for a probe. The gp_request_fts_probe_scan() call was missing, so the 60 iterations re-read an unchanged catalog in a couple of seconds and gave up. gpinitsystem then exited 0 with all mirrors still down.
3. The default ETCD/FTS config was logged as [WARN]. gpinitsystem's own log scan made every run end with "warnings or errors were generated". These messages are now logged at INFO.
4. gpinitstandby got a bare -f. On any multi-host cluster without external FTS hosts, gpinitsystem passed -f with no value, gpinitstandby failed parsing its options, and gpinitsystem still exited 0 with no standby. -f is now passed only when there is a value, and the list builder's always-false [ -z FTS_HOST_STRING_LIST ] is fixed. This was first seen on a five-node EC2 cluster.
5. [ERROR] did not set the exit status in LOG_MSG. As a result, gpinitsystem --unknown-option exited 0. The scenario that should have caught this was itself passing --ignore-warnings, which is not a gpinitsystem option. That stray argument is dropped, and a scenario for the exit status is added.
6. SED_PG_CONF substitutions were unanchored. gpinitsystem -p with work_mem=999 rewrote maintenance_work_mem. The anchors are restored. CHECK_FTS's argument-count guard (copied from CHECK_ETCD) and CLEANUP_ETCD's error message are also corrected, because once fix 5 lands, the wrong guard would make gpinitsystem -F report a failure after it succeeded.

Split out of #1639 (originally b971d48, by @tuhaihe).

Test Plan

Rocky 9 build image, built from #2170's tree plus #2177 (767f12685a4, absolute TRUSTED_SHELL), with a demo cluster created the way the Behave workflow creates it. I removed the @needs_fix tags that point at either fix, then ran each feature with CI's arguments. Run A is without this PR; run B is with it, with the three scripts installed into $GPHOME/bin.

	A: without	B: with this PR
gpinitsystem.feature (--tags ~@extended)	17 passed, 16 failed	33 passed, 0 failed
gpaddmirrors.feature (--tags ~@concourse_cluster)	10 passed, 2 failed	10 passed, 2 failed (same two; not this fix)

Why the scenarios fail in A:

• "exits with status 1 when the user provides invalid option" (the new scenario) exits 0. Fix 5.
• The two "should warn but not fail when standby cannot be instantiated" scenarios fail with "all segments are not currently running". Fix 2.
• "should print FQDN in pg_hba.conf when HBA_HOSTNAMES=1": gpinitsystem's own gpstop is refused with no pg_hba.conf entry for host "::1". That is the loopback part of fix 1.
• The 12 scenarios after it fail because each one starts by stopping that cluster. gpstop is refused the same way, and the Behave suite's pg_ctl fallback raises NameError (fixed in CI: run the gpMgmt Behave suite #2170), so all of them start from a cluster that can't be stopped.
  • The run I did for gpdemo: write an absolute TRUSTED_SHELL into the generated config #2177 applied only that fix and still skipped every scenario up to and including the DCA one. In it, the locale, banner, -D and both trusted_shell scenarios passed.
  • "consistent port entry on segments postgresql.conf" was the exception: it failed with expected 15437, got 15434, which fix 1 addresses.
  • I have no clean result without this PR for the HBA_HOSTNAMES=0, standby HBA_HOSTNAMES=1 and DCA scenarios. They pass with it.

#2170 tags these scenarios @needs_fix pointing at this change. Whichever PR lands second removes those tags.

Type of Change

• Bug fix (non-breaking change)

Checklist

• This PR contains AI-assisted code generation

[leborchuk]

LGTM, old code just didn't work with some options. Of course we should fix it.

[Alena0704]

LGTM. I suggest adding regression test for SED_PG_CONF, particularly preserving maintenance_work_mem when changing work_mem.

test

diff --git a/gpMgmt/test/behave/mgmt_utils/gp_bash_functions.feature b/gpMgmt/test/behave/mgmt_utils/gp_bash_functions.feature
index 35e6629020e..bfecbbc4a89 100644
--- a/gpMgmt/test/behave/mgmt_utils/gp_bash_functions.feature
+++ b/gpMgmt/test/behave/mgmt_utils/gp_bash_functions.feature
@@ -1,5 +1,16 @@
 @gp_bash_functions.sh
 Feature: gp_bash_funtions.sh unit test cases
+    @sed_pg_conf
+    Scenario Outline: SED_PG_CONF changes only the requested parameter
+        Then SED_PG_CONF preserves unrelated settings on "<host>" with KEEP_PREV "<keep_prev>"
+
+        Examples:
+            | host      | keep_prev |
+            | local     | 0         |
+            | local     | 1         |
+            | localhost | 0         |
+            | localhost | 1         |
+
     Scenario: REMOTE_EXECUTE_AND_GET_OUTPUT returns proper output when no banner is set
         Given a standard local demo cluster is running
         Then source gp_bash_functions and run simple echo
diff --git a/gpMgmt/test/behave/mgmt_utils/steps/gp_bash_functions.py b/gpMgmt/test/behave/mgmt_utils/steps/gp_bash_functions.py
new file mode 100644
index 00000000000..084afc01c87
--- /dev/null
+++ b/gpMgmt/test/behave/mgmt_utils/steps/gp_bash_functions.py
@@ -0,0 +1,48 @@
+import os
+from pathlib import Path
+import subprocess
+import tempfile
+
+from behave import then
+
+
+@then('SED_PG_CONF preserves unrelated settings on "{host}" with KEEP_PREV "{keep_prev}"')
+def check_sed_pg_conf_parameter_boundary(context, host, keep_prev):
+    # Use an active work_mem setting so the replacement branch is exercised;
+    # a file containing only the default #work_mem would take the append path.
+    unrelated_lines = [
+        "maintenance_work_mem = 65536",
+        "# work_mem should not change inside a comment",
+    ]
+    original = "  work_mem = 4096\n" + "\n".join(unrelated_lines) + "\n"
+    functions = Path(os.environ["GPHOME"]) / "bin/lib/gp_bash_functions.sh"
+
+    # localhost uses the real SSH branch against the same temporary file.
+    # No database is required; SSH must be configured as for the other tests
+    # in gp_bash_functions.feature.
+    with tempfile.TemporaryDirectory(prefix="sed_pg_conf_") as temp_dir:
+        config = Path(temp_dir) / "postgresql.conf"
+        config.write_text(original)
+        script = '''
+source "$1" || exit $?
+LOG_FILE="$2"
+SED_PG_CONF "$3" work_mem work_mem=999 "$4" "$5"
+'''
+        result = subprocess.run(
+            ["bash", "-c", script, "sed_pg_conf_test", str(functions),
+             str(Path(temp_dir) / "test.log"), str(config), keep_prev,
+             "" if host == "local" else host],
+            stdout=subprocess.PIPE, stderr=subprocess.PIPE,
+            universal_newlines=True, timeout=60,
+        )
+        assert result.returncode == 0, (
+            "SED_PG_CONF failed: stdout={!r}, stderr={!r}".format(
+                result.stdout, result.stderr)
+        )
+        lines = config.read_text().splitlines()
+        assert lines[1:] == unrelated_lines, (
+            "SED_PG_CONF modified unrelated settings or comments: {!r}".format(lines)
+        )
+        assert lines[0].split("#", 1)[0].strip() == "work_mem=999", (
+            "SED_PG_CONF did not update work_mem: {!r}".format(lines)
+        )

[MisterRaindrop]

Thanks, added your scenario as is in 8752fe3, credited in the commit message. All four examples fail against main's gp_bash_functions.sh, where maintenance_work_mem and the comment are both rewritten, and pass with this PR.

[Alena0704]

No more issues. LGTM. Thank you!)