ci: the nightly's coverage job installs pg_partman, hll and tdigest (#1382)

.github/workflows/nightly.yml

@@ -419,7 +419,7 @@ jobs:
         run: |
           set -euo pipefail
           v=$(apt-cache policy \
-                postgresql-18 postgresql-server-dev-18 \
+                postgresql-18 postgresql-server-dev-18 postgresql-18-partman postgresql-18-hll postgresql-18-tdigest \
                 liblz4-dev libzstd-dev zlib1g-dev python3-pip lcov \
               | awk '/Candidate:/ {print $2}' | sort)
           # $ImageOS, from the shell, where it actually has a value. Reported
@@ -449,7 +449,7 @@ jobs:
           mkdir -p "$HOME/apt-archives/partial"
           sudo timeout 600 apt-get -o Acquire::Retries=5 \
             -o Dir::Cache::archives="$HOME/apt-archives" install -y --no-install-recommends \
-            postgresql-18 postgresql-server-dev-18 \
+            postgresql-18 postgresql-server-dev-18 postgresql-18-partman postgresql-18-hll postgresql-18-tdigest \
             liblz4-dev libzstd-dev zlib1g-dev python3-pip lcov 2>&1 | tee /tmp/apt-install.log
           # Installed with sudo because the suites run under sudo; a plain
           # pip3 install lands in ~/.local where root cannot see it, which is how

test/check_ledger.tsv

@@ -1790,6 +1790,12 @@ harness_selftest	630-a-build-with-no-pg-config-must	and the refusal is make's ow
 harness_selftest	630-a-build-with-no-pg-config-must	and with stdin closed it refuses, rather than exiting 0 having done nothing	15;16;17;18;19	2026-10-13	sed given </dev/null but no refusal: make goes on with an empty PGXS and exits 0;the Makefile before #1344: an empty PGXS hands the CC probe's sed no file, so it reads stdin;the refusal fires but does not name PG_CONFIG
 harness_selftest	630-a-build-with-no-pg-config-must	premise: a reader of this stdin waits rather than reaching EOF	15;16;17;18;19	never	-
 harness_selftest	630-a-build-with-no-pg-config-must	premise: the selftest knows where the Makefile it drives is	15;16;17;18;19	never	-
+harness_selftest	640-every-suite-job-installs-the-extensions	every job that runs the suites installs every extension package (#1382)	15;16;17;18;19	2026-10-14	nightly.yml's coverage job without postgresql-18-partman, -hll and -tdigest (main before #1382's fix)
+harness_selftest	640-every-suite-job-installs-the-extensions	premise: both workflows are present	15;16;17;18;19	never	-
+harness_selftest	640-every-suite-job-installs-the-extensions	the detector names a suite job that lacks a package the others install	15;16;17;18;19	2026-10-14	640's _ej_missing never reports a missing package;640's job parser matches no job key
+harness_selftest	640-every-suite-job-installs-the-extensions	the detector names a workflow in which it finds no suite job	15;16;17;18;19	2026-10-14	640's _ej_missing without its no-suite-job guard
+harness_selftest	640-every-suite-job-installs-the-extensions	the detector names a workflow it cannot read	15;16;17;18;19	2026-10-14	640's _ej_missing without its unreadable-file guard
+harness_selftest	640-every-suite-job-installs-the-extensions	the suite-running jobs are exactly the five this part knows	15;16;17;18;19	2026-10-14	640's job scan no longer counts a cluster job that runs pytest
 import_export_privilege	import_export_privilege	export_arrow refuses SELECT of a table the caller cannot read	15;16;17;18;19	never	-
 import_export_privilege	import_export_privilege	export_parquet refuses SELECT of a table the caller cannot read	15;16;17;18;19	never	-
 import_export_privilege	import_export_privilege	get_storage_id refuses a caller who cannot SELECT the relation	15;16;17;18;19	never	-

test/check_ledger_budget.txt

@@ -1955,4 +1955,16 @@ suites_not_covered 244
 #     never         1895 - 1 = 1894
 #     dated          571 + 1 =  572
 #     1894 + 572 == 2466
-checks_never_observed_red 1894
+#
+# #1382, the nightly's coverage job installs the extension packages (jdatcmd):
+# 1894 -> 1895. harness_selftest part 640 enters with six checks from five-major
+# runs (1245/1245 on each of 15-19). Five are dated: the job check by main's
+# nightly.yml (coverage lacked postgresql-18-partman, -hll and -tdigest), and the
+# detector, unreadable-file, no-suite-job and job-set checks by four mutations of
+# the part itself. The premise that both workflow files exist arrives never.
+#
+#     ledger rows   2466 + 6 = 2472
+#     never         1894 + 1 = 1895
+#     dated          572 + 5 =  577
+#     1895 + 577 == 2472
+checks_never_observed_red 1895

test/pytest/TESTS.md

@@ -151,6 +151,7 @@ behaviour, the source of that number is named.
 - [103. test_collation_change.py: a collation change without a rewrite leaves columnar answering as heap does](#103-test_collation_changepy-a-collation-change-without-a-rewrite-leaves-columnar-answering-as-heap-does)
 - [104. test_build_needs_pg_config.py: a build with no pg_config refuses, it does not wait on stdin](#104-test_build_needs_pg_configpy-a-build-with-no-pg_config-refuses-it-does-not-wait-on-stdin)
 - [105. test_import_export_privilege.py: import and export check the table's privilege, serial and parallel](#105-test_import_export_privilegepy-import-and-export-check-the-tables-privilege-serial-and-parallel)
+- [106. test_every_suite_job_installs_the_extensions.py: every job that runs the suites installs the extension packages](#106-test_every_suite_job_installs_the_extensionspy-every-job-that-runs-the-suites-installs-the-extension-packages)
 
 ## 1. How to read a test in here
 
@@ -6829,3 +6830,20 @@ The pytest twin of `test/import_export_privilege.sh` (#559, #566), with the para
 | `test_a_table_the_caller_holds_no_grant_on_is_refused` | the seven deny arms, serial and parallel, in the shell's order; then `victim` keeps exactly its rows and no file of it was written |
 | `test_the_owner_is_not_refused` | the owner's own exports, storage id and import are not refused |
 | `test_the_owner_runs_the_parallel_twins` | the owner's `parallel_copy` loads its 50 rows, and its `parallel_export_parquet` writes some |
+
+## 106. test_every_suite_job_installs_the_extensions.py: every job that runs the suites installs the extension packages
+
+The pytest half of selftest part 640 (#1382). Suites such as `partman_columnar` and `approx_aggregates` need a PGDG extension package (`postgresql-N-partman`, `-hll`, `-tdigest`) and record a FAIL without it. The nightly's coverage job ran those suites without the packages, and the nightly went red. So the list is derived, not trusted: every job that runs the suites must install every extension package any such job installs.
+
+READ, WITH ITS OWN READER. Python reads `ci.yml` and `nightly.yml` line by line, where the part uses awk. A job runs the suites when a line that is not a comment calls `run_all_versions.sh` or `run_coverage.sh`, or runs `pytest -q` in a job whose key names a cluster. The detector names a workflow it cannot read or finds no suite job in, so an empty answer can only mean nothing is missing. The check names match the part's.
+
+Needs no cluster and no driver.
+
+### Every test
+
+| test | what it holds |
+| --- | --- |
+| `test_both_workflows_are_present` | both workflow files exist (premise) |
+| `test_the_detector_names_what_it_finds_and_what_it_cannot_read` | on workflows written to fail, the detector names the job that lacks a package, a workflow it cannot read, and a workflow with no suite job |
+| `test_the_suite_jobs_are_the_five_known` | the suite-running jobs are exactly `ci.yml`'s `pytest-cluster` and `suites`, and `nightly.yml`'s `coverage`, `pytest-cluster` and `suites`, so a rename is red |
+| `test_every_suite_job_installs_every_extension_package` | no suite job lacks a package another installs |

test/pytest/expected_tests.txt

@@ -281,7 +281,9 @@
 # of selftest part 630, collected over the NO_CLUSTER files on this tree.
 # guard_tests 427 -> 429 for #1236's two gate arms in test_mutation_ledger.py (a dated
 # premise is refused, and its control), by collection over the NO_CLUSTER files on this tree.
-guard_tests 429
+# guard_tests 429 -> 433 for #1382's test_every_suite_job_installs_the_extensions.py, four
+# tests, the twin of selftest part 640, by collection over the NO_CLUSTER files.
+guard_tests 433
 
 # The complement: tests that need the driver and a throwaway cluster. Until #1016 these ran
 # in no CI job at all -- a quarter of the corpus, green when somebody ran them by hand and

test/pytest/test_every_suite_job_installs_the_extensions.py

@@ -0,0 +1,127 @@
+"""Every job that runs the suites installs the same extension packages (#1382).
+
+The pytest half of selftest part 640, with its own reader: Python, line by line,
+where the part uses awk. The check names match.
+
+Suites such as partman_columnar and approx_aggregates need a PGDG extension
+package (postgresql-N-partman, -hll, -tdigest), and record a FAIL without it
+(#1215). The nightly's coverage job ran those suites without the packages, and
+the nightly went red (#1382). So the list is derived, not trusted: every job that
+runs the suites must install every extension package any such job installs.
+
+A job runs the suites when a line that is not a comment calls run_all_versions.sh
+or run_coverage.sh, or runs `pytest -q` in a job whose key names a cluster. The
+suite jobs are also named as a set, so a renamed cluster job cannot drop out
+quietly. Needs no cluster and no driver.
+"""
+
+import pathlib
+import re
+
+HERE = pathlib.Path(__file__).resolve().parent
+WORKFLOWS = HERE.parent.parent / ".github" / "workflows"
+CI = WORKFLOWS / "ci.yml"
+NIGHTLY = WORKFLOWS / "nightly.yml"
+
+JOB_KEY = re.compile(r"^  ([A-Za-z_-]+):\s*$")
+RUNS = re.compile(r"run_all_versions\.sh|run_coverage\.sh")
+PACKAGE = re.compile(r"postgresql-(?:\$\{\{ matrix\.pg \}\}|\d+)-([a-z0-9]+)")
+
+
+def suite_jobs(path):
+    """{job: set of package suffixes} for each job in `path` that runs the suites."""
+    jobs, job, in_jobs = {}, None, False
+    runs, packages = set(), {}
+    for line in path.read_text().splitlines():
+        if line.startswith("jobs:"):
+            in_jobs = True
+            continue
+        if not in_jobs:
+            continue
+        key = JOB_KEY.match(line)
+        if key:
+            job = key.group(1)
+            packages[job] = set()
+            continue
+        if job is None or line.lstrip().startswith("#"):
+            continue
+        if RUNS.search(line) or ("/pytest -q" in line and "cluster" in job):
+            runs.add(job)
+        packages[job].update(PACKAGE.findall(line))
+    for name in packages:
+        if name in runs:
+            jobs[name] = packages[name]
+    return jobs
+
+
+def missing(*paths):
+    """Space-joined "file:job:pkg" for each package a suite job lacks, sorted.
+
+    A workflow it could not compare is named as "file:unreadable" or
+    "file:no-suite-job", so an empty answer can only mean nothing is missing.
+    """
+    found, out = {}, []
+    for path in paths:
+        if not path.is_file():
+            out.append(f"{path.name}:unreadable")
+            continue
+        found[path] = suite_jobs(path)
+        if not found[path]:
+            out.append(f"{path.name}:no-suite-job")
+    every = set().union(*(p for jobs in found.values() for p in jobs.values()))
+    for path, jobs in found.items():
+        for job, packages in jobs.items():
+            out.extend(f"{path.name}:{job}:{p}" for p in every - packages)
+    return " ".join(sorted(out))
+
+
+FIXTURE = """\
+jobs:
+  a:
+    steps:
+      - run: apt-get install postgresql-${{ matrix.pg }}-partman postgresql-${{ matrix.pg }}-hll
+      - run: bash test/run_all_versions.sh x
+  b:
+    steps:
+      - run: apt-get install postgresql-18 postgresql-server-dev-18 postgresql-18-partman
+      - run: bash test/run_coverage.sh x
+  c:
+    steps:
+      # c runs no suite, though this comment names run_all_versions.sh
+      - run: apt-get install postgresql-18
+"""
+
+
+def test_both_workflows_are_present(expect):
+    expect.text("yes" if CI.is_file() and NIGHTLY.is_file() else "no", "yes",
+                "premise: both workflows are present")
+
+
+def test_the_detector_names_what_it_finds_and_what_it_cannot_read(tmp_path, expect):
+    # Proven on workflows made to fail, so the real check cannot pass by finding
+    # nothing to compare.
+    w = tmp_path / "w.yml"
+    w.write_text(FIXTURE)
+    expect.text(missing(w), "w.yml:b:hll",
+                "the detector names a suite job that lacks a package the others install")
+
+    expect.text(missing(tmp_path / "absent.yml"), "absent.yml:unreadable",
+                "the detector names a workflow it cannot read")
+
+    nosuite = tmp_path / "nosuite.yml"
+    nosuite.write_text("jobs:\n  c:\n" + FIXTURE.split("  c:\n", 1)[1])
+    expect.text(missing(nosuite), "nosuite.yml:no-suite-job",
+                "the detector names a workflow in which it finds no suite job")
+
+
+def test_the_suite_jobs_are_the_five_known(expect):
+    # Not a premise: renaming, adding or deleting a job reddens it (#1236).
+    got = " ".join(sorted(f"{p.name}:{job}" for p in (CI, NIGHTLY) for job in suite_jobs(p)))
+    expect.text(got, "ci.yml:pytest-cluster ci.yml:suites nightly.yml:coverage "
+                     "nightly.yml:pytest-cluster nightly.yml:suites",
+                "the suite-running jobs are exactly the five this part knows")
+
+
+def test_every_suite_job_installs_every_extension_package(expect):
+    expect.text(missing(CI, NIGHTLY) or "none", "none",
+                "every job that runs the suites installs every extension package (#1382)")

test/pytest/test_harness_deps.py

@@ -83,6 +83,9 @@
     # #1344: drives `make -n` on the Makefile with no pg_config. Needs no cluster
     # and no driver; it never builds and never runs the shell harness.
     "test_build_needs_pg_config.py",
+    # #1382: reads the two workflow files as text, with its own Python reader. Needs
+    # no cluster and no driver, and never runs the shell harness.
+    "test_every_suite_job_installs_the_extensions.py",
     # The classifier's own controls, split out of this file so the job whose file
     # list IS this list actually runs them. This file cannot be in the list: it
     # hands cluster-bound file names to pytest, so it needs what they need.
@@ -1282,6 +1285,10 @@ def installs(text):
         "with its own Python reader, to find an EXIT trap that replaces the one setup "
         "arms (#1321). It never sources or executes either: the rule is about what "
         "the shell text says, so the text is the only place to read it",
+    "test_every_suite_job_installs_the_extensions.py":
+        "names the matrix and coverage runners' file names as data, to recognise a "
+        "workflow step that runs the suites (#1382). It reads only the two workflow "
+        "files and never reads or executes the runners themselves",
 }
 
 _SHELL_NAMES = re.compile(

test/selftest/640-every-suite-job-installs-the-extensions.sh

@@ -0,0 +1,99 @@
+# ---- every job that runs the suites installs the same extension packages -----
+#
+# Suites such as partman_columnar and approx_aggregates need a PGDG extension
+# package (postgresql-N-partman, -hll, -tdigest). Without it they record a FAIL,
+# not a pass (#1215). #1291 and #1296 added those packages to the per-PR jobs and
+# to the nightly's suites and pytest jobs, but not to the nightly's coverage job,
+# which runs the same suites under run_coverage.sh. The nightly went red (#1382).
+#
+# So the list is derived, not trusted: every job that runs the suites must
+# install every extension package any such job installs. A job runs the suites
+# when it calls run_all_versions.sh or run_coverage.sh, or runs pytest against a
+# cluster (a job whose key names a cluster). Comment lines are skipped: two jobs
+# that run no suite mention run_all_versions.sh in their comments.
+
+# _ej_scan <workflow>: one line per suite-running job, "job pkg pkg ...", where
+# each pkg is the suffix of postgresql-<major>-<suffix> on the job's lines.
+_ej_scan() {
+	awk '
+	/^jobs:/ { inj = 1; next }
+	inj && /^  [a-zA-Z_-]+:[[:space:]]*$/ { flush(); job = $1; sub(/:$/, "", job); runs = 0; delete pk; next }
+	inj && job != "" && $0 !~ /^[[:space:]]*#/ {
+		if ($0 ~ /run_all_versions\.sh|run_coverage\.sh/ || ($0 ~ /\/pytest -q/ && job ~ /cluster/)) runs = 1
+		line = $0
+		while (match(line, /postgresql-(\$\{\{ matrix\.pg \}\}|[0-9]+)-[a-z0-9]+/)) {
+			p = substr(line, RSTART, RLENGTH); sub(/.*-/, "", p); pk[p] = 1
+			line = substr(line, RSTART + RLENGTH)
+		}
+	}
+	function flush(  s, p) {
+		if (job != "" && runs) { s = job; for (p in pk) s = s " " p; print s }
+	}
+	END { flush() }' "$1"
+}
+
+# _ej_missing <workflow>...: "file:job:pkg" for each package a suite job lacks, and
+# "file:unreadable" or "file:no-suite-job" for a workflow it could not compare, so
+# an empty answer can only mean nothing is missing.
+_ej_missing() {
+	local f all
+	all="$(for f in "$@"; do [ -r "$f" ] && _ej_scan "$f" | awk '{ $1 = ""; print }' | tr ' ' '\n'; done | grep . | sort -u)"
+	for f in "$@"; do
+		if [ ! -r "$f" ]; then echo "${f##*/}:unreadable"; continue; fi
+		[ -n "$(_ej_scan "$f")" ] || echo "${f##*/}:no-suite-job"
+		_ej_scan "$f" | while read -r job pkgs; do
+			for p in $all; do
+				case " $pkgs " in *" $p "*) ;; *) echo "${f##*/}:$job:$p" ;; esac
+			done
+		done
+	done | sort | tr '\n' ' ' | sed 's/ $//'
+}
+
+_ej_ci="$TESTDIR/../.github/workflows/ci.yml"
+_ej_nl="$TESTDIR/../.github/workflows/nightly.yml"
+
+check "premise: both workflows are present" \
+	"$([ -r "$_ej_ci" ] && [ -r "$_ej_nl" ] && echo yes || echo no)" "yes"
+
+_ej_jobs="$(for f in "$_ej_ci" "$_ej_nl"; do _ej_scan "$f" | sed "s|^|${f##*/}:|"; done)"
+echo "-- suite jobs and their extension packages:"
+printf '%s\n' "$_ej_jobs" | sed 's/^/     /'
+
+# The detector proven on a workflow made to lack one package, so the real check
+# below cannot pass by finding nothing to compare.
+_ej_tmp="$(mktemp -d)"
+cat >"$_ej_tmp/w.yml" <<'EOF'
+jobs:
+  a:
+    steps:
+      - run: apt-get install postgresql-${{ matrix.pg }}-partman postgresql-${{ matrix.pg }}-hll
+      - run: bash test/run_all_versions.sh x
+  b:
+    steps:
+      - run: apt-get install postgresql-18 postgresql-server-dev-18 postgresql-18-partman
+      - run: bash test/run_coverage.sh x
+  c:
+    steps:
+      # c runs no suite, though this comment names run_all_versions.sh
+      - run: apt-get install postgresql-18
+EOF
+check "the detector names a suite job that lacks a package the others install" \
+	"$(_ej_missing "$_ej_tmp/w.yml")" "w.yml:b:hll"
+
+# An empty answer must mean "nothing missing", never "nothing read" (#1385 review):
+# a workflow the detector cannot read, or in which it finds no suite job, is named.
+sed -n '/^  c:/,$p' "$_ej_tmp/w.yml" | sed '1i jobs:' >"$_ej_tmp/nosuite.yml"
+check "the detector names a workflow it cannot read" \
+	"$(_ej_missing "$_ej_tmp/absent.yml")" "absent.yml:unreadable"
+check "the detector names a workflow in which it finds no suite job" \
+	"$(_ej_missing "$_ej_tmp/nosuite.yml")" "nosuite.yml:no-suite-job"
+rm -rf "${_ej_tmp:?}"
+
+# Not a premise: renaming, adding or deleting a job reddens it (#1236). Named as a
+# set, so a cluster job renamed without "cluster" in its key cannot drop out quietly.
+check "the suite-running jobs are exactly the five this part knows" \
+	"$(printf '%s\n' "$_ej_jobs" | awk '{ print $1 }' | sort | tr '\n' ' ' | sed 's/ $//')" \
+	"ci.yml:pytest-cluster ci.yml:suites nightly.yml:coverage nightly.yml:pytest-cluster nightly.yml:suites"
+
+check "every job that runs the suites installs every extension package (#1382)" \
+	"$(_ej_m="$(_ej_missing "$_ej_ci" "$_ej_nl")"; echo "${_ej_m:-none}")" "none"

test/selftest/parts.manifest

@@ -60,3 +60,4 @@
 610-a-declined-capability-is-not-silent.sh
 620-a-suite-s-own-trap-must-stop.sh
 630-a-build-with-no-pg-config-must.sh
+640-every-suite-job-installs-the-extensions.sh