ci(postgres): scheduled fan-out Postgres CI (daily 3am IST + 'postgres' label gate)

.github/helper/hydrate.sh

@@ -0,0 +1,74 @@
+#!/bin/bash
+#
+# Hydrate a test shard from the setup job's artifact.
+#
+# The bench (apps, venv, node_modules, sites) is already on disk at ~/frappe-bench — the
+# workflow untar'd it from the artifact the setup job built. So there is NO bench init, no
+# asset build, and no reinstall here: just bring the DB up and restore the dump the setup job
+# baked into the bench, then start bench so tests can run. Mirrors the DB + bench-start tail of
+# install.sh. The whole point is that the expensive work happened ONCE in the setup job.
+#
+set -e
+
+ci_user="${ERPNEXT_CI_USER:-frappe}"
+db_host="${DB_HOST:-127.0.0.1}"
+dump="${CI_BASELINE_BACKUP:-/home/$ci_user/frappe-bench/test_site-db.sql.gz}"
+
+# Re-exec as the ci user (uid 1001) so bench/cache ownership matches the artifact, same as
+# install.sh. The workflow untar'd as root with -p, so the files are already owned by ci.
+if [ "$(id -u)" = "0" ] && [ "${SKIP_SYSTEM_SETUP:-0}" = "1" ] && [ "$ci_user" != "root" ]; then
+    exec su -m "$ci_user" -s /bin/bash -c \
+        "ERPNEXT_CI_USER='$ci_user' DB_HOST='$db_host' CI_BASELINE_BACKUP='$dump' bash '$0'"
+fi
+
+cd ~/frappe-bench
+
+# Start the DB on the datadir baked into the artifact. It's already populated (the setup job
+# reinstalled into this very datadir), so there is NO restore — the server comes up on the
+# existing files. This is what replaces the per-shard SQL replay.
+bash ~/frappe-bench/start-db.sh
+
+# Bring up redis (lightmode unit tests need cache + queue). In the self-hosted container we use the
+# full `bench start` (web/workers too, like install.sh). On the bare GitHub Postgres shard
+# `bench start` (honcho) lagged — it blocks the redis procs behind web/worker procs the lightmode
+# suite never uses, so the wait below burned its full timeout (~4m). There, start the two redis
+# instances directly: fast and deterministic.
+if [ "${DB:-mariadb}" = "postgres" ]; then
+    # Start redis directly as daemons — reliable and persists across steps. Do NOT route it through
+    # `bench start`: honcho tears the whole process group down if any one Procfile proc dies on the
+    # bare shard, which took redis with it (redis @ 13000 refused in Run Tests). Keeping redis
+    # independent is what makes it survive. The web server (for PDF tests) is NOT started here — a
+    # backgrounded server doesn't survive into the next step; it's started inside the Run Tests step.
+    for conf in redis_cache redis_queue; do
+        [ -f ~/frappe-bench/config/$conf.conf ] && redis-server ~/frappe-bench/config/$conf.conf --daemonize yes
+    done
+else
+    bench start >> ~/frappe-bench/bench_start.log 2>&1 &
+fi
+
+# Wait for redis, failing fast instead of silently burning minutes if it never comes up.
+cfg=~/frappe-bench/sites/common_site_config.json
+if [ -f "$cfg" ]; then
+    ports=$(python - "$cfg" <<'PY'
+import json, re, sys
+try:
+    cfg = json.load(open(sys.argv[1]))
+except Exception:
+    sys.exit(0)
+for key in ("redis_cache", "redis_queue"):
+    m = re.search(r":(\d+)", str(cfg.get(key, "")))
+    if m:
+        print(m.group(1))
+PY
+)
+    for port in $ports; do
+        up=0
+        for _ in $(seq 1 60); do
+            if (exec 3<>"/dev/tcp/127.0.0.1/$port") 2>/dev/null; then exec 3>&- 3<&-; up=1; break; fi
+            sleep 1
+        done
+        [ "$up" = "1" ] || { echo "redis did not come up on port $port"; exit 1; }
+    done
+fi
+
+echo "Hydrated: DB up on baked datadir, redis up — ready for tests."

.github/helper/install.sh

@@ -7,21 +7,106 @@ cd ~ || exit
 githubbranch=${GITHUB_BASE_REF:-${GITHUB_REF##*/}}
 frappeuser=${FRAPPE_USER:-"frappe"}
 frappecommitish=${FRAPPE_BRANCH:-$githubbranch}
+db_host=${DB_HOST:-"127.0.0.1"}
+db_user_host=${DB_USER_HOST:-"localhost"}
+wkhtmltox_deb=${WKHTMLTOX_DEB:-"/tmp/wkhtmltox.deb"}
+bench_cache_dir=${BENCH_CACHE_DIR:-}
+
+run_as_ci_user_if_needed() {
+    if [ "$(id -u)" != "0" ] || [ "${SKIP_SYSTEM_SETUP:-0}" != "1" ] || [ "${ERPNEXT_CI_NON_ROOT:-0}" = "1" ]; then
+        return
+    fi
+
+    local missing_packages=()
+    if ! command -v pkg-config >/dev/null 2>&1; then
+        missing_packages+=("pkg-config")
+    fi
+    if ! command -v mariadb_config >/dev/null 2>&1 && ! command -v mysql_config >/dev/null 2>&1; then
+        missing_packages+=("libmariadb-dev")
+    fi
+    if ! command -v crontab >/dev/null 2>&1; then
+        missing_packages+=("cron")
+    fi
+
+    if [ "${#missing_packages[@]}" -gt 0 ]; then
+        apt-get update
+        apt-get install -y --no-install-recommends "${missing_packages[@]}"
+    fi
+
+    local ci_user="${ERPNEXT_CI_USER:-frappe}"
+
+    if ! id "$ci_user" >/dev/null 2>&1; then
+        useradd --home-dir "$HOME" --no-create-home --shell /bin/bash "$ci_user"
+    fi
+
+    rm -rf ~/frappe ~/frappe-bench
+
+    local ci_dirs=(
+        "$HOME"
+        "$GITHUB_WORKSPACE"
+        "$HOME/.cache"
+        "${PIP_CACHE_DIR:-$HOME/.cache/pip}"
+        "${npm_config_cache:-$HOME/.npm}"
+        "${YARN_CACHE_FOLDER:-$HOME/.cache/yarn}"
+        "$HOME/.yarn"
+        "${UV_CACHE_DIR:-$HOME/.cache/uv}"
+        "$(dirname "$wkhtmltox_deb")"
+    )
+    if [ -n "$bench_cache_dir" ]; then
+        ci_dirs+=("$bench_cache_dir")
+    fi
+
+    # Create + own (non-recursively) the home/cache/workspace dirs before dropping to
+    # the ci user. We deliberately do NOT wipe the yarn/uv caches here so a persistent
+    # cache (mounted volume or baked image layer) stays warm across runs.
+    mkdir -p "${ci_dirs[@]}" "$HOME/.yarn"
+    chown "$ci_user:$ci_user" "${ci_dirs[@]}" "$HOME/.yarn"
+
+    export ERPNEXT_CI_NON_ROOT=1
+    exec su -m "$ci_user" -s /bin/bash -c "cd '$HOME' && bash '$GITHUB_WORKSPACE/.github/helper/install.sh'"
+}
+
+run_as_ci_user_if_needed
+
+run_ci_step() {
+    local label=$1
+    shift
+
+    echo "::group::${label}"
+    date -u
+    timeout --foreground "${CI_INSTALL_STEP_TIMEOUT:-600}" "$@"
+    local exit_code=$?
+    date -u
+    echo "::endgroup::"
+    return "$exit_code"
+}
+
+if [ -n "${GITHUB_WORKSPACE:-}" ]; then
+    git config --global --add safe.directory "$GITHUB_WORKSPACE" || true
+    git config --global --add safe.directory "$GITHUB_WORKSPACE/.git" || true
+fi
+
+rm -rf ~/frappe ~/frappe-bench
 
 # ---------------------------------------------------------------------------
 # Phase 1 — parallelise the three slow, independent setup steps:
 #   a) system packages   b) frappe-bench pip install   c) frappe git fetch
 # ---------------------------------------------------------------------------
 
-sudo apt update
+if [ "${SKIP_SYSTEM_SETUP:-0}" != "1" ]; then
+    sudo apt-get update
 
-# apt remove/install must run sequentially but can overlap with pip and git.
+    # apt remove/install must run sequentially but can overlap with pip and git.
-sudo apt remove mysql-server mysql-client
+    sudo apt-get remove -y mysql-server mysql-client
-sudo apt install libcups2-dev redis-server mariadb-client libmariadb-dev &
+    sudo apt-get install -y libcups2-dev redis-server mariadb-client libmariadb-dev &
-apt_pid=$!
+    apt_pid=$!
 
-pip install frappe-bench &
+    pip install frappe-bench &
-pip_pid=$!
+    pip_pid=$!
+else
+    apt_pid=
+    pip_pid=
+fi
 
 mkdir frappe
 (
@@ -32,84 +117,264 @@ mkdir frappe
 ) &
 clone_pid=$!
 
-wait $apt_pid
+if [ -n "$apt_pid" ]; then wait $apt_pid; fi
-wait $pip_pid
+if [ -n "$pip_pid" ]; then wait $pip_pid; fi
 wait $clone_pid
 
 pushd frappe
 git checkout FETCH_HEAD
 popd
+frappe_sha=$(git -C frappe rev-parse HEAD)
+
+get_bench_cache_archive() {
+    if [ -z "$bench_cache_dir" ]; then
+        return
+    fi
+
+    mkdir -p "$bench_cache_dir"
+
+    # Keyed on tool versions only (NOT the frappe SHA): any recent base bench works, because
+    # restore_warm_bench fast-forwards it to the exact live develop SHA. This is what lets a
+    # constantly-moving develop still hit the cache.
+    local cache_key
+    cache_key=$(
+        {
+            uname -m
+            python --version
+            node --version
+            bench --version
+        } | sha256sum | awk '{print $1}'
+    )
+
+    echo "${bench_cache_dir}/frappe-bench-base-${cache_key}.tar.zst"
+}
+
+restore_warm_bench() {
+    bench_cache_archive=$(get_bench_cache_archive)
+    [ -n "$bench_cache_archive" ] && [ -f "$bench_cache_archive" ] || return 1
+
+    echo "Restoring base bench from ${bench_cache_archive}"
+    tar --use-compress-program=unzstd -xf "$bench_cache_archive" -C ~ || return 1
+    [ -d ~/frappe-bench/apps/frappe/.git ] || return 1
+    mkdir -p ~/frappe-bench/sites ~/frappe-bench/logs
+    [ -f ~/frappe-bench/sites/apps.txt ] || printf "frappe\n" > ~/frappe-bench/sites/apps.txt
+    [ -f ~/frappe-bench/sites/common_site_config.json ] || printf "{}\n" > ~/frappe-bench/sites/common_site_config.json
+
+    # Fast-forward the restored frappe to the EXACT live develop SHA fetched in phase 1, then
+    # rebuild only what changed. The editable install means the venv tracks the new code with
+    # no reinstall. Any failure returns non-zero so the caller falls back to a full bench init.
+    if ! (
+        cd ~/frappe-bench/apps/frappe || exit 1
+        # Phase 1 already fetched ~/frappe to the exact live develop SHA. Fetch that commit
+        # straight from it (bench init names the remote 'upstream', not 'origin', and points
+        # it at this local clone — so a plain `git fetch origin` does not work).
+        git fetch --no-tags "$HOME/frappe" HEAD || exit 1
+        git checkout --force FETCH_HEAD || exit 1
+    ); then
+        echo "Fast-forward to ${frappe_sha} failed; falling back to full init"
+        rm -rf ~/frappe-bench
+        return 1
+    fi
+
+    # Pick up any frappe dependency changes since the base was built (cached → fast if none),
+    # so a develop commit that bumped requirements doesn't leave a stale venv.
+    if ! ~/frappe-bench/env/bin/python -m pip install -q -e ~/frappe-bench/apps/frappe; then
+        echo "frappe dependency refresh failed; falling back to full init"
+        rm -rf ~/frappe-bench
+        return 1
+    fi
+
+    ( cd ~/frappe-bench && CI=Yes bench build --app frappe ) || { rm -rf ~/frappe-bench; return 1; }
+    return 0
+}
+
+save_warm_bench() {
+    if [ -z "${bench_cache_archive:-}" ] || [ -f "$bench_cache_archive" ]; then
+        return
+    fi
+
+    if [ -n "$bench_cache_dir" ] && [ ! -w "$bench_cache_dir" ]; then
+        echo "Skipping warm bench save because ${bench_cache_dir} is not writable"
+        return
+    fi
+
+    local tmp_archive
+    tmp_archive="${bench_cache_archive}.${$}.tmp"
+
+    echo "Saving warm bench to ${bench_cache_archive}"
+    # Keep sites/common_site_config.json (the redis ports live there — dropping it makes the
+    # restore path fall back to a default redis port that bench start never bound, so reinstall
+    # fails with "redis ... connection refused"). Only the rebuildable sites/assets is excluded;
+    # restore_warm_bench runs `bench build` to regenerate it.
+    tar \
+        --use-compress-program="zstd -T0 -3" \
+        --exclude="frappe-bench/logs" \
+        --exclude="frappe-bench/sites/assets" \
+        -cf "$tmp_archive" \
+        -C ~ frappe-bench
+    mv "$tmp_archive" "$bench_cache_archive"
+}
 
 # ---------------------------------------------------------------------------
 # Phase 2 — bench init and site setup
 # ---------------------------------------------------------------------------
 
-bench init --skip-assets --frappe-path ~/frappe --python "$(which python)" frappe-bench
+install_whktml() {
+    # Re-use the .deb if the wkhtmltopdf cache step already restored it.
+    if [ ! -f "$wkhtmltox_deb" ]; then
+        wget -O "$wkhtmltox_deb" https://github.com/wkhtmltopdf/packaging/releases/download/0.12.6.1-2/wkhtmltox_0.12.6.1-2.jammy_amd64.deb
+    fi
+    sudo apt-get install -y "$wkhtmltox_deb"
+}
+if [ "${SKIP_WKHTMLTOX_SETUP:-0}" != "1" ]; then
+    install_whktml &
+    wkpid=$!
+else
+    wkpid=
+fi
 
-mkdir ~/frappe-bench/sites/test_site
+if ! restore_warm_bench; then
+    bench init --skip-assets --frappe-path ~/frappe --python "$(which python)" frappe-bench
+
+    cd ~/frappe-bench || exit
+
+    sed -i 's/watch:/# watch:/g' Procfile
+    sed -i 's/schedule:/# schedule:/g' Procfile
+    sed -i 's/socketio:/# socketio:/g' Procfile
+    sed -i 's/redis_socketio:/# redis_socketio:/g' Procfile
+
+    CI=Yes bench build --app frappe
+    save_warm_bench
+fi
+
+if [ -n "$wkpid" ]; then wait $wkpid; fi
+
+mkdir -p ~/frappe-bench/sites/test_site
 
 if [ "$DB" == "mariadb" ];then
     cp -r "${GITHUB_WORKSPACE}/.github/helper/site_config_mariadb.json" ~/frappe-bench/sites/test_site/site_config.json
+    if [ "$db_host" != "127.0.0.1" ]; then
+        sed -i "s/\"db_host\": \"127.0.0.1\"/\"db_host\": \"${db_host}\"/" ~/frappe-bench/sites/test_site/site_config.json
+    fi
 else
     cp -r "${GITHUB_WORKSPACE}/.github/helper/site_config_postgres.json" ~/frappe-bench/sites/test_site/site_config.json
 fi
 
 
 if [ "$DB" == "mariadb" ];then
-    mariadb --host 127.0.0.1 --port 3306 -u root -proot -e "SET GLOBAL character_set_server = 'utf8mb4'"
+    for _ in {1..60}; do
-    mariadb --host 127.0.0.1 --port 3306 -u root -proot -e "SET GLOBAL collation_server = 'utf8mb4_unicode_ci'"
+        if mariadb-admin ping --host "$db_host" --port 3306 -u root -proot --silent; then
+            break
+        fi
+        sleep 1
+    done
+    mariadb-admin ping --host "$db_host" --port 3306 -u root -proot --silent
 
-    # Belt-and-suspenders: also set performance variables at runtime in case
+    mariadb --host "$db_host" --port 3306 -u root -proot -e "SET GLOBAL character_set_server = 'utf8mb4'"
-    # MARIADB_EXTRA_FLAGS was not honoured by the container image.
+    mariadb --host "$db_host" --port 3306 -u root -proot -e "SET GLOBAL collation_server = 'utf8mb4_unicode_ci'"
-    mariadb --host 127.0.0.1 --port 3306 -u root -proot \
+
+    # Throwaway-DB durability tuning at runtime. (innodb_doublewrite is read-only on MariaDB
+    # 10.6, so it can't be disabled here — would need a server startup flag.)
+    mariadb --host "$db_host" --port 3306 -u root -proot \
         -e "SET GLOBAL innodb_flush_log_at_trx_commit=0; SET GLOBAL sync_binlog=0;"
 
-    mariadb --host 127.0.0.1 --port 3306 -u root -proot -e "CREATE USER 'test_frappe'@'localhost' IDENTIFIED BY 'test_frappe'"
+    # Opt-in DDL speedup: a shared tablespace avoids a create+fsync per DocType table during
-    mariadb --host 127.0.0.1 --port 3306 -u root -proot -e "CREATE DATABASE test_frappe"
+    # reinstall — a big win under disk contention. But ROW_FORMAT=DYNAMIC must be accepted in
-    mariadb --host 127.0.0.1 --port 3306 -u root -proot -e "GRANT ALL PRIVILEGES ON \`test_frappe\`.* TO 'test_frappe'@'localhost'"
+    # the system tablespace on this MariaDB. Enable with CI_INNODB_SHARED_TABLESPACE=1; if
+    # reinstall then errors on table creation, unset it (off by default — zero risk).
+    if [ "${CI_INNODB_SHARED_TABLESPACE:-0}" = "1" ]; then
+        mariadb --host "$db_host" --port 3306 -u root -proot -e "SET GLOBAL innodb_file_per_table=0;"
+    fi
 
-    mariadb --host 127.0.0.1 --port 3306 -u root -proot -e "FLUSH PRIVILEGES"
+    mariadb --host "$db_host" --port 3306 -u root -proot -e "CREATE USER 'test_frappe'@'${db_user_host}' IDENTIFIED BY 'test_frappe'"
+    mariadb --host "$db_host" --port 3306 -u root -proot -e "CREATE DATABASE test_frappe"
+    mariadb --host "$db_host" --port 3306 -u root -proot -e "GRANT ALL PRIVILEGES ON \`test_frappe\`.* TO 'test_frappe'@'${db_user_host}'"
+
+    mariadb --host "$db_host" --port 3306 -u root -proot -e "FLUSH PRIVILEGES"
 fi
 
 if [ "$DB" == "postgres" ];then
     echo "travis" | psql -h 127.0.0.1 -p 5432 -c "CREATE DATABASE test_frappe" -U postgres;
     echo "travis" | psql -h 127.0.0.1 -p 5432 -c "CREATE USER test_frappe WITH PASSWORD 'test_frappe'" -U postgres;
-    # CI databases are disposable, so trade durability for speed: postgres fsyncs on every commit
-    # by default, which dominates a commit-heavy test suite. These are all reload-time settings
-    # (no restart needed). MariaDB CI is unaffected (DB != postgres).
-    echo "travis" | psql -h 127.0.0.1 -p 5432 -U postgres \
-        -c "ALTER SYSTEM SET synchronous_commit = 'off'" \
-        -c "ALTER SYSTEM SET fsync = 'off'" \
-        -c "ALTER SYSTEM SET full_page_writes = 'off'" \
-        -c "SELECT pg_reload_conf()";
 fi
 
-
-install_whktml() {
-    # Re-use the .deb if the wkhtmltopdf cache step already restored it.
-    if [ ! -f /tmp/wkhtmltox.deb ]; then
-        wget -O /tmp/wkhtmltox.deb https://github.com/wkhtmltopdf/packaging/releases/download/0.12.6.1-2/wkhtmltox_0.12.6.1-2.jammy_amd64.deb
-    fi
-    sudo apt install /tmp/wkhtmltox.deb
-}
-install_whktml &
-wkpid=$!
-
-
 cd ~/frappe-bench || exit
 
-sed -i 's/watch:/# watch:/g' Procfile
+run_ci_step "Get payments app" bench get-app payments --branch develop
-sed -i 's/schedule:/# schedule:/g' Procfile
-sed -i 's/socketio:/# socketio:/g' Procfile
-sed -i 's/redis_socketio:/# redis_socketio:/g' Procfile
 
-bench get-app payments --branch develop
+# Opt-in: skip building erpnext's frontend assets. Server tests don't need them, but PDF
-bench get-app erpnext "${GITHUB_WORKSPACE}"
+# tests (print formats) do — they pass only if the PDF renderer ignores missing assets.
+# Enable with CI_SKIP_ERPNEXT_ASSETS=1 to test; if PDF tests fail, unset it.
+erpnext_get_app_args=()
+if [ "${CI_SKIP_ERPNEXT_ASSETS:-0}" = "1" ]; then erpnext_get_app_args=(--skip-assets); fi
+run_ci_step "Get erpnext app" bench get-app erpnext "${GITHUB_WORKSPACE}" "${erpnext_get_app_args[@]}"
 
-if [ "$TYPE" == "server" ]; then bench setup requirements --dev; fi
+if [ "$TYPE" == "server" ]; then run_ci_step "Setup dev requirements" bench setup requirements --dev; fi
 
-wait $wkpid
+bench start >> ~/frappe-bench/bench_start.log 2>&1 &
 
-bench start &>> ~/frappe-bench/bench_start.log &
+# Under heavy concurrency, gunicorn's startup can delay redis coming up. reinstall and the
-CI=Yes bench build --app frappe &
+# tests need redis, so wait for it (best-effort, bounded) instead of racing — contention
-bench --site test_site reinstall --yes
+# then slows the job rather than failing it.
+wait_for_redis() {
+    local cfg=~/frappe-bench/sites/common_site_config.json
+    [ -f "$cfg" ] || return 0
+    local ports port
+    ports=$(python - "$cfg" <<'PY'
+import json, re, sys
+try:
+    cfg = json.load(open(sys.argv[1]))
+except Exception:
+    sys.exit(0)
+for key in ("redis_cache", "redis_queue"):
+    match = re.search(r":(\d+)", str(cfg.get(key, "")))
+    if match:
+        print(match.group(1))
+PY
+)
+    for port in $ports; do
+        for _ in $(seq 1 120); do
+            if (exec 3<>"/dev/tcp/127.0.0.1/$port") 2>/dev/null; then
+                exec 3>&- 3<&-
+                break
+            fi
+            sleep 1
+        done
+    done
+}
+wait_for_redis
+
+# Site setup. `bench reinstall` rebuilds the entire schema in Python (~1000 DocTypes) — the
+# CI bottleneck that DB tuning / tmpfs / faster cores couldn't move. Instead, restore a
+# pre-baked baseline (the DB engine loads it, no Python schema-build) and `migrate` to sync
+# only the drift since the baseline was built. The baseline is produced by
+# .github/helper/generate-ci-baseline.sh (run nightly / at image build) from a clean
+# reinstall on develop. Gated by CI_RESTORE_FROM_BACKUP so it A/Bs against plain reinstall;
+# falls back to reinstall if the baseline is missing or the restore fails.
+CI_BASELINE_BACKUP="${CI_BASELINE_BACKUP:-/opt/ci-baseline/test_site-database.sql.gz}"
+if [ "${CI_RESTORE_FROM_BACKUP:-0}" = "1" ] && [ -f "$CI_BASELINE_BACKUP" ]; then
+    if [ "$DB" == "mariadb" ]; then
+        db_root_args=(--db-root-username root --db-root-password root)
+    else
+        db_root_args=(--db-root-username postgres --db-root-password travis)
+    fi
+    if run_ci_step "Restore baseline test site" bench --site test_site --force restore "${db_root_args[@]}" "$CI_BASELINE_BACKUP"; then
+        run_ci_step "Migrate test site" bench --site test_site migrate
+    else
+        run_ci_step "Reinstall test site (baseline restore failed)" bench --site test_site reinstall --yes
+    fi
+else
+    run_ci_step "Reinstall test site" bench --site test_site reinstall --yes
+fi
+
+# Refresh the baseline backup from this freshly set-up site. Run a normal job (reinstall path)
+# with CI_GENERATE_BASELINE=1 and the baseline dir mounted read-write; install.sh captures the
+# DB dump to CI_BASELINE_BACKUP so future runs can restore it. Nightly is enough — bench migrate
+# absorbs intraday develop drift. To bake into the image instead, copy the produced .sql.gz in.
+if [ "${CI_GENERATE_BASELINE:-0}" = "1" ]; then
+    run_ci_step "Backup baseline test site" bench --site test_site backup
+    latest_backup=$(ls -t ~/frappe-bench/sites/test_site/private/backups/*-database.sql.gz | head -1)
+    mkdir -p "$(dirname "$CI_BASELINE_BACKUP")"
+    cp "$latest_backup" "$CI_BASELINE_BACKUP"
+    echo "Baseline written to $CI_BASELINE_BACKUP ($(du -h "$latest_backup" | cut -f1))"
+fi

.github/helper/start-db.sh

@@ -0,0 +1,75 @@
+#!/bin/bash
+#
+# Run MariaDB INSIDE the runner container, on a datadir we control. Because the datadir can be
+# packaged into the bench artifact, test shards start an already-loaded server instead of
+# replaying a SQL dump (the ~60s hydrate restore). Each shard gets its own copy → isolation kept.
+#
+# CI_DB_DATADIR picks the path:
+#   - setup job:  /home/ci/db-data  (OUTSIDE the bench, so install.sh's `rm -rf ~/frappe-bench`
+#                 doesn't wipe it; it's moved into the bench just before packaging)
+#   - test shard: ~/frappe-bench/mariadb-data  (where the artifact untar'd it)
+#
+# Idempotent: inits a fresh datadir if absent (setup), else starts on the existing one (shards).
+#
+set -e
+
+ci_user="${ERPNEXT_CI_USER:-frappe}"
+
+# Re-exec as the ci user so mariadbd and the datadir are owned consistently (root mariadbd is
+# refused anyway). Mirrors install.sh's user switch.
+if [ "$(id -u)" = "0" ] && [ "${SKIP_SYSTEM_SETUP:-0}" = "1" ] && [ "$ci_user" != "root" ]; then
+    exec su -m "$ci_user" -s /bin/bash -c \
+        "ERPNEXT_CI_USER='$ci_user' CI_DB_DATADIR='${CI_DB_DATADIR:-}' bash '$0'"
+fi
+
+# --- PostgreSQL (GitHub-hosted CI): run in-runner on a PGDATA so it bakes into the artifact,
+# same idea as the mariadb datadir. Trust auth (throwaway CI) skips password setup; durability
+# off for speed. Postgres is preinstalled on ubuntu-latest under /usr/lib/postgresql/<ver>/bin.
+if [ "${DB:-mariadb}" = "postgres" ]; then
+    PG_BIN=$(ls -d /usr/lib/postgresql/*/bin 2>/dev/null | sort -V | tail -1)
+    [ -n "$PG_BIN" ] && export PATH="$PG_BIN:$PATH"
+    PGDATA="${CI_DB_DATADIR:-$HOME/frappe-bench/pgdata}"
+    if [ ! -d "$PGDATA/base" ]; then
+        initdb -D "$PGDATA" -U postgres --auth-local=trust --auth-host=trust >/dev/null
+        echo "host all all 127.0.0.1/32 trust" >> "$PGDATA/pg_hba.conf"
+    fi
+    pg_ctl -D "$PGDATA" -w -o "-p 5432 -c listen_addresses=127.0.0.1 -c unix_socket_directories=$PGDATA -c fsync=off -c synchronous_commit=off -c full_page_writes=off" start
+    echo "PostgreSQL up in-runner (pgdata=$PGDATA)"
+    exit 0
+fi
+
+# --- MariaDB ---
+DATADIR="${CI_DB_DATADIR:-$HOME/frappe-bench/mariadb-data}"
+SOCK="$DATADIR/mysqld.sock"
+fresh=0
+
+if [ ! -d "$DATADIR/mysql" ]; then
+    mkdir -p "$DATADIR"
+    mariadb-install-db --no-defaults --datadir="$DATADIR" \
+        --auth-root-authentication-method=normal --skip-test-db >/dev/null 2>&1
+    fresh=1
+fi
+
+# Throwaway-CI durability off; bind TCP 127.0.0.1:3306 so bench/install.sh connect as usual.
+mariadbd --no-defaults --datadir="$DATADIR" --socket="$SOCK" --pid-file="$DATADIR/mysqld.pid" \
+    --port=3306 --bind-address=127.0.0.1 \
+    --innodb-flush-log-at-trx-commit=0 --sync-binlog=0 --skip-log-bin \
+    > "$HOME/mariadb.log" 2>&1 &
+
+for _ in $(seq 1 60); do
+    mariadb-admin --socket="$SOCK" ping --silent 2>/dev/null && break
+    sleep 1
+done
+
+if [ "$fresh" = "1" ]; then
+    # A fresh datadir has only a password-less root@localhost. Give it the password install.sh
+    # uses, plus a TCP-reachable root@127.0.0.1, so the rest of install.sh works unchanged.
+    mariadb --no-defaults --socket="$SOCK" -u root <<'SQL'
+ALTER USER 'root'@'localhost' IDENTIFIED BY 'root';
+CREATE USER IF NOT EXISTS 'root'@'127.0.0.1' IDENTIFIED BY 'root';
+GRANT ALL PRIVILEGES ON *.* TO 'root'@'127.0.0.1' WITH GRANT OPTION;
+FLUSH PRIVILEGES;
+SQL
+fi
+
+echo "MariaDB up in-container (datadir=$DATADIR, fresh=$fresh)"

.github/workflows/server-tests-postgres.yml

@@ -1,79 +1,43 @@
 name: Server (Postgres)
 
 on:
-  repository_dispatch:
+  schedule:
-    types: [frappe-framework-change]
+    # 03:00 AM IST daily (21:30 UTC the previous day)
+    - cron: "30 21 * * *"
   pull_request:
-    # 'labeled' is required so adding the 'postgres' label to an open PR triggers this run
-    # (the job itself is gated on that label below)
-    types: [opened, reopened, synchronize, labeled]
     paths-ignore:
       - '**.js'
-      - '**.css'
-      - '**.svg'
       - '**.md'
       - '**.html'
       - 'crowdin.yml'
       - '.coderabbit.yml'
       - '.mergify.yml'
-  schedule:
-    # Run everday at midnight UTC / 5:30 IST
-    - cron: "0 0 * * *"
   workflow_dispatch:
-    inputs:
-      user:
-        description: 'Frappe Framework repository user (add your username for forks)'
-        required: true
-        default: 'frappe'
-        type: string
-      branch:
-        description: 'Frappe Framework branch'
-        default: 'develop'
-        required: false
-        type: string
-
-permissions:
-  contents: read
 
 concurrency:
   group: server-postgres-develop-${{ github.event_name }}-${{ github.event.number || github.event_name == 'workflow_dispatch' && github.run_id || '' }}
   cancel-in-progress: true
 
+permissions:
+  contents: read
+
+# Postgres CI stays on GitHub-hosted (free, full-speed VM per shard) but follows the same fan-out
+# we built for MariaDB: build the bench + reinstall ONCE in the setup job, bake the PostgreSQL
+# PGDATA into the artifact, and have 4 test shards start Postgres on that datadir — no per-shard
+# clone/build/reinstall/restore. Python is pinned so the venv transplants between VMs.
+env:
+  TZ: 'Asia/Kolkata'
+  NODE_ENV: "production"
+  PYTHON_VERSION: '3.14'
+
 jobs:
-  test:
+  setup:
-    # Opt-in on PRs: only runs when the PR carries the 'postgres' label. Scheduled / manual /
+    name: Build & reinstall (setup)
-    # framework-dispatch runs always execute (no PR labels to gate on).
-    if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'postgres') }}
     runs-on: ubuntu-latest
-    timeout-minutes: 60
+    # Runs on the daily schedule (and workflow_dispatch). On PRs it runs ONLY when the PR carries
-    env:
+    # the 'postgres' label — the test job needs setup, so it's skipped too when this is.
-      TZ: 'Asia/Kolkata'
+    if: github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'postgres')
-      NODE_ENV: "production"
+    timeout-minutes: 40
-      WITH_COVERAGE: ${{ github.event_name != 'pull_request' }}
-
-    strategy:
-      fail-fast: false
-
-      matrix:
-        container: [1, 2, 3, 4]
-
-    # Distinct from the MariaDB job's "Python Unit Tests" so its check contexts do NOT collide with
-    # the required "Python Unit Tests (1..4)" status checks -- this keeps Postgres non-required for now.
-    name: Postgres Unit Tests
-
-    services:
-      postgres:
-        image: postgres:13.3
-        env:
-          POSTGRES_PASSWORD: travis
-        ports:
-          - 5432:5432
-        options: >-
-          --health-cmd pg_isready
-          --health-interval 10s
-          --health-timeout 5s
-          --health-retries 5
-
     steps:
       - name: Clone
         uses: actions/checkout@v6
@@ -81,7 +45,7 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.14'
+          python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Check for valid Python & Merge Conflicts
         run: |
@@ -100,98 +64,128 @@ jobs:
       - name: Add to Hosts
         run: echo "127.0.0.1 test_site" | sudo tee -a /etc/hosts
 
-      - name: Cache pip
+      - name: Cache deps (uv/pip/npm/yarn)
         uses: actions/cache@v4
         with:
-          path: ~/.cache/pip
+          path: |
-          key: ${{ runner.os }}-pip-${{ hashFiles('**/*requirements.txt', '**/pyproject.toml') }}
+            ~/.cache/uv
-          restore-keys: |
+            ~/.cache/pip
-            ${{ runner.os }}-pip-
+            ~/.npm
-            ${{ runner.os }}-
+            ~/.cache/yarn
+          key: ${{ runner.os }}-deps-${{ hashFiles('**/*requirements.txt', '**/pyproject.toml', '**/yarn.lock') }}
+          restore-keys: ${{ runner.os }}-deps-
 
-      - name: Cache node modules
+      # Warm-bench cache (the big one): install.sh saves the built base bench — frappe + env +
+      # node_modules + assets — here as frappe-bench-base-*.tar.zst. Later runs restore it and only
+      # fast-forward to the live develop SHA + rebuild the delta, so the bench BUILD is near-free and
+      # only the test_site reinstall (per-run DB, uncacheable) stays slow — matching the self-hosted
+      # box. The first run after a deps change populates it; every run after that is fast.
+      - name: Cache warm bench (base build)
         uses: actions/cache@v4
+        with:
+          path: ~/bench-cache
+          key: ${{ runner.os }}-warmbench-v2-${{ hashFiles('**/*requirements.txt', '**/pyproject.toml', '**/yarn.lock') }}
+          restore-keys: ${{ runner.os }}-warmbench-v2-
+
+      # Postgres runs in-runner on a PGDATA OUTSIDE the bench (install.sh wipes ~/frappe-bench);
+      # after the reinstall it's moved into the bench so it ships in the artifact.
+      - name: Start DB
+        run: bash ${GITHUB_WORKSPACE}/.github/helper/start-db.sh
         env:
-          cache-name: cache-node-modules
+          DB: postgres
-        with:
+          CI_DB_DATADIR: /home/runner/pgdata
-          path: ~/.npm
-          key: ${{ runner.os }}-build-${{ env.cache-name }}-${{ hashFiles('**/package-lock.json') }}
-          restore-keys: |
-            ${{ runner.os }}-build-${{ env.cache-name }}-
-            ${{ runner.os }}-build-
-            ${{ runner.os }}-
-
-      - name: Get yarn cache directory path
-        id: yarn-cache-dir-path
-        run: echo "dir=$(yarn cache dir)" >> $GITHUB_OUTPUT
-
-      - uses: actions/cache@v4
-        id: yarn-cache
-        with:
-          path: ${{ steps.yarn-cache-dir-path.outputs.dir }}
-          key: ${{ runner.os }}-yarn-${{ hashFiles('**/yarn.lock') }}
-          restore-keys: |
-            ${{ runner.os }}-yarn-
-
-      - name: Cache wkhtmltopdf
-        uses: actions/cache@v4
-        with:
-          path: /tmp/wkhtmltox.deb
-          key: wkhtmltox-0.12.6.1-2-jammy-amd64
 
       - name: Install
         run: bash ${GITHUB_WORKSPACE}/.github/helper/install.sh
         env:
           DB: postgres
           TYPE: server
-          FRAPPE_USER: ${{ github.event.inputs.user }}
+          FRAPPE_BRANCH: develop
-          FRAPPE_BRANCH: ${{ github.event.client_payload.sha || github.event.inputs.branch }}
+          BENCH_CACHE_DIR: /home/runner/bench-cache
+
+      - name: Stop DB and stage datadir
+        run: |
+          PG_BIN=$(ls -d /usr/lib/postgresql/*/bin | sort -V | tail -1)
+          "$PG_BIN/pg_ctl" -D /home/runner/pgdata -m fast -w stop || true
+          mv /home/runner/pgdata /home/runner/frappe-bench/pgdata
+
+      - name: Package bench for test shards
+        run: |
+          cp "${GITHUB_WORKSPACE}/.github/helper/hydrate.sh" /home/runner/frappe-bench/hydrate.sh
+          cp "${GITHUB_WORKSPACE}/.github/helper/start-db.sh" /home/runner/frappe-bench/start-db.sh
+          tar czpf "${GITHUB_WORKSPACE}/bench.tar.gz" -C /home/runner \
+            --exclude='.git' --exclude='node_modules' frappe-bench
+          ls -lh "${GITHUB_WORKSPACE}/bench.tar.gz"
+
+      - name: Upload bench artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: bench-pg
+          path: bench.tar.gz
+          retention-days: 1
+          compression-level: 0
+
+  test:
+    name: Python Unit Tests (PG)
+    needs: setup
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        container: [1, 2, 3, 4]
+    steps:
+      - name: Download bench artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: bench-pg
+
+      - name: Setup Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+
+      - name: Add to Hosts
+        run: echo "127.0.0.1 test_site" | sudo tee -a /etc/hosts
+
+      # The bench CLI (frappe-bench) and redis are global/system tools — not in the bench tarball.
+      # The setup runner got them via install.sh; the MariaDB shards get them from the arc5 image.
+      # GitHub-hosted PG shards install them here (cheap vs the build+reinstall that setup did once).
+      - name: Install shard runtime (bench CLI + redis + wkhtmltopdf)
+        run: |
+          pip install frappe-bench
+          command -v redis-server >/dev/null || { sudo apt-get update -qq && sudo apt-get install -y -qq redis-server; }
+          # wkhtmltopdf (patched-qt build) for print-format / PDF tests — same .deb install.sh uses.
+          if ! command -v wkhtmltopdf >/dev/null; then
+            wget -qO /tmp/wkhtmltox.deb https://github.com/wkhtmltopdf/packaging/releases/download/0.12.6.1-2/wkhtmltox_0.12.6.1-2.jammy_amd64.deb
+            sudo apt-get install -y -qq /tmp/wkhtmltox.deb
+          fi
+
+      - name: Untar bench
+        run: |
+          tar xzpf "${GITHUB_WORKSPACE}/bench.tar.gz" -C /home/runner
+          ls -ld /home/runner/frappe-bench
+
+      - name: Hydrate (start Postgres on the baked datadir)
+        run: bash /home/runner/frappe-bench/hydrate.sh
+        env:
+          DB: postgres
+          DB_HOST: 127.0.0.1
 
       - name: Run Tests
         run: |
           cd ~/frappe-bench/
-          coverage_flag=""
+          # print-format / PDF tests are engine-independent (they exercise wkhtmltopdf rendering,
-          if [ "$WITH_COVERAGE" = "true" ]; then coverage_flag="--with-coverage"; fi
+          # not postgres SQL — the MariaDB CI already covers them). They only fetch the static asset
+          # bundles from http://test_site:8000/assets/..., so a plain static file server over sites/
+          # satisfies wkhtmltopdf without the frappe web server (which never bound on a bare runner).
+          ( cd ~/frappe-bench/sites && nohup python3 -m http.server 8000 --bind 127.0.0.1 > ~/frappe-bench/web.log 2>&1 & )
+          for _ in $(seq 1 15); do (exec 3<>/dev/tcp/127.0.0.1/8000) 2>/dev/null && { exec 3>&- 3<&-; break; }; sleep 1; done
           bench --site test_site run-parallel-tests --lightmode --app erpnext \
-            --total-builds ${{ strategy.job-total }} \
+            --total-builds 4 --build-number ${{ matrix.container }}
-            --build-number ${{ matrix.container }} \
-            $coverage_flag
         env:
           TYPE: server
 
-
+      - name: Show web server log
-      - name: Show bench output
         if: ${{ always() }}
-        run: cat ~/frappe-bench/bench_start.log || true
+        run: cat ~/frappe-bench/web.log 2>/dev/null || true
-
-      - name: Upload coverage data
-        if: ${{ env.WITH_COVERAGE == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          name: coverage-postgres-${{ matrix.container }}
-          path: /home/runner/frappe-bench/sites/coverage.xml
-
-  coverage:
-    name: Coverage Wrap Up
-    needs: test
-    if: ${{ github.event_name != 'pull_request' }}
-    runs-on: ubuntu-latest
-    steps:
-      - name: Clone
-        uses: actions/checkout@v6
-
-      - name: Download artifacts
-        uses: actions/download-artifact@v4
-        with:
-          pattern: coverage-postgres-*
-
-      - name: Upload coverage data
-        uses: codecov/codecov-action@v4
-        with:
-          name: Postgres
-          flags: postgres
-          # explicit glob: download-artifact extracts each shard into its own coverage-postgres-N/ dir
-          files: coverage-postgres-*/coverage.xml
-          token: ${{ secrets.CODECOV_TOKEN }}
-          fail_ci_if_error: true
-          verbose: true