scripts/stage-rc-from-prod.sh

Rama master · réplica de lectura del repositorio original.

#!/usr/bin/env bash
# Clone the prod Sublarr Postgres DB into the RC Postgres. DESTRUCTIVE on the
# target. Guarded so it can only ever hit sublarr-rc-postgres, and only with
# FORCE=1. See docs/superpowers/specs/2026-07-09-sublarr-3tier-design.md.
set -euo pipefail

# NB: rely on ssh's default key discovery (~/.ssh/id_ed25519) rather than an
# explicit -i path — a HOME with spaces (Windows git-bash "Dennis Wittke")
# word-splits inside the unquoted $SSH expansion and breaks the -i argument.
SSH="ssh -o ConnectTimeout=10 root@192.168.178.36"
SOURCE_PG="sublarr-postgres"
TARGET_PG="${TARGET_PG:-sublarr-rc-postgres}"
DB="sublarr"
DBUSER="sublarr"
RC_COMPOSE_DIR="/mnt/user/appdata/sublarr-rc"
RC_PROJECT="sublarr-rc"
RC_APP="sublarr-rc"
RC_PORT="5766"  # RC's port on the Cardinal host (container itself listens on 5765)
# Config-at-rest encryption key lives in {config_dir}/.encryption_key, deliberately
# OUTSIDE the DB — so a DB clone alone can't decrypt prod's encrypted config values
# and the RC app crashes on boot ("Failed to decrypt"). Copy prod's key too.
PROD_KEY="/mnt/user/appdata/sublarr/config/.encryption_key"
RC_KEY="/mnt/user/appdata/sublarr-rc/config/.encryption_key"

# --- Guards (the critical safety logic; run BEFORE any docker/ssh) ---
if [[ "${FORCE:-}" != "1" ]]; then
  echo "REFUSE: set FORCE=1 to run the destructive prod->RC clone" >&2
  exit 2
fi
case "$TARGET_PG" in
  sublarr-postgres | sublarr-beta-postgres)
    echo "REFUSE: target '$TARGET_PG' is prod/beta — never clone onto it" >&2
    exit 3
    ;;
  sublarr-rc-postgres) : ;; # the only allowed target
  *)
    echo "REFUSE: unknown target '$TARGET_PG' (allowed: sublarr-rc-postgres)" >&2
    exit 4
    ;;
esac
if [[ "$TARGET_PG" == "$SOURCE_PG" ]]; then
  echo "REFUSE: target equals source" >&2
  exit 5
fi

# --- Clone: stop RC app (release connections) -> drop+recreate -> dump|restore -> start ---
echo "Cloning $SOURCE_PG -> $TARGET_PG ..."
$SSH "cd $RC_COMPOSE_DIR && docker compose -p $RC_PROJECT stop $RC_APP"
$SSH "docker exec $TARGET_PG psql -U $DBUSER -d postgres -c 'DROP DATABASE IF EXISTS $DB;' -c 'CREATE DATABASE $DB;'"
# ON_ERROR_STOP is the whole point of this line. Without it psql walks past a
# failed statement and still exits 0 -- and because a pg_dump puts the data
# first and the ALTER TABLE ... ADD CONSTRAINT, CREATE INDEX and setval calls
# LAST, "mostly worked" means a clone with rows but no primary keys and
# sequences left at 1. That is what RC had been running on, undetected, until
# 2026-09-09: 73 tables without a primary key against prod's 8, the mappings
# sequence at 42 while rows reached 1171, and duplicate ids as a result.
# --single-transaction makes a failure leave nothing behind to puzzle over.
if ! $SSH "set -o pipefail; docker exec $SOURCE_PG pg_dump -U $DBUSER -d $DB --no-owner --no-privileges | docker exec -i $TARGET_PG psql -U $DBUSER -d $DB -v ON_ERROR_STOP=1 --single-transaction -q"; then
  echo "REFUSE: the restore failed — RC is NOT a copy of prod, leaving it stopped" >&2
  exit 6
fi
# --- Prove the copy carries the schema, not just the rows ------------------
# A clone is only a mirror if what constrains the data came with it. Compare
# the two databases on the things a silent restore drops first.
pkless() { $SSH "docker exec $1 psql -U $DBUSER -d $DB -Atc \"SELECT COUNT(*) FROM information_schema.tables t WHERE t.table_schema='public' AND NOT EXISTS (SELECT 1 FROM information_schema.table_constraints c WHERE c.table_name=t.table_name AND c.constraint_type='PRIMARY KEY');\"" | tr -d '
'; }
SRC_PKLESS="$(pkless "$SOURCE_PG")"
DST_PKLESS="$(pkless "$TARGET_PG")"
echo "tables without a primary key: source=$SRC_PKLESS target=$DST_PKLESS"
if [[ "$SRC_PKLESS" != "$DST_PKLESS" ]]; then
  echo "REFUSE: the target lost constraints in the restore ($DST_PKLESS vs $SRC_PKLESS)" >&2
  echo "        Testing against this clone proves nothing — fix the restore first." >&2
  exit 7
fi

# --- Do not let the clone inherit prod's identity in the telemetry ---------
# config_entries carries usage_stats_install_id, so a straight dump gives RC
# prod's id. Both instances then ping stats.sublarr.de under it and overwrite
# each other — measured 2026-09-28: prod pinged 17:03:37, RC 17:05:38, and the
# single "postgres" install in the public aggregate was therefore describing
# the idle mirror rather than the live instance. Blanking it makes RC mint its
# own id on first ping. Consent stays as it is; the id is the only thing here
# that identifies a machine rather than a decision.
$SSH "docker exec $TARGET_PG psql -U $DBUSER -d $DB -q -c \"UPDATE config_entries SET value='' WHERE key='usage_stats_install_id';\"" \
  || echo "WARN: could not blank the cloned install id — RC will report as prod" >&2

# --- Keep the first boot quiet -------------------------------------------
# The config PUT below only lands once RC is up, but the boot itself reads
# prod's wanted_scan_on_startup / wanted_search_on_startup and the live
# intervals. Observed 2026-10-26: RC started a wanted search over prod's
# backlog before the jobs were paused. Writing the quiet values into the
# clone before the start closes that window; the PUT below stays as the
# belt-and-braces path through the app.
if [[ "${TRANSLATE:-}" == "1" ]]; then BOOT_AUTOMATION=True; else BOOT_AUTOMATION=False; fi
BOOT_SQL="INSERT INTO config_entries (key, value, updated_at) VALUES
  ('wanted_scan_on_startup','False',now()), ('wanted_search_on_startup','False',now()),
  ('wanted_search_interval_hours','0',now()), ('upgrade_scan_interval_hours','0',now()),
  ('wanted_scan_interval_hours','0',now()), ('wanted_auto_translate','False',now()),
  ('subtitle_automation_enabled','$BOOT_AUTOMATION',now())
  ON CONFLICT (key) DO UPDATE SET value=EXCLUDED.value, updated_at=EXCLUDED.updated_at;"
$SSH "docker exec $TARGET_PG psql -U $DBUSER -d $DB -q -v ON_ERROR_STOP=1 -c \"$BOOT_SQL\"" || {
  echo "REFUSE: could not write the quiet boot settings — leaving RC stopped" >&2
  exit 8
}

$SSH "cp -a '$PROD_KEY' '$RC_KEY' && chmod 600 '$RC_KEY'"
$SSH "cd $RC_COMPOSE_DIR && docker compose -p $RC_PROJECT up -d $RC_APP"

# --- Silence the automation the clone just armed -------------------------
# The dump carries config_entries, and the encryption key copied above makes
# them decryptable — so RC comes up holding prod's *real* DeepL/OpenSubtitles
# credentials, pointed at prod's full backlog (thousands of wanted items).
# subtitle_automation then re-translates that backlog on RC, billing the
# owner's live DeepL quota for work prod already paid for. Observed
# 2026-08-01: 1468 DeepL calls in six hours from a single staging round.
#
# Pause every job that can reach a paid API. Everything else (cleanup,
# stats, health sweeps) is local and stays running so the RC still behaves
# like a real instance. Unpause individually in Settings → System →
# Scheduler when a round genuinely needs to exercise translation.
# NB: 5766 is RC's port *on the host* — the container listens on 5765, but
# these curls run on Cardinal, not inside it.
echo "Waiting for RC to accept requests ..."
$SSH "for i in \$(seq 1 60); do curl -sf --max-time 3 http://localhost:${RC_PORT}/api/v1/health >/dev/null 2>&1 && exit 0; sleep 2; done; exit 1" || {
  echo "REFUSE: RC did not become healthy — pause the paid-API jobs by hand" >&2
  exit 6
}

# A pause alone does NOT hold. _apply_intervals_to_apscheduler runs on every
# settings-save and on every boot, and re-arms any job whose configured
# interval is greater than zero — so the next config write, or the next
# restart, quietly puts the paused job back on the clock with prod's live
# credentials. Setting the interval to 0 is what sticks: the boot path pauses
# on interval <= 0.
#
# wanted_scan_interval_hours and wanted_auto_translate joined this list on
# 2026-08-29, when prod switched both on. The clone carries them, and
# wanted_scanner was in nobody's pause list.
# TRANSLATE=1 leaves the translation pipeline running. It is off by default
# because a clone re-translating prod's backlog is duplicated work, but the
# original reason -- billing the owner's live DeepL quota -- no longer
# applies: since 2026-08-04 both prod and RC run translation_default_backend
# = ollama with an empty fallback chain, so no paid backend is reachable at
# all. What it still costs is the Mac mini, which prod shares.
#
# It exists because an RC that translates nothing cannot fail at translating,
# and the promote gate reads "three days without ERNST". On 2026-09-11 RC had
# been quiet for 29 h with ev_ok_24h=0 and ev_err_24h=0 -- silence that said
# nothing about the build.
echo "Disabling the load- and quota-generating settings on RC ..."
if [[ "${TRANSLATE:-}" == "1" ]]; then
  echo "  TRANSLATE=1 -- leaving subtitle_automation ON (ollama, no paid backend)"
  AUTOMATION_JSON='"subtitle_automation_enabled":true'
  PAUSE_JOBS="wanted_search wanted_scanner mt_reseek upgrade_scan"
else
  AUTOMATION_JSON='"subtitle_automation_enabled":false'
  PAUSE_JOBS="subtitle_automation wanted_search wanted_scanner mt_reseek upgrade_scan"
fi
CFG_JSON="{\"wanted_search_interval_hours\":0,\"upgrade_scan_interval_hours\":0,\"wanted_scan_interval_hours\":0,\"wanted_auto_translate\":false,${AUTOMATION_JSON}}"
$SSH "docker exec $RC_APP sh -lc 'curl -sS -m 30 -o /dev/null -w \"  config: %{http_code}\n\" -X PUT -H \"X-Api-Key: \$SUBLARR_API_KEY\" -H \"Content-Type: application/json\" -d '\''$CFG_JSON'\'' http://localhost:5765/api/v1/config'" || \
  echo "  config write failed — set the intervals to 0 by hand" >&2

# Belt and braces: the pause bites immediately, the interval keeps it off
# across restarts. mt_reseek has no interval setting, so there the pause is
# all there is.
echo "Pausing paid-API scheduler jobs on RC ..."
for _job in $PAUSE_JOBS; do
  $SSH "curl -sS -X POST 'http://localhost:${RC_PORT}/api/v1/scheduler/jobs/${_job}/pause' -o /dev/null -w '  ${_job}: %{http_code}\n'" || \
    echo "  ${_job}: pause failed (non-fatal — pause it by hand)" >&2
done

if [[ "${TRANSLATE:-}" == "1" ]]; then
  echo "Done. RC mirrors prod DB; provider jobs off, translation ON."
else
  echo "Done. RC mirrors prod DB (config_entries included); paid-API jobs off."
fi