erplibre/script/odoo/migration/reset_stale_cow_views.py
Mathieu Benoit e214df7329 [FIX] smoke: judge the back office AFTER the repair, and say why it failed
No timeout was hit and no test was skipped. The back office failed for the
same reason the site did: the COW copy of website.submenu breaks
/web/login exactly as it breaks the public pages — both go through the
same layout. The pass simply ran BEFORE the reset, and unlike the URLs it
was never looked at again. It is now, on the same second server.

« Session expired » described the consequence and hid the cause. The login
page returned 500 and nothing said so, so one looks at the password. The
status of that page, and a missing CSRF token, are now named.

Diagnosing this, I started Odoo 18 on a 17 database and got 500 on all
thirty-seven URLs and on /web/login: a report that says the site is
entirely broken when nothing is. database_cleanup already refused that;
the smoke tools now share the same guard rather than reimplementing it.

--- FR ---

[FIX] smoke : juger le back-office APRÈS la réparation, et dire pourquoi

Aucun délai n'a été atteint et aucun test n'a été sauté. Le back-office a
échoué pour la raison même qui cassait le site : la copie COW de
website.submenu casse /web/login comme elle casse les pages publiques —
les deux passent par le même gabarit. La passe tournait avant la
réinitialisation et, contrairement aux URL, n'était jamais revue.

« Session expired » décrivait la conséquence et cachait la cause : la page
de connexion rendait 500 sans que rien ne le dise, et l'on cherchait du
côté du mot de passe.

En diagnostiquant, j'ai lancé un Odoo 18 sur une base 17 : 500 partout, un
rapport qui déclare le site entièrement cassé quand rien ne l'est.
database_cleanup refusait déjà cela ; les outils de fumée partagent
désormais sa garde au lieu d'en écrire une seconde.

Assisted-by: Claude Opus 5
2026-08-22 07:23:59 -04:00

517 lines
19 KiB
Python
Executable file
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#!/usr/bin/env python3
# © 2021-2026 TechnoLibre (http://www.technolibre.ca)
# License AGPL-3.0 or later (http://www.gnu.org/licenses/agpl)
"""Find the website COW copies that have gone stale, and reset them.
The problem
-----------
A COW copy freezes the module view it was copied from. Version after version,
the MODULE view is modernised while the copy is not — it is user data, nothing
rewrites it. The copy keeps working right up to the moment a CHILD view does an
``xpath`` onto something the module added, renamed or re-anchored::
Element '<xpath expr="//head/script[@id='web.layout.odooscript']">'
cannot be located in parent view
Observed on ``web.layout`` during 14.0 -> 15.0: the copy dated from 12.0, where
the script tag had no ``id`` and QWeb still said ``t-raw``. It had survived
until then only because the 14.0 module xpath carried a fallback
(``//head/script[@id='…'] | //head/script[last()]``) that 15.0 removed.
What this checks
----------------
Not a heuristic on deprecated syntax, and not an absolute one either: the
DRIFT between a copy and its module twin. For every active child of a COW
copy, each ``xpath`` expression is resolved twice — against the module twin
and against the copy. An expression the twin can satisfy and the copy cannot
is an anchor the copy has lost.
The comparison has to be differential. Odoo resolves an xpath against the
COMBINED arch of the whole inheritance chain, so « does not resolve in this
parent » proves nothing on its own — a child of ``website.layout`` targets
``//header``, which comes from an ancestor.
When it bites
-------------
A reported copy is not necessarily failing right now. Odoo re-validates a COW
copy when its module twin is REWRITTEN — which is what a version bump does —
or when the page is rendered for that website. So a finding is a breakage
already present in the data, waiting for the next bump to surface it. That is
precisely when it is cheap to fix.
Resetting
---------
``--reset`` copies the MODULE arch over the stale copy. The previous arch is
always written to a backup file first, and the diff is printed, because a copy
can hold a genuine customisation — one line among fifty of drift. Read the
diff, reset, then re-apply what mattered as an INHERITING view rather than a
full copy, so the next version bump cannot make it stale again.
Plain psql on purpose: this runs on databases whose Odoo registry does not
load, which is precisely when it is needed.
"""
import argparse
import datetime
import difflib
import json
import os
import re
import subprocess
import sys
sys.path.append(
os.path.normpath(os.path.join(os.path.dirname(__file__), "..", "..", ".."))
)
try:
from script.todo.todo_i18n import t
except Exception: # pragma: no cover - repli si i18n indisponible
def t(key: str) -> str:
return key
def run_psql(database, sql):
"""Run a statement and return stdout, raising on failure."""
result = subprocess.run(
["psql", "-d", database, "-tAc", sql],
capture_output=True,
text=True,
)
if result.returncode:
raise RuntimeError(
f"Query failed on '{database}': {result.stderr.strip()}"
)
return result.stdout
def normalise_arch(value):
"""The arch as a string, whatever the column type.
Odoo stores ``arch_db`` as text up to 15.0 and as jsonb (one entry per
language) from 16.0. Casting to text gives ``{"en_US": "<t …"}`` in the
second case, so unwrap it — the structure is the same in every language,
and the structure is all this tool looks at.
"""
if not isinstance(value, str):
return "" if value is None else str(value)
text = value.strip()
if text.startswith("{") and '"' in text:
try:
data = json.loads(text)
except ValueError:
return value
if isinstance(data, dict) and data:
for lang in ("en_US", *sorted(data)):
if lang in data and isinstance(data[lang], str):
return data[lang]
return value
def fetch_views(database):
"""Every view that matters: COW copies, their module twin, their children.
One query rather than one per view — these databases hold thousands of
views. The result comes back as JSON: an arch is multi-line XML, so any
line-based format would need a record separator the data could forge.
"""
raw = run_psql(
database,
"SELECT COALESCE(json_agg(json_build_object("
"'id', id, 'key', COALESCE(key, ''), 'inherit_id', inherit_id,"
"'website_id', website_id, 'active', active,"
"'arch', arch_db::text))::text, '[]')"
" FROM ir_ui_view WHERE arch_db IS NOT NULL",
)
rows = {}
for item in json.loads(raw.strip() or "[]"):
item["arch"] = normalise_arch(item.get("arch"))
rows[item["id"]] = item
return rows
XPATH_RE = re.compile(r"<xpath\b[^>]*\bexpr=([\"'])(.*?)\1", re.S)
def child_xpaths(arch):
"""The xpath expressions a view applies to its parent."""
return [match.group(2) for match in XPATH_RE.finditer(arch)]
def require_lxml():
"""The XPath engine, or a loud failure.
Never degrade to « everything resolves » when lxml is missing: a checker
that cannot check must not answer « all clean ». Odoo's own venvs all ship
lxml; the bare system python3 usually does not.
"""
try:
from lxml import etree
return etree
except ImportError:
sys.exit(
f"❌ {t('lxml is required to resolve the xpath expressions.')}\n"
" Run this with an interpreter that has it, e.g.\n"
" ./.venv.erplibre/bin/python3 " + os.path.relpath(__file__)
)
def resolve(etree, arch, expr):
"""True if `expr` matches something in `arch`.
An arch that does not parse, or an expression lxml cannot evaluate, counts
as resolvable: this tool reports views that will CERTAINLY break, and must
never invent a failure it cannot substantiate.
"""
try:
tree = etree.fromstring(arch.encode("utf-8"))
except etree.XMLSyntaxError:
return True
try:
return bool(tree.xpath(expr))
except etree.XPathEvalError:
return True
def analyse(database):
"""[(copy, module_twin, [(child_id, failing_expr), ...])] — the COW copies
that have DRIFTED away from their module twin.
The test is differential, and it has to be. Odoo resolves an xpath against
the COMBINED arch of the whole inheritance chain, not against the parent's
own arch, so « does not resolve here » proves nothing on its own: a child
of ``website.layout`` legitimately targets ``//header``, which comes from
an ancestor. Comparing the copy with its module twin removes that whole
class of noise:
resolves in the module twin, not in the copy -> the copy lost an
anchor: real drift
resolves in neither -> the anchor lives
further up the chain
resolves in both -> nothing to see
Inactive children are skipped: Odoo never applies them, so they cannot
break a load.
"""
etree = require_lxml()
views = fetch_views(database)
module_by_key = {
v["key"]: v
for v in views.values()
if v["website_id"] is None and v["key"]
}
children = {}
for view in views.values():
if view["inherit_id"]:
children.setdefault(view["inherit_id"], []).append(view)
findings = []
for view in sorted(views.values(), key=lambda v: v["id"]):
twin = module_by_key.get(view["key"])
if view["website_id"] is None or twin is None:
continue
broken = []
for child in children.get(view["id"], []):
if not child.get("active", True):
continue
for expr in child_xpaths(child["arch"]):
if resolve(etree, twin["arch"], expr) and not resolve(
etree, view["arch"], expr
):
broken.append((child["id"], expr))
if broken:
findings.append((view, twin, broken))
return findings
def find_copy_by_key(views, key):
"""(copie COW, jumelle module) pour cette clé, ou (None, None).
La détection différentielle ne voit qu'une copie dont un ENFANT casse.
Une copie périmée sans enfant lui échappe — mesuré : la copie de
`website_crm.contactus_form` était plus petite d'un tiers que sa
jumelle, et c'est elle qui rendait /contactus en 500. Demander sa
réinitialisation par clé doit donc marcher même hors détection.
"""
copy = twin = None
for row in views.values():
if row.get("key") != key:
continue
if row.get("website_id"):
if copy is None or row["id"] < copy["id"]:
copy = row
elif twin is None or row["id"] < twin["id"]:
twin = row
return copy, twin
def render_diff(module_view, cow_view, indent=" "):
"""Module arch vs copy : ce qu'une réinitialisation rend et abandonne.
Rendu en texte plutôt qu'imprimé : la TUI montre exactement le même
diff que la ligne de commande. Deux rendus séparés dériveraient sans
que rien ne le signale.
"""
diff = difflib.unified_diff(
module_view["arch"].splitlines(),
cow_view["arch"].splitlines(),
fromfile=f"module id={module_view['id']}",
tofile=f"cow id={cow_view['id']}",
lineterm="",
)
return "\n".join(f"{indent}{line}" for line in diff)
def render_broken(cow_view, module_view, broken):
"""Pourquoi ça casse : les enfants dont l'xpath ne trouve plus son point."""
lines = [
f"── id={cow_view['id']} {cow_view['key']}"
f" (website={cow_view.get('website_id')}) ──",
"",
]
if not broken:
lines.append(
f" {t('No child fails on this copy: it drifted without')}"
f" {t('breaking anything yet.')}"
)
else:
lines.append(
f" {t('These children no longer find their anchor in the copy')}"
" :"
)
for child_id, expr in broken:
lines.append(f" {t('child')} {child_id} : {expr}")
lines += [
"",
f" {t('The anchor exists in the module view')}"
f" (id={module_view['id'] if module_view else '?'})"
f" {t('but not in this copy, frozen on an older version.')}",
"",
f" {t('Resetting restores the module arch; the customization the')}",
f" {t('copy carried is saved first, and is yours to re-apply as an')}",
f" {t('INHERITING view with its own key.')}",
]
return "\n".join(lines)
def show_diff(module_view, cow_view):
"""Imprimer le diff, pour la ligne de commande."""
print(render_diff(module_view, cow_view))
def backup(database, cow_view, directory):
"""Store the arch about to be replaced, and return the file path."""
os.makedirs(directory, exist_ok=True)
stamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
path = os.path.join(
directory, f"{cow_view['key']}_{cow_view['id']}_{stamp}.json"
)
with open(path, "w", encoding="utf-8") as fh:
json.dump(
{
"database": database,
"id": cow_view["id"],
"key": cow_view["key"],
"website_id": cow_view["website_id"],
"saved_at": stamp,
"arch_db": cow_view["arch"],
},
fh,
indent=2,
ensure_ascii=False,
)
return path
def reset(database, cow_view, module_view):
"""Copy the module arch over the stale copy. Returns rows updated."""
sql = (
"UPDATE ir_ui_view c SET arch_db = m.arch_db "
"FROM ir_ui_view m "
f"WHERE c.id = {cow_view['id']} AND m.id = {module_view['id']}"
)
run_psql(database, sql)
return 1
def main():
parser = argparse.ArgumentParser(
description=(
"Report the website COW copies whose children can no longer be"
" applied, and optionally reset them onto the module view."
)
)
parser.add_argument("-d", "--database", required=True)
parser.add_argument(
"--tui",
action="store_true",
help="browse the differences full screen",
)
parser.add_argument(
"--list-keys",
action="store_true",
help="print one drifted key per line, and nothing else",
)
parser.add_argument(
"--reset",
metavar="KEY",
action="append",
default=[],
help="reset this key onto its module view ('all' for every finding)",
)
parser.add_argument(
"--apply",
action="store_true",
help="really write; without it --reset only shows what it would do",
)
parser.add_argument(
"--backup-dir",
default=None,
help="where the replaced arch is saved"
" (default private/odoo/migration/<db>/cow_reset)",
)
config = parser.parse_args()
findings = analyse(config.database)
# Une sortie sans décor, pour que l'appelant construise un menu : on ne
# connaît pas les clés de tête, et les recopier depuis un diff de mille
# lignes est le genre de recopie où l'on se trompe.
if config.list_keys:
for cow_view, _module_view, _broken in findings:
print(cow_view["key"])
return 1 if findings else 0
# « Rien détecté » ne veut pas dire « rien à faire » quand une clé est
# demandée : la détection différentielle ne voit qu'une copie dont un
# ENFANT casse, et une copie périmée sans enfant lui échappe. Sortir ici
# faisait taire la demande — on croyait la copie réinitialisée.
if not findings and not (set(config.reset) - {"all"}):
print(f"✅ {t('No COW copy has drifted from its module view.')}")
return 0
if not findings:
# Le même texte que ci-dessus se lisait comme une contradiction :
# « aucune copie n'a dérivé », puis « ✅ réinitialisé id=2656 ».
# Les deux sont vrais — la détection différentielle ne voit qu'une
# copie dont un ENFANT casse — mais mis côte à côte sans un mot,
# on croit l'outil incohérent et l'on cesse de le lire.
print(
f"ℹ {t('The differential detection found nothing; resetting')}"
f" {t('the requested key(s) anyway.')}"
)
if findings:
print(
f"⚠️ {len(findings)}"
f" {t('COW copy(ies) drifted from their module')}"
f" {t('view in')} {config.database}"
)
print(
f" {t('Odoo surfaces this when the module view is rewritten')}"
f" {t('(a version bump) or when the page is rendered.')}\n"
)
for cow_view, module_view, broken in findings:
twin = (
f"module id={module_view['id']}"
if module_view
else t("NO module view with this key")
)
print(
f" id={cow_view['id']} key={cow_view['key']}"
f" website_id={cow_view['website_id']} [{twin}]"
)
for child_id, expr in broken:
print(f" {t('child')} {child_id} {t('cannot apply')}: {expr}")
if module_view:
show_diff(module_view, cow_view)
print()
if not config.reset:
print(
f"{t('Nothing changed. Re-run with --reset <key> --apply to')}"
f" {t('reset a copy onto its module view.')}"
)
return 1
if config.tui and findings:
from reset_stale_cow_tui import run_tui
# False n'est pas un échec : l'écran a dit pourquoi, et le rapport
# texte ci-dessus porte déjà la même information.
run_tui(findings, config.database)
wanted = set(config.reset)
directory = config.backup_dir or os.path.join(
"private", "odoo", "migration", config.database, "cow_reset"
)
done = 0
honoured = set()
missed = []
# Les vues brutes servent aux clés que la détection ne voit pas ; on ne
# les lit que si l'on va s'en servir.
views = fetch_views(config.database) if wanted - {"all"} else {}
for cow_view, module_view, _broken in findings:
if "all" not in wanted and cow_view["key"] not in wanted:
continue
honoured.add(cow_view["key"])
if not module_view:
print(
f"⏭ {cow_view['key']} :"
f" {t('no module view to reset onto, skipped.')}"
)
continue
if not config.apply:
print(
f"[{t('dry-run')}] {t('would reset')} id={cow_view['id']}"
f" ({cow_view['key']}) {t('onto')} id={module_view['id']}"
)
continue
path = backup(config.database, cow_view, directory)
reset(config.database, cow_view, module_view)
done += 1
print(f"✅ {t('reset')} id={cow_view['id']} ({cow_view['key']})")
print(f" {t('previous arch saved to')} {path}")
# Une clé demandée qui ne correspond à rien N'EST PAS un succès. Elle
# l'était : la commande tournait, ne faisait rien, et se taisait. On a
# donc cru une copie réinitialisée alors que /contactus rendait encore
# 500 — mesuré sur une vraie migration.
for key in sorted(wanted - honoured - {"all"}):
copy, twin = find_copy_by_key(views, key)
if copy is None:
print(f"⚠️ {t('No COW copy carries this key')} : {key}")
missed.append(key)
continue
if twin is None:
print(f"⚠️ {key} : {t('no module view to reset onto, skipped.')}")
missed.append(key)
continue
if copy["arch"] == twin["arch"]:
print(f"ℹ {key} : {t('already identical to the module view.')}")
continue
if not config.apply:
print(
f"[{t('dry-run')}] {t('would reset')} id={copy['id']}"
f" ({key}) {t('onto')} id={twin['id']}"
)
continue
path = backup(config.database, copy, directory)
reset(config.database, copy, twin)
done += 1
print(f"✅ {t('reset')} id={copy['id']} ({key})")
print(f" {t('previous arch saved to')} {path}")
if config.apply and done:
print(
f"\n{done} {t('copy(ies) reset. Re-apply any real customisation')}"
f" {t('as an INHERITING view, not a copy, so it cannot go stale')}"
f" {t('again.')}"
)
# Une demande non honorée doit se voir jusque dans le code de sortie :
# l'appelant qui enchaîne ne lit pas le texte.
return 2 if missed else 0
if __name__ == "__main__":
sys.exit(main())