Files
GGThedandClaude Opus 5 bda23dbb67 fix(ops): la sauvegarde des contrats archivait le mauvais repertoire
OBS-006. Trois racines etaient baties sur os.getcwd() : le magasin de
documents, les journaux et les sauvegardes. La vague G a corrige la
premiere, parce qu'elle bloquait aussi OPS-011, et a laisse les deux autres.
C'est la lecon deja consignee deux fois : un motif fautif corrige dans une
seule couche reste dans les autres.

Le plus serieux n'est pas le motif, c'est l'ecart qu'il a ouvert.
backup.py gardait sa propre constante DOCUMENTS_DIR sur os.getcwd(), donc
il ignorait DOCUMENTS_ROOT -- la variable que la vague G a introduite et que
docs/deployment.md dit maintenant de regler pour sortir les televersements
des repertoires de version. Des qu'un exploitant suit cette consigne, le
script archive un repertoire ou l'application n'a jamais rien ecrit. Et
comme il repond a un repertoire absent par une ligne d'information et un
code de sortie 0, une tache planifiee qui surveille le code de sortie voit
vert indefiniment.

Autrement dit : plus l'exploitant suivait correctement la documentation de
deploiement, plus surement ses sauvegardes de contrats etaient vides.

Les trois racines viennent desormais d'app/storage.py, resolues a l'appel et
non a l'import, et la sauvegarde imprime la source qu'elle a utilisee. Le
message d'absence nomme le chemin ou elle a cherche : "No documents
directory found" se lisait comme "il n'y a pas de documents" plutot que
comme "je regarde au mauvais endroit".

Le test qui porte est celui qui ouvre l'archive : un zip vide est un fichier
de taille non nulle, donc verifier qu'un fichier a ete produit ne prouvait
rien. Verifie par mutation.

Co-Authored-By: Claude Opus 5 <[email protected]>
2026-08-11 18:34:52 -04:00

150 lines
5.7 KiB
Python

"""Where the application's files live, and how the database refers to them.
Three roots: the project itself, the document store, and the log directory.
They are here together because they are the same defect three times over
(OBS-006) — `os.path.join(os.getcwd(), …)`, evaluated at import or upload
time, so every one of them moved with whatever directory the process was
started from. The document store was fixed first, in wave G, because it was
the one that also blocked OPS-011; the other two were left behind, which is
the repeated lesson of this project: a faulty pattern corrected in one layer
stays in the others.
Contracts were stored at `os.path.join(os.getcwd(), 'documents', …)`,
evaluated at upload time, and the resulting absolute path was written into
`Contract.file_path`. The storage root therefore moved with whatever
directory the process happened to be started from. Two consequences:
- one latent: start the server from elsewhere and new contracts land in a
new tree while the old ones become unreadable — with the database still
saying they are there, so the failure surfaces as a 500 on download
rather than as anything a person could act on;
- one blocking: it rules out a release-directory deployment (OPS-011)
outright. Every stored path would point inside a release that is about
to be replaced, so the first switch would take every contract ever
uploaded with it.
New rows keep a path *relative* to the document root. Old rows keep their
absolute path and are returned untouched, so this change needs no data
migration and can ship before Alembic does (DB-002).
"""
import os
#: Environment override for the document root. What a release-directory
#: deployment sets, to a path outside the releases — alongside them, not
#: inside whichever one is current.
DOCUMENTS_ROOT_ENV = 'DOCUMENTS_ROOT'
#: Environment override for the log directory. Same reasoning: a release
#: directory that carries its own logs loses them at the next switch.
LOG_DIR_ENV = 'LOG_DIR'
#: Environment override for the backup directory.
BACKUP_DIR_ENV = 'BACKUP_DIR'
#: Sub-directory holding uploaded contracts, under the document root.
CONTRACTS_DIR = 'contrats signés'
def project_root():
"""Absolute path of the project, derived from this file's location.
The anchor every other root falls back to. Not `os.getcwd()`: the
process is started by Waitress under Pterodactyl, by pytest, by a task
scheduler and by a person in a shell, and only one of those four is
reliably in the project directory.
"""
package_dir = os.path.dirname(os.path.abspath(__file__))
return os.path.dirname(package_dir)
def documents_root():
"""Absolute path of the document store.
Falls back to `documents/` beside this package — the project root
wherever it is installed, rather than wherever the process was launched.
"""
return _rooted(DOCUMENTS_ROOT_ENV, 'documents')
def logs_root():
"""Absolute path of the log directory.
Was `os.path.join(os.getcwd(), 'logs')`. Starting the server from
another directory sent the logs somewhere new without a word, which is
the worst possible failure mode for the one file you go and read when
something else has gone wrong.
"""
return _rooted(LOG_DIR_ENV, 'logs')
def backups_root():
"""Absolute path of the backup directory."""
return _rooted(BACKUP_DIR_ENV, 'backups')
def _rooted(env_name, default_name):
"""The configured path, or `default_name` under the project root."""
configured = os.getenv(env_name)
if configured:
return os.path.abspath(configured)
return os.path.join(project_root(), default_name)
def discard_documents(stored_paths):
"""Remove these documents from disk. Returns how many went (DATA-012).
`delete_user` removed the Contract rows and left the PDFs. Signed,
named contracts therefore stayed on the server after the account was
deleted, with nothing in the database pointing at them — invisible to
the application, unmanageable through it, and still personal data.
Call this **after** the commit that removed the rows, never before: a
failure between the two should leave a file with no row (recoverable,
and what the previous behaviour produced anyway) rather than a row with
no file (a download that 500s for ever).
A path that cannot be removed is logged and skipped. Nothing here should
be able to abort the deletion of an account.
"""
import logging
logger = logging.getLogger(__name__)
removed = 0
for stored_path in stored_paths:
if not stored_path:
continue
target = document_path(stored_path)
try:
os.remove(target)
removed += 1
except FileNotFoundError:
# Already gone. Two contracts sharing a stem, or a previous
# attempt: not a problem, and not worth an error line.
logger.info('Document already absent: %s', target)
except OSError as exc:
logger.error(
'Could not remove %s (%s). It is now an orphan: no database row '
'refers to it, so nothing in the application will ever offer to '
'delete it again.',
target,
exc,
)
return removed
def document_path(stored_path):
"""Absolute path of a document, from what the database holds.
Args:
stored_path: The value of Contract.file_path or signed_file_path.
Relative for rows written since this module existed, absolute
for the ones written before.
Returns:
str: An absolute path.
"""
if os.path.isabs(stored_path):
return stored_path
return os.path.join(documents_root(), stored_path)