Merge pull request #2799 from arc53/feat/install-4-backup

docsgpt backup and restore

docsgpt/cli.py

@@ -230,6 +230,18 @@ def _add_deploy_commands(commands) -> None:
     uninstall.add_argument("-y", "--yes", action="store_true", help="do not ask for confirmation")
     uninstall.add_argument("--purge", action="store_true", help="also delete the settings and all data")
 
+    backup = stack_command("backup", "backup", "write a backup of the database and the uploaded data")
+    backup.add_argument("--out", help="directory for the archive (default: <stack>/backups)")
+    backup.add_argument("--with-settings", action="store_true",
+                        help="include .env in the archive; it holds this install's secrets")
+
+    restore = stack_command("restore", "restore", "restore a backup over this install")
+    restore.add_argument("archive", help="the .tar.gz written by `docsgpt backup`")
+    restore.add_argument("-y", "--yes", action="store_true", help="do not ask for confirmation")
+    restore.add_argument("--force", action="store_true", help="restore a backup taken with a newer DocsGPT")
+    restore.add_argument("--timeout", type=int, default=300,
+                         help="seconds to wait for the API afterwards (default: 300)")
+
     env = stack_command("env", "env", "show, get or set the stack's settings")
     env_actions = env.add_subparsers(dest="env_action", metavar="<action>")
     get = env_actions.add_parser("get", help="print one setting")

docsgpt/deploy/backup.py

@@ -0,0 +1,149 @@
+"""The archive ``docsgpt backup`` writes and ``docsgpt restore`` reads.
+
+One gzipped tar holds a SQL dump of the database, a tar per data volume, and a
+manifest saying which version and image the backup came from. The settings file
+is left out unless it is asked for: it holds the secrets.
+
+``postgres_data`` is not tarred, because the dump is the database backup and a
+copy of a running data directory would be a torn one. Caddy's volumes are left
+out too: they hold certificates it obtains again on the next start.
+"""
+
+from __future__ import annotations
+
+import io
+import json
+import os
+import tarfile
+import time
+from collections.abc import Mapping
+from datetime import datetime
+from pathlib import Path
+from typing import Optional
+
+from docsgpt.deploy.docker import DeployError
+
+FORMAT = 1
+MANIFEST = "manifest.json"
+DUMP = "database.sql"
+SETTINGS = "settings.env"
+VOLUME_DIR = "volumes"
+DATA_VOLUMES = ("indexes", "inputs", "vectors")
+
+
+def archive_name(when: datetime) -> str:
+    """The file name for a backup taken at ``when``."""
+    return f"docsgpt-{when.strftime('%Y%m%d-%H%M%S')}.tar.gz"
+
+
+def volume_member(name: str) -> str:
+    """Where a volume's tar sits inside the archive."""
+    return f"{VOLUME_DIR}/{name}.tar"
+
+
+def write_archive(
+    path: Path,
+    *,
+    dump: Path,
+    volume_tars: Mapping[str, Path],
+    manifest: Mapping[str, object],
+    settings: Optional[Path] = None,
+) -> None:
+    """Write the backup archive; the manifest is added last so a truncated file has none."""
+    path.parent.mkdir(parents=True, exist_ok=True)
+    # 0600 from the start: the dump is the install's data, and --with-settings adds its secrets.
+    descriptor = os.open(path, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600)
+    os.fchmod(descriptor, 0o600)
+    with os.fdopen(descriptor, "wb") as handle, tarfile.open(fileobj=handle, mode="w:gz") as archive:
+        archive.add(dump, arcname=DUMP)
+        for name, tar_path in sorted(volume_tars.items()):
+            archive.add(tar_path, arcname=volume_member(name))
+        if settings is not None:
+            archive.add(settings, arcname=SETTINGS)
+        body = json.dumps(dict(manifest), indent=2).encode("utf-8")
+        info = tarfile.TarInfo(MANIFEST)
+        info.size = len(body)
+        info.mtime = int(time.time())
+        info.mode = 0o600
+        archive.addfile(info, io.BytesIO(body))
+
+
+def read_manifest(path: Path) -> dict:
+    """The archive's manifest, or a DeployError naming what is wrong with the file."""
+    if not path.is_file():
+        raise DeployError(f"{path} does not exist")
+    try:
+        with tarfile.open(path, "r:gz") as archive:
+            member = archive.extractfile(MANIFEST)
+            if member is None:
+                raise KeyError(MANIFEST)
+            manifest = json.loads(member.read().decode("utf-8"))
+    except (tarfile.TarError, KeyError, ValueError, OSError) as exc:
+        raise DeployError(f"{path} is not a DocsGPT backup: {exc}") from exc
+    if not isinstance(manifest, dict) or "volumes" not in manifest:
+        raise DeployError(f"{path} is not a DocsGPT backup: its manifest is missing what to restore")
+    return manifest
+
+
+def validate(path: Path, manifest: Mapping[str, object]) -> list[str]:
+    """The volumes to restore, once the archive is known to hold everything it declares.
+
+    Checked before the stack is stopped: a damaged or hand-made archive must fail while DocsGPT is
+    still running, not after ``docker compose down``. Only the volumes a backup is made of are
+    accepted, so a manifest cannot name ``postgres_data`` and have it emptied on the way in.
+    """
+    volumes = manifest.get("volumes")
+    if not isinstance(volumes, list) or not all(isinstance(name, str) for name in volumes):
+        raise DeployError(f"{path} is not a DocsGPT backup: its manifest does not list the volumes it holds")
+    unsupported = sorted(set(volumes) - set(DATA_VOLUMES))
+    if unsupported:
+        raise DeployError(
+            f"{path} names volumes that are not part of a backup: {', '.join(unsupported)}. "
+            f"A DocsGPT backup holds {', '.join(DATA_VOLUMES)}."
+        )
+    required = [DUMP, *(volume_member(name) for name in volumes)]
+    try:
+        with tarfile.open(path, "r:gz") as archive:
+            present = {member.name for member in archive.getmembers() if member.isfile()}
+    except (tarfile.TarError, OSError) as exc:
+        raise DeployError(f"{path} is not a DocsGPT backup: {exc}") from exc
+    missing = [name for name in required if name not in present]
+    if missing:
+        raise DeployError(f"{path} is missing {', '.join(missing)}, so there is nothing to restore from")
+    return list(volumes)
+
+
+def check_volume_tar(name: str, path: Path) -> None:
+    """Read a volume tar through, so a damaged one is found before any volume is replaced."""
+    try:
+        with tarfile.open(path, "r:*") as archive:
+            for _ in archive:
+                pass
+    except (tarfile.TarError, OSError) as exc:
+        raise DeployError(f"the {name} volume in this backup is damaged: {exc}") from exc
+
+
+def extract(path: Path, destination: Path) -> None:
+    """Unpack the archive into ``destination`` (data filter: no paths outside it, no devices)."""
+    with tarfile.open(path, "r:gz") as archive:
+        archive.extractall(destination, filter="data")
+
+
+def _parts(version: str) -> tuple[int, ...]:
+    numbers = []
+    for chunk in str(version).split("."):
+        digits = "".join(character for character in chunk if character.isdigit())
+        numbers.append(int(digits) if digits else 0)
+    return tuple(numbers)
+
+
+def check_version(manifest: Mapping[str, object], current: str, force: bool) -> None:
+    """Refuse a backup from a newer DocsGPT: its data may not fit this version's schema."""
+    taken_with = str(manifest.get("version") or "")
+    if force or not taken_with:
+        return
+    if _parts(taken_with) > _parts(current):
+        raise DeployError(
+            f"this backup is from DocsGPT {taken_with}, newer than the installed {current}. "
+            "Upgrade first with `docsgpt upgrade`, or pass --force to restore it anyway."
+        )

docsgpt/deploy/commands.py

@@ -9,6 +9,7 @@ import re
 import shutil
 import subprocess
 import sys
+import tempfile
 import webbrowser
 from collections.abc import Callable, Mapping
 from dataclasses import dataclass
@@ -16,6 +17,7 @@ from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any, Optional
 
+from docsgpt.deploy import backup as backup_format
 from docsgpt.deploy import envfile, stack
 from docsgpt.deploy.docker import DeployError, Docker, lan_ip, wait_healthy
 
@@ -350,6 +352,149 @@ def env(args, context: Optional[Context] = None) -> int:
     return 0
 
 
+def _stack_image(env: Mapping[str, str]) -> str:
+    """The image this install runs; the volume tars go through it, so nothing extra is pulled."""
+    tag = env.get("DOCSGPT_IMAGE_TAG") or "latest"
+    return f"arc53/docsgpt:{tag}{env.get('DOCSGPT_IMAGE_VARIANT', '')}"
+
+
+def backup(args, context: Optional[Context] = None) -> int:
+    """Write a dump of the database and a tar of each data volume into one archive."""
+    context = context or Context.default(args)
+    directory = stack.stack_dir(args.dir)
+    env = _installed(directory)
+    if env is None:
+        return 1
+
+    out_dir = Path(args.out).expanduser() if args.out else directory / "backups"
+    taken_at = datetime.now(timezone.utc)
+    target = out_dir / backup_format.archive_name(taken_at)
+    image = _stack_image(env)
+
+    try:
+        print("Pausing the backend and the worker so the database and the files match ...")
+        context.docker.compose(directory, "stop", "backend", "worker")
+        _write_backup(args, context, directory, env, target, image, taken_at)
+    finally:
+        print("Starting the backend and the worker again ...")
+        context.docker.compose(directory, "up", "-d", "backend", "worker")
+
+    size = target.stat().st_size / 1_000_000
+    print(f"\nBackup written to {target} ({size:.1f} MB)")
+    if args.with_settings:
+        print("It contains .env, so it holds this install's secrets: keep it somewhere private.")
+    else:
+        print("Settings are not in it; `docsgpt backup --with-settings` includes .env, secrets and all.")
+    print(f"Restore it with: docsgpt restore {target}")
+    return 0
+
+
+def _write_backup(args, context: Context, directory: Path, env, target: Path, image: str, taken_at) -> None:
+    """Dump the database and the data volumes into ``target``, with the writers stopped."""
+    with tempfile.TemporaryDirectory() as workspace:
+        work = Path(workspace)
+        dump = work / backup_format.DUMP
+        print("Dumping the database ...")
+        with dump.open("w", encoding="utf-8") as handle:
+            context.docker.compose(
+                directory, "exec", "-T", "postgres",
+                "pg_dump", "--clean", "--if-exists", "-U", "docsgpt", "-d", "docsgpt",
+                stdout=handle,
+            )
+        volume_tars = {}
+        for name in backup_format.DATA_VOLUMES:
+            print(f"Archiving the {name} volume ...")
+            tar_path = work / f"{name}.tar"
+            context.docker.export_volume(f"{PROJECT}_{name}", tar_path, image)
+            volume_tars[name] = tar_path
+        manifest = {
+            "format": backup_format.FORMAT,
+            "created_at": taken_at.isoformat(timespec="seconds"),
+            "version": context.version,
+            "image_tag": env.get("DOCSGPT_IMAGE_TAG", ""),
+            "volumes": list(backup_format.DATA_VOLUMES),
+            "settings_included": bool(args.with_settings),
+        }
+        settings = directory / ".env" if args.with_settings else None
+        backup_format.write_archive(target, dump=dump, volume_tars=volume_tars, manifest=manifest, settings=settings)
+
+
+def _restore_data(context: Context, directory: Path, archive: Path, volumes: list, image: str) -> None:
+    """Put the archive's volumes and database back, with the stack stopped."""
+    with tempfile.TemporaryDirectory() as workspace:
+        work = Path(workspace)
+        backup_format.extract(archive, work)
+        # Read every volume tar through before replacing any of them: a damaged third tar must not
+        # be discovered with the first two already swapped in.
+        tars = {}
+        for name in volumes:
+            tar_path = work / backup_format.volume_member(name)
+            if not tar_path.is_file():
+                raise DeployError(f"{archive} is missing the {name} volume it says it contains")
+            backup_format.check_volume_tar(name, tar_path)
+            tars[name] = tar_path
+        for name, tar_path in tars.items():
+            print(f"Restoring the {name} volume ...")
+            context.docker.import_volume(f"{PROJECT}_{name}", tar_path, image)
+
+        dump = work / backup_format.DUMP
+        if not dump.is_file():
+            raise DeployError(f"{archive} is missing its database dump")
+        print("Starting the database ...")
+        context.docker.compose(directory, "up", "-d", "--wait", "postgres")
+        print("Restoring the database ...")
+        with dump.open("r", encoding="utf-8") as handle:
+            context.docker.compose(
+                directory, "exec", "-T", "postgres",
+                "psql", "--quiet", "--set", "ON_ERROR_STOP=on", "-U", "docsgpt", "-d", "docsgpt",
+                stdin=handle,
+            )
+
+
+def restore(args, context: Optional[Context] = None) -> int:
+    """Put a backup's database and data volumes back over this install."""
+    context = context or Context.default(args)
+    archive = Path(args.archive).expanduser()
+    manifest = backup_format.read_manifest(archive)
+    backup_format.check_version(manifest, context.version, args.force)
+    # Everything the archive declares is checked here, while DocsGPT is still up.
+    volumes = backup_format.validate(archive, manifest)
+
+    directory = stack.stack_dir(args.dir)
+    env = _installed(directory)
+    if env is None:
+        return 1
+
+    taken_at = manifest.get("created_at", "an unknown time")
+    if not args.yes:
+        if not context.interactive:
+            raise DeployError("restore needs --yes when there is no terminal to confirm on")
+        question = f"Replace the data in {directory} with the backup from {taken_at}? This cannot be undone."
+        if not context.prompter.confirm(question, default=False):
+            print("Nothing restored.")
+            return 1
+
+    image = _stack_image(env)
+    try:
+        print("Stopping the stack ...")
+        context.docker.compose(directory, *_EVERY_PROFILE, "down")
+        _restore_data(context, directory, archive, volumes, image)
+    except BaseException:
+        # The stack is down by now. A corrupt payload inside an otherwise well-formed archive, or a
+        # statement psql refuses, must not leave DocsGPT stopped: start it again, then report.
+        print("The restore failed. Starting DocsGPT again ...", file=sys.stderr)
+        context.docker.compose(directory, "up", "-d", "--remove-orphans", check=False)
+        raise
+
+    print("Starting DocsGPT ...")
+    context.docker.compose(directory, "up", "-d", "--remove-orphans")
+    if not context.wait(stack.health_url(env), args.timeout):
+        print("DocsGPT did not answer after the restore. See `docsgpt logs backend`.", file=sys.stderr)
+        return 1
+    print(f"\nRestored the backup from {taken_at}. DocsGPT is running at {stack.url(env, context.lan_ip())}")
+    return 0
+
+
 def _uninstall_hint(installer: str) -> str:
     return {"uv": "uv tool uninstall docsgpt", "pipx": "pipx uninstall docsgpt"}.get(installer, "pip uninstall docsgpt")
 

docsgpt/deploy/docker.py

@@ -22,10 +22,16 @@ class DeployError(Exception):
     """A problem the user can act on; the command prints it without a traceback."""
 
 
-def run(args: Sequence[str], *, cwd: Optional[Path] = None, capture: bool = False, check: bool = True):
-    """Run a command, streaming its output unless ``capture``; with ``check`` a failure raises DeployError."""
+def run(args: Sequence[str], *, cwd: Optional[Path] = None, capture: bool = False, check: bool = True,
+        stdout=None, stdin=None):
+    """Run a command, streaming its output unless ``capture`` or a file is given for ``stdout``.
+
+    ``stdout`` and ``stdin`` take open files, so a dump goes straight to disk and
+    back in again without passing through this process.
+    """
+    streams = {"capture_output": True} if capture else {"stdout": stdout, "stdin": stdin}
     try:
-        result = subprocess.run(list(args), cwd=cwd, text=True, capture_output=capture, check=False)
+        result = subprocess.run(list(args), cwd=cwd, text=True, check=False, **streams)
     except FileNotFoundError as exc:
         raise DeployError(f"{args[0]} is not installed or not on PATH") from exc
     if check and result.returncode != 0:
@@ -94,9 +100,37 @@ class Docker:
             raise DeployError("Docker is not running. Start it with `sudo systemctl start docker` and run this again.")
         raise DeployError("Docker is not running. Start Docker Desktop and run this again.")
 
-    def compose(self, directory: Path, *args: str, capture: bool = False, check: bool = True):
+    def compose(self, directory: Path, *args: str, capture: bool = False, check: bool = True,
+                stdout=None, stdin=None):
         """``docker compose <args>`` in ``directory``, which holds the Compose file and its ``.env``."""
-        return self._run(["docker", "compose", *args], cwd=directory, capture=capture, check=check)
+        return self._run(["docker", "compose", *args], cwd=directory, capture=capture, check=check,
+                         stdout=stdout, stdin=stdin)
+
+    def export_volume(self, volume: str, dest: Path, image: str) -> None:
+        """Write ``volume`` to ``dest`` as a tar, through an image the install already has."""
+        dest.parent.mkdir(parents=True, exist_ok=True)
+        with dest.open("wb") as handle:
+            self._run(
+                ["docker", "run", "--rm", "-v", f"{volume}:/data:ro", image, "tar", "cf", "-", "-C", "/data", "."],
+                stdout=handle,
+            )
+
+    def import_volume(self, volume: str, source: Path, image: str) -> None:
+        """Replace ``volume``'s contents with the tar at ``source``; the volume is created when missing."""
+        # Unpack into a throwaway directory inside the container first, so a truncated or corrupt
+        # tar fails before the live volume is touched rather than halfway through emptying it. The
+        # container makes the directory itself: the image does not run as root, and a fixed path
+        # would be both a guess about what is writable and a temp-file smell.
+        script = (
+            'set -e; stage=$(mktemp -d); tar xf - -C "$stage"; '
+            'find /data -mindepth 1 -delete; tar cf - -C "$stage" . | tar xf - -C /data; '
+            'rm -rf "$stage"'
+        )
+        with source.open("rb") as handle:
+            self._run(
+                ["docker", "run", "--rm", "-i", "-v", f"{volume}:/data", image, "sh", "-c", script],
+                stdin=handle,
+            )
 
     def volume_exists(self, name: str) -> bool:
         return self._run(["docker", "volume", "inspect", name], capture=True, check=False).returncode == 0