mirror of
https://github.com/Garmelon/PFERD.git
synced 2023-12-21 10:23:01 +01:00
256 lines
10 KiB
Python
256 lines
10 KiB
Python
"""
|
|
Convenience functions for using PFERD.
|
|
"""
|
|
|
|
import logging
|
|
from pathlib import Path
|
|
from typing import Callable, List, Optional, Union
|
|
|
|
from .cookie_jar import CookieJar
|
|
from .diva import (DivaDownloader, DivaDownloadStrategy, DivaPlaylistCrawler,
|
|
diva_download_new)
|
|
from .errors import FatalException, swallow_and_print_errors
|
|
from .ilias import (IliasAuthenticator, IliasCrawler, IliasDirectoryFilter,
|
|
IliasDownloader, IliasDownloadInfo, IliasDownloadStrategy,
|
|
KitShibbolethAuthenticator, download_modified_or_new)
|
|
from .location import Location
|
|
from .logging import PrettyLogger
|
|
from .organizer import Organizer
|
|
from .tmp_dir import TmpDir
|
|
from .transform import TF, Transform, apply_transform
|
|
from .utils import PathLike, to_path
|
|
|
|
# TODO save known-good cookies as soon as possible
|
|
|
|
|
|
LOGGER = logging.getLogger(__name__)
|
|
PRETTY = PrettyLogger(LOGGER)
|
|
|
|
|
|
class Pferd(Location):
|
|
# pylint: disable=too-many-arguments
|
|
"""
|
|
The main entrypoint in your Pferd usage: This class combines a number of
|
|
useful shortcuts for running synchronizers in a single interface.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
base_dir: Path,
|
|
tmp_dir: Path = Path(".tmp"),
|
|
test_run: bool = False
|
|
):
|
|
super().__init__(Path(base_dir))
|
|
|
|
self._tmp_dir = TmpDir(self.resolve(tmp_dir))
|
|
self._test_run = test_run
|
|
|
|
@staticmethod
|
|
def _print_transformables(transformables: List[TF]) -> None:
|
|
LOGGER.info("")
|
|
LOGGER.info("Results of the test run:")
|
|
for transformable in transformables:
|
|
LOGGER.info(transformable.path)
|
|
|
|
def _ilias(
|
|
self,
|
|
target: PathLike,
|
|
base_url: str,
|
|
crawl_function: Callable[[IliasCrawler], List[IliasDownloadInfo]],
|
|
authenticator: IliasAuthenticator,
|
|
cookies: Optional[PathLike],
|
|
dir_filter: IliasDirectoryFilter,
|
|
transform: Transform,
|
|
download_strategy: IliasDownloadStrategy,
|
|
clean: bool = True
|
|
) -> Organizer:
|
|
# pylint: disable=too-many-locals
|
|
cookie_jar = CookieJar(to_path(cookies) if cookies else None)
|
|
session = cookie_jar.create_session()
|
|
tmp_dir = self._tmp_dir.new_subdir()
|
|
organizer = Organizer(self.resolve(to_path(target)))
|
|
|
|
crawler = IliasCrawler(base_url, session, authenticator, dir_filter)
|
|
downloader = IliasDownloader(tmp_dir, organizer, session, authenticator, download_strategy)
|
|
|
|
cookie_jar.load_cookies()
|
|
info = crawl_function(crawler)
|
|
cookie_jar.save_cookies()
|
|
|
|
transformed = apply_transform(transform, info)
|
|
if self._test_run:
|
|
self._print_transformables(transformed)
|
|
return organizer
|
|
|
|
downloader.download_all(transformed)
|
|
cookie_jar.save_cookies()
|
|
|
|
if clean:
|
|
organizer.cleanup()
|
|
|
|
return organizer
|
|
|
|
@swallow_and_print_errors
|
|
def ilias_kit(
|
|
self,
|
|
target: PathLike,
|
|
course_id: str,
|
|
dir_filter: IliasDirectoryFilter = lambda x: True,
|
|
transform: Transform = lambda x: x,
|
|
cookies: Optional[PathLike] = None,
|
|
username: Optional[str] = None,
|
|
password: Optional[str] = None,
|
|
download_strategy: IliasDownloadStrategy = download_modified_or_new,
|
|
clean: bool = True,
|
|
) -> Organizer:
|
|
"""
|
|
Synchronizes a folder with the ILIAS instance of the KIT.
|
|
|
|
Arguments:
|
|
target {Path} -- the target path to write the data to
|
|
course_id {str} -- the id of the main course page (found in the URL after ref_id
|
|
when opening the course homepage)
|
|
|
|
Keyword Arguments:
|
|
dir_filter {IliasDirectoryFilter} -- A filter for directories. Will be applied on the
|
|
crawler level, these directories and all of their content is skipped.
|
|
(default: {lambdax:True})
|
|
transform {Transform} -- A transformation function for the output paths. Return None
|
|
to ignore a file. (default: {lambdax:x})
|
|
cookies {Optional[Path]} -- The path to store and load cookies from.
|
|
(default: {None})
|
|
username {Optional[str]} -- The SCC username. If none is given, it will prompt
|
|
the user. (default: {None})
|
|
password {Optional[str]} -- The SCC password. If none is given, it will prompt
|
|
the user. (default: {None})
|
|
download_strategy {DownloadStrategy} -- A function to determine which files need to
|
|
be downloaded. Can save bandwidth and reduce the number of requests.
|
|
(default: {download_modified_or_new})
|
|
clean {bool} -- Whether to clean up when the method finishes.
|
|
"""
|
|
# This authenticator only works with the KIT ilias instance.
|
|
authenticator = KitShibbolethAuthenticator(username=username, password=password)
|
|
PRETTY.starting_synchronizer(target, "ILIAS", course_id)
|
|
return self._ilias(
|
|
target=target,
|
|
base_url="https://ilias.studium.kit.edu/",
|
|
crawl_function=lambda crawler: crawler.crawl_course(course_id),
|
|
authenticator=authenticator,
|
|
cookies=cookies,
|
|
dir_filter=dir_filter,
|
|
transform=transform,
|
|
download_strategy=download_strategy,
|
|
clean=clean,
|
|
)
|
|
|
|
@swallow_and_print_errors
|
|
def ilias_kit_personal_desktop(
|
|
self,
|
|
target: PathLike,
|
|
dir_filter: IliasDirectoryFilter = lambda x: True,
|
|
transform: Transform = lambda x: x,
|
|
cookies: Optional[PathLike] = None,
|
|
username: Optional[str] = None,
|
|
password: Optional[str] = None,
|
|
download_strategy: IliasDownloadStrategy = download_modified_or_new,
|
|
clean: bool = True,
|
|
) -> Organizer:
|
|
"""
|
|
Synchronizes a folder with the ILIAS instance of the KIT. This method will crawl the ILIAS
|
|
"personal desktop" instead of a single course.
|
|
|
|
Arguments:
|
|
target {Path} -- the target path to write the data to
|
|
|
|
Keyword Arguments:
|
|
dir_filter {IliasDirectoryFilter} -- A filter for directories. Will be applied on the
|
|
crawler level, these directories and all of their content is skipped.
|
|
(default: {lambdax:True})
|
|
transform {Transform} -- A transformation function for the output paths. Return None
|
|
to ignore a file. (default: {lambdax:x})
|
|
cookies {Optional[Path]} -- The path to store and load cookies from.
|
|
(default: {None})
|
|
username {Optional[str]} -- The SCC username. If none is given, it will prompt
|
|
the user. (default: {None})
|
|
password {Optional[str]} -- The SCC password. If none is given, it will prompt
|
|
the user. (default: {None})
|
|
download_strategy {DownloadStrategy} -- A function to determine which files need to
|
|
be downloaded. Can save bandwidth and reduce the number of requests.
|
|
(default: {download_modified_or_new})
|
|
clean {bool} -- Whether to clean up when the method finishes.
|
|
"""
|
|
# This authenticator only works with the KIT ilias instance.
|
|
authenticator = KitShibbolethAuthenticator(username=username, password=password)
|
|
PRETTY.starting_synchronizer(target, "ILIAS", "Personal Desktop")
|
|
return self._ilias(
|
|
target=target,
|
|
base_url="https://ilias.studium.kit.edu/",
|
|
crawl_function=lambda crawler: crawler.crawl_personal_desktop(),
|
|
authenticator=authenticator,
|
|
cookies=cookies,
|
|
dir_filter=dir_filter,
|
|
transform=transform,
|
|
download_strategy=download_strategy,
|
|
clean=clean,
|
|
)
|
|
|
|
@swallow_and_print_errors
|
|
def diva_kit(
|
|
self,
|
|
target: Union[PathLike, Organizer],
|
|
playlist_location: str,
|
|
transform: Transform = lambda x: x,
|
|
download_strategy: DivaDownloadStrategy = diva_download_new,
|
|
clean: bool = True
|
|
) -> Organizer:
|
|
"""
|
|
Synchronizes a folder with a DIVA playlist.
|
|
|
|
Arguments:
|
|
organizer {Organizer} -- The organizer to use.
|
|
playlist_location {str} -- the playlist id or the playlist URL
|
|
in the format 'https://mediaservice.bibliothek.kit.edu/#/details/DIVA-2019-271'
|
|
|
|
Keyword Arguments:
|
|
transform {Transform} -- A transformation function for the output paths. Return None
|
|
to ignore a file. (default: {lambdax:x})
|
|
download_strategy {DivaDownloadStrategy} -- A function to determine which files need to
|
|
be downloaded. Can save bandwidth and reduce the number of requests.
|
|
(default: {diva_download_new})
|
|
clean {bool} -- Whether to clean up when the method finishes.
|
|
"""
|
|
tmp_dir = self._tmp_dir.new_subdir()
|
|
|
|
if playlist_location.startswith("http"):
|
|
playlist_id = DivaPlaylistCrawler.fetch_id(playlist_link=playlist_location)
|
|
else:
|
|
playlist_id = playlist_location
|
|
|
|
if target is None:
|
|
PRETTY.starting_synchronizer("None", "DIVA", playlist_id)
|
|
raise FatalException("Got 'None' as target directory, aborting")
|
|
|
|
if isinstance(target, Organizer):
|
|
organizer = target
|
|
else:
|
|
organizer = Organizer(self.resolve(to_path(target)))
|
|
|
|
PRETTY.starting_synchronizer(organizer.path, "DIVA", playlist_id)
|
|
|
|
crawler = DivaPlaylistCrawler(playlist_id)
|
|
downloader = DivaDownloader(tmp_dir, organizer, download_strategy)
|
|
|
|
info = crawler.crawl()
|
|
|
|
transformed = apply_transform(transform, info)
|
|
if self._test_run:
|
|
self._print_transformables(transformed)
|
|
return organizer
|
|
|
|
downloader.download_all(transformed)
|
|
|
|
if clean:
|
|
organizer.cleanup()
|
|
|
|
return organizer
|