# Copyright (C) 2015-2024 The Software Heritage developers
# See the AUTHORS file at the top-level directory of this distribution
# License: GNU General Public License version 3, or any later version
# See top-level LICENSE file for more information
import logging
import queue
import threading
from typing import Dict, Iterable, Iterator, List, Mapping, Optional, Set, Tuple, Union
# note: it's required to access the statsd object form the statsd module to
# help mocking it in tests...
from swh.core import statsd
from swh.core.api import RemoteException
from swh.model.model import Sha1
from swh.objstorage.exc import (
NoBackendsLeftError,
ObjCorruptedError,
ObjNotFoundError,
ObjStorageAPIError,
ReadOnlyObjStorageError,
)
from swh.objstorage.factory import get_objstorage
from swh.objstorage.interface import CompositeObjId, ObjId, ObjStorageInterface
from swh.objstorage.objstorage import ObjStorage, timed
from swh.objstorage.utils import format_obj_id
MP_COUNTER_METRICS = "swh_objstorage_multiplexer_backend_total"
MP_BACKEND_DISABLED_METRICS = "swh_objstorage_multiplexer_backend_disabled_total"
MP_BACKEND_ENABLED_METRICS = "swh_objstorage_multiplexer_backend_enabled_total"
logger = logging.getLogger(__name__)
DEFAULT_TRANSIENT_READ_EXCEPTIONS = (
"requests.exceptions.ConnectTimeout",
"requests.exceptions.ReadTimeout",
"requests.exceptions.SSLError",
"requests.exceptions.ConnectionError",
"builtins.TimeoutError",
"builtins.IOError",
)
[docs]
class ObjStorageThread(threading.Thread):
def __init__(self, storage):
super().__init__(daemon=True)
self.storage = storage
self.commands = queue.Queue()
[docs]
def run(self):
while True:
try:
mailbox, command, args, kwargs = self.commands.get(True, 0.05)
except queue.Empty:
continue
try:
ret = getattr(self.storage, command)(*args, **kwargs)
except Exception as exc:
self.queue_result(mailbox, "exception", exc)
else:
self.queue_result(mailbox, "result", ret)
[docs]
def queue_command(self, command, *args, mailbox=None, **kwargs):
"""Enqueue a new command to be processed by the thread.
Args:
command (str): one of the method names for the underlying storage.
mailbox (queue.Queue): explicit mailbox if the calling thread wants
to override it.
args, kwargs: arguments for the command.
Returns:
queue.Queue: The mailbox you can read the response from
"""
if not mailbox:
mailbox = queue.Queue()
self.commands.put((mailbox, command, args, kwargs))
return mailbox
[docs]
def queue_result(self, mailbox, result_type, result):
"""Enqueue a new result in the mailbox
This also provides a reference to the storage, which can be useful when
an exceptional condition arises.
Args:
mailbox (queue.Queue): the mailbox to which we need to enqueue the
result
result_type (str): one of 'result', 'exception'
result: the result to pass back to the calling thread
"""
mailbox.put(
{
"type": result_type,
"result": result,
}
)
[docs]
@staticmethod
def get_result_from_mailbox(mailbox, *args, **kwargs):
"""Unpack the result from the mailbox.
Args:
mailbox (queue.Queue): A mailbox to unpack a result from
args: positional arguments to :func:`mailbox.get`
kwargs: keyword arguments to :func:`mailbox.get`
Returns:
the next result unpacked from the queue
Raises:
either the exception we got back from the underlying storage,
or :exc:`queue.Empty` if :func:`mailbox.get` raises that.
"""
result = mailbox.get(*args, **kwargs)
if result["type"] == "exception":
raise result["result"] from None
else:
return result["result"]
[docs]
@staticmethod
def collect_results(mailbox, num_results):
"""Collect num_results from the mailbox"""
collected = 0
ret = []
while collected < num_results:
try:
ret.append(
ObjStorageThread.get_result_from_mailbox(mailbox, True, 0.05)
)
except queue.Empty:
continue
collected += 1
return ret
def __getattr__(self, attr):
def call(*args, **kwargs):
mailbox = self.queue_command(attr, *args, **kwargs)
return self.get_result_from_mailbox(mailbox)
return call
def __contains__(self, *args, **kwargs):
mailbox = self.queue_command("__contains__", *args, **kwargs)
return self.get_result_from_mailbox(mailbox)
[docs]
class MultiplexerObjStorage(ObjStorage):
"""Implementation of ObjStorage that distributes between multiple
storages.
The multiplexer object storage allows an input to be demultiplexed
among multiple storages that will or will not accept it by
themselves.
As the ids can be different, no pre-computed ids should be
submitted. Also, there are no guarantees that the returned ids
can be used directly into the storages that the multiplexer
manage.
Use case examples follow.
Example::
storage_v1 = ReadOnlyProxyObjStorage(
PathSlicingObjStorage('/dir1', '0:2/2:4/4:6')
)
storage_v2 = PathSlicingObjStorage('/dir2', '0:1/0:5')
storage = MultiplexerObjStorage([storage_v1, storage_v2])
When using 'storage', all the new contents will only be added to the v2
storage, while it will be retrievable from both.
"""
name: str = "multiplexer"
def __init__(
self,
*,
objstorages: Iterable[Union[ObjStorageInterface, Dict]],
read_exception_cooldown: float = 5,
transient_read_exceptions: Optional[List[str]] = None,
readonly: bool = False,
**kwargs,
):
super().__init__(**kwargs)
self.readonly = readonly
self.storages = [
get_objstorage(**sto) if isinstance(sto, dict) else sto
for sto in objstorages
]
self.storage_threads = [ObjStorageThread(storage) for storage in self.storages]
for thread in self.storage_threads:
thread.start()
self.write_storage_threads = []
if not readonly:
for thread, storage in zip(self.storage_threads, self.storages):
try:
checked = storage.check_config(check_write=True)
except PermissionError:
checked = False
except RemoteException:
logger.warning(
"Received RemoteException when calling check_config on backend %s:",
storage.name,
exc_info=True,
)
checked = False
if checked:
self.write_storage_threads.append(thread)
self.read_exception_cooldown = read_exception_cooldown
self.transient_read_exceptions = set(
transient_read_exceptions or DEFAULT_TRANSIENT_READ_EXCEPTIONS
)
self.active_readers: Set[int] = set()
self.reset_timers: Dict[int, threading.Timer] = {}
self.reset_active_readers()
[docs]
def reset_active_readers(self):
"""Reset the active readers set to all storages, and cancel all reset_failed_threads"""
for t in self.reset_timers.values():
if t.is_alive():
t.cancel()
self.reset_timers = {}
self.active_readers = set(range(len(self.storages)))
[docs]
def disable_backend(self, endpoint: str, name: str, i: int) -> None:
"""Mark read backend `name` at index `i` as failed, from endpoint `endpoint`."""
if i not in self.active_readers:
return
statsd.statsd.increment(
MP_BACKEND_DISABLED_METRICS,
1,
tags={
"endpoint": endpoint,
"name": self.name,
"backend": name,
"backend_number": i,
},
)
self.active_readers.remove(i)
if i not in self.reset_timers:
reset_failed_timer = threading.Timer(
self.read_exception_cooldown,
self.enable_backend,
kwargs={"name": name, "i": i},
)
reset_failed_timer.start()
self.reset_timers[i] = reset_failed_timer
if not self.active_readers:
raise NoBackendsLeftError(
"All backends disabled due to transient failures!"
)
[docs]
def enable_backend(self, name: str, i: int):
"""Mark a reader as available again"""
statsd.statsd.increment(
MP_BACKEND_ENABLED_METRICS,
1,
tags={
"name": self.name,
"backend": name,
"backend_number": i,
},
)
self.active_readers.add(i)
if i in self.reset_timers:
del self.reset_timers[i]
[docs]
def wrap_call(self, threads, call, *args, **kwargs):
threads = list(threads)
mailbox = queue.Queue()
for thread in threads:
thread.queue_command(call, *args, mailbox=mailbox, **kwargs)
return ObjStorageThread.collect_results(mailbox, len(threads))
[docs]
def get_read_threads(self, obj_id=None):
yield from self.storage_threads
[docs]
def get_write_threads(self, obj_id=None):
yield from self.write_storage_threads
[docs]
def check_config(self, *, check_write):
"""Check whether the object storage is properly configured.
If check_write is True, return True if at least one object storage
returned True.
Args:
check_write (bool): if True, check if writes to the object storage
can succeed.
Returns:
True if the configuration check worked, an exception if it didn't.
"""
if not check_write:
return all(
self.wrap_call(
self.storage_threads, "check_config", check_write=check_write
)
)
else:
return any(
self.wrap_call(
self.storage_threads, "check_config", check_write=check_write
)
)
@timed
def __contains__(self, obj_id: ObjId) -> bool:
"""Indicate if the given object is present in the storage.
Args:
obj_id (bytes): object identifier.
Returns:
True if and only if the object is present in the current object
storage.
"""
for storage in self.get_read_threads(obj_id):
if obj_id in storage:
return True
return False
def __iter__(self) -> Iterator[CompositeObjId]:
def obj_iterator():
for i, storage in enumerate(self.storages):
if i in self.active_readers:
yield from storage
return obj_iterator()
[docs]
@timed
def add(self, content: bytes, obj_id: ObjId, check_presence: bool = True) -> None:
"""Add a new object to the object storage.
If the adding step works in all the storages that accept this content,
this is a success. Otherwise, the full adding step is an error even if
it succeed in some of the storages.
Args:
content: content of the object to be added to the storage.
obj_id: checksum of [bytes] using [ID_HASH_ALGO] algorithm. When
given, obj_id will be trusted to match the bytes. If missing,
obj_id will be computed on the fly.
check_presence: indicate if the presence of the content should be
verified before adding the file.
Returns:
an id of the object into the storage. As the write-storages are
always readable as well, any id will be valid to retrieve a
content.
"""
if self.readonly:
raise ReadOnlyObjStorageError("add")
# note: we do not have per-backend statsd metrics here because the
# threading scaffolding to manage IO with backends makes it a bit
# harder to do in a nice manner; plus metrics should be available in
# the backend objstorages themselves.
self.wrap_call(
self.get_write_threads(obj_id),
"add",
content,
obj_id=obj_id,
check_presence=check_presence,
)
[docs]
def add_batch(
self,
contents: Union[Mapping[Sha1, bytes], Iterable[Tuple[ObjId, bytes]]],
check_presence: bool = True,
) -> Dict:
"""Add a batch of new objects to the object storage."""
if self.readonly:
raise ReadOnlyObjStorageError("add_batch")
write_threads = list(self.get_write_threads())
results = self.wrap_call(
write_threads,
"add_batch",
contents,
check_presence=check_presence,
)
summed = {"object:add": 0, "object:add:bytes": 0}
for result in results:
summed["object:add"] += result["object:add"]
summed["object:add:bytes"] += result["object:add:bytes"]
return {
"object:add": summed["object:add"] // len(results),
"object:add:bytes": summed["object:add:bytes"] // len(results),
}
[docs]
def restore(self, content: bytes, obj_id: ObjId) -> None:
if self.readonly:
raise ReadOnlyObjStorageError("restore")
return self.wrap_call(
self.get_write_threads(obj_id),
"restore",
content,
obj_id=obj_id,
).pop()
[docs]
@timed
def get(self, obj_id: ObjId) -> bytes:
corrupted_exc: Optional[ObjCorruptedError] = None
for i, storage in enumerate(self.get_read_threads(obj_id)):
if i not in self.active_readers:
continue
try:
obj = storage.get(obj_id)
statsd.statsd.increment(
MP_COUNTER_METRICS,
1,
tags={
"endpoint": "get",
"name": self.name,
"backend": storage.storage.name,
"backend_number": i,
},
)
return obj
except ObjNotFoundError:
continue
except ObjCorruptedError as exc:
logger.warning(
"Object %s was reported as corrupted by backend '%s': %s",
format_obj_id(obj_id),
storage.storage.name,
str(exc),
)
# Hoist exception, mypy doesn't like when we directly use
# `except ObjCorruptedError as corrupted_exc`
corrupted_exc = exc
# Try reading from another storage
continue
except Exception as exc:
if isinstance(exc, ObjStorageAPIError):
exc = exc.args[0]
exc_class = f"{exc.__class__.__module__}.{exc.__class__.__name__}"
if exc_class in self.transient_read_exceptions:
logger.warning(
"While reading object %s, received transient read "
"exception on backend %s, marking backend as failed",
format_obj_id(obj_id),
storage.storage.name,
exc_info=True,
)
self.disable_backend(endpoint="get", name=storage.storage.name, i=i)
continue
raise
if corrupted_exc:
# The only objects we've found were corrupted, raise that exception
raise corrupted_exc
else:
# No storage contains this content, raise the error
raise ObjNotFoundError(obj_id)
[docs]
def check(self, obj_id: ObjId) -> None:
nb_present = 0
nb_corrupted = 0
exception: Optional[Exception] = None
for storage in self.get_read_threads(obj_id):
try:
storage.check(obj_id)
except ObjNotFoundError as e:
exception = e
continue
except ObjCorruptedError as e:
exception = e
nb_corrupted += 1
else:
nb_present += 1
# Raise exception only if the content could not be found in all the storages
# or is corrupted in all the storages
if exception and (nb_present == 0 or nb_corrupted == len(self.storages)):
raise exception
[docs]
def delete(self, obj_id: ObjId):
if self.readonly:
raise ReadOnlyObjStorageError("delete")
super().delete(obj_id) # Check delete permission
return all(self.wrap_call(self.get_write_threads(obj_id), "delete", obj_id))
