mirror of
https://github.com/MISP/misp-galaxy.git
synced 2024-11-30 02:37:17 +00:00
67 lines
2.2 KiB
Python
67 lines
2.2 KiB
Python
from asyncio import sleep
|
|
from typing import TYPE_CHECKING, Any, Awaitable, Callable, Tuple, Type, TypeVar
|
|
|
|
from redis.exceptions import ConnectionError, RedisError, TimeoutError
|
|
|
|
if TYPE_CHECKING:
|
|
from redis.backoff import AbstractBackoff
|
|
|
|
|
|
T = TypeVar("T")
|
|
|
|
|
|
class Retry:
|
|
"""Retry a specific number of times after a failure"""
|
|
|
|
__slots__ = "_backoff", "_retries", "_supported_errors"
|
|
|
|
def __init__(
|
|
self,
|
|
backoff: "AbstractBackoff",
|
|
retries: int,
|
|
supported_errors: Tuple[Type[RedisError], ...] = (
|
|
ConnectionError,
|
|
TimeoutError,
|
|
),
|
|
):
|
|
"""
|
|
Initialize a `Retry` object with a `Backoff` object
|
|
that retries a maximum of `retries` times.
|
|
`retries` can be negative to retry forever.
|
|
You can specify the types of supported errors which trigger
|
|
a retry with the `supported_errors` parameter.
|
|
"""
|
|
self._backoff = backoff
|
|
self._retries = retries
|
|
self._supported_errors = supported_errors
|
|
|
|
def update_supported_errors(self, specified_errors: list):
|
|
"""
|
|
Updates the supported errors with the specified error types
|
|
"""
|
|
self._supported_errors = tuple(
|
|
set(self._supported_errors + tuple(specified_errors))
|
|
)
|
|
|
|
async def call_with_retry(
|
|
self, do: Callable[[], Awaitable[T]], fail: Callable[[RedisError], Any]
|
|
) -> T:
|
|
"""
|
|
Execute an operation that might fail and returns its result, or
|
|
raise the exception that was thrown depending on the `Backoff` object.
|
|
`do`: the operation to call. Expects no argument.
|
|
`fail`: the failure handler, expects the last error that was thrown
|
|
"""
|
|
self._backoff.reset()
|
|
failures = 0
|
|
while True:
|
|
try:
|
|
return await do()
|
|
except self._supported_errors as error:
|
|
failures += 1
|
|
await fail(error)
|
|
if self._retries >= 0 and failures > self._retries:
|
|
raise error
|
|
backoff = self._backoff.compute(failures)
|
|
if backoff > 0:
|
|
await sleep(backoff)
|