mirror of
https://github.com/Evil0ctal/Douyin_TikTok_Download_API.git
synced 2026-10-29 12:59:49 +08:00
Evil0ctal
e72bc75b42
Three changes, all found by measuring rather than reasoning. Rotation was not rotating. The tie-break seed came from the retry count, so every concurrent first attempt ranked the candidate list identically and callers walked it in lockstep; and recency was compared exactly, so no two candidates ever tied and the tie-break was never consulted at all. Over 600 concurrent requests across 12 identities that gave a 26..74 split. With a per-call seed and recency quantized to half a second the median spread is 40% against a uniform random baseline of 46% - the ordering now earns its place by beating chance, which is the only honest yardstick once perfect round-robin is unreachable. The lock now uses SET NX instead of EXISTS-then-SET. Both are safe inside a Lua script, but NX states the requirement in one indivisible operation rather than two statements that happen to add up to it. The no-token path releases the lock it just took, which the two-statement form did not need. The policy table is process-global mutable state and tests that registered an endpoint left it there, where the next test read it as configuration for an endpoint that does not exist. snapshot/restore plus an autouse fixture. Honest note on an unresolved observation: while three other workloads were saturating this machine, the exclusivity assertion intermittently reported more leases than identities. Redis is single-threaded and SET NX is indivisible, so that should be impossible without the key disappearing - yet expired_keys and evicted_keys stay at zero across a failing run, and it does not reproduce on an idle machine or outside pytest. Four separate probes were built to chase it and three of them turned out to have bugs of their own (in-process ordering mistaken for server ordering, an off-by-one in a Lua key slice, a single-sample threshold on a wide distribution), which is most of why this took as long as it did. The test is left asserting the real invariant rather than relaxed into something that always passes.
23 lines
600 B
Python
23 lines
600 B
Python
"""Shared fixtures.
|
|
|
|
The scheduler's policy table is process-global mutable state. A test that
|
|
registers an endpoint leaves it there, where the next test sees it as
|
|
configuration for an endpoint that does not exist. Snapshotting around every
|
|
test keeps that from turning into a false failure somewhere else.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
from dtk.scheduler.policies import restore_policies, snapshot_policies
|
|
|
|
|
|
@pytest.fixture(autouse=True)
|
|
def _isolated_policy_registry():
|
|
saved = snapshot_policies()
|
|
try:
|
|
yield
|
|
finally:
|
|
restore_policies(saved)
|