Skip to content

DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.

3.2 KB
Raw Blame
from __future__ import annotations

import pytest

from deeptutor.services.keypool import KeyPool, primary_api_key


def test_keypool_rotates_in_round_robin_order() -> None:
    pool = KeyPool(["key-a", "key-b", "key-c"])

    assert len(pool) == 3
    assert [pool.next() for _ in range(5)] == [
        "key-a",
        "key-b",
        "key-c",
        "key-a",
        "key-b",
    ]


def test_keypool_cools_key_after_two_429s_and_restores_it(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    from deeptutor.services import keypool as keypool_module

    now = {"value": 100.0}
    monkeypatch.setattr(keypool_module, "monotonic", lambda: now["value"])
    pool = KeyPool(["key-a", "key-b"], cooldown_s=60)

    assert pool.next() == "key-a"
    pool.mark_429("key-a")
    pool.mark_429("key-a")
    assert [pool.next(), pool.next()] == ["key-b", "key-b"]

    now["value"] = 161.0
    assert pool.next() == "key-a"


def test_keypool_still_serves_a_single_cooling_key(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """A single-key setup must never be refused service.

    Cooling is only useful as a hint about which key to prefer. With one key
    there is nothing to prefer, so refusing to hand it out would turn a
    retryable provider 429 into a hard failure for every LLM and embedding
    call until the cooldown expires.
    """
    from deeptutor.services import keypool as keypool_module

    now = {"value": 10.0}
    monkeypatch.setattr(keypool_module, "monotonic", lambda: now["value"])
    pool = KeyPool(["only-key"], cooldown_s=5)

    assert pool.next() == "only-key"
    pool.mark_429("only-key")
    assert pool.next() == "only-key"
    pool.mark_429("only-key")
    assert pool.next() == "only-key"

    now["value"] = 16.0
    assert pool.next() == "only-key"


def test_keypool_prefers_the_soonest_recovering_key_when_all_are_cooling(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    from deeptutor.services import keypool as keypool_module

    now = {"value": 0.0}
    monkeypatch.setattr(keypool_module, "monotonic", lambda: now["value"])
    pool = KeyPool(["key-a", "key-b"], cooldown_s=60)

    # key-a cools at t=0 (until 60), key-b at t=10 (until 70).
    pool.mark_429("key-a")
    pool.mark_429("key-a")
    now["value"] = 10.0
    pool.mark_429("key-b")
    pool.mark_429("key-b")

    now["value"] = 20.0
    assert pool.next() == "key-a"


@pytest.mark.parametrize(
    ("value", "expected"),
    [
        ("sk-a", "sk-a"),
        (["sk-a"], "sk-a"),
        (["sk-a", "sk-b"], "sk-a"),
        # An empty key is not a key, in either spelling — so a caller can test
        # the result instead of knowing which shape the config happened to use.
        ("", None),
        (None, None),
        ([], None),
        ([""], None),
    ],
)
def test_primary_api_key_reduces_every_configured_shape(value, expected) -> None:
    assert primary_api_key(value) == expected


def test_llm_config_get_api_key_uses_the_same_reduction() -> None:
    """``LLMConfig.get_api_key`` is a ``-> str`` front door, not a second copy."""
    from deeptutor.services.llm.config import LLMConfig

    for value in ("sk-a", ["sk-a", "sk-b"], "", [], [""]):
        config = LLMConfig(api_key=value, model="gpt-5")
        assert config.get_api_key() == (primary_api_key(value) or "")