Skip to content

The fastest, litest AI Gateway. Rust core with Python SDK. Call 100+ LLM APIs in OpenAI (or native) format with cost tracking, guardrails, load balancing, and logging [Bedrock, Azure, OpenAI, Anthropic, OpenAI, VertexAI, vLLM, Nvidia NIM]

11.4 KB
Raw Blame
from __future__ import annotations

import json
import uuid
from hashlib import sha256
from collections.abc import Iterator, Sequence
from dataclasses import dataclass
from pathlib import Path
from typing import Final

import httpx
import pytest
import yaml
from litellm.constants import LITELLM_PROXY_MASTER_KEY_ALIAS
from litellm.types.integrations.prometheus import LATENCY_BUCKETS

from tests.integration._support.client import Gateway, eventually, gateway_from_environment, object_value
from tests.integration._support.process import owned_proxy
from tests.integration._support.prometheus_series import Sample, label_values, scrape
from tests.integration._support.wire import Reply, Request, Wire, wire_server

_GOOD: Final = "prometheus-good-endpoint"
_LIMITED: Final = "prometheus-rate-limited-endpoint"
_FAILING: Final = "prometheus-failing-endpoint"
_LATENCY: Final = "prometheus-latency-endpoint"
_END_USER: Final = f"prometheus-end-user-{uuid.uuid4().hex}"


@dataclass(frozen=True, slots=True)
class _Rig:
    gateway: Gateway
    upstream: Wire


def _respond(request: Request) -> Reply:
    if json.loads(request.body)["model"] == "429":
        return Reply(
            status=429,
            body=json.dumps({"error": {"message": "rate limited", "type": "rate_limit_error", "code": "429"}}).encode(),
        )
    return Reply(
        body=json.dumps(
            {
                "id": f"chatcmpl-{uuid.uuid4().hex}",
                "object": "chat.completion",
                "created": 1,
                "model": "gpt-4o-mini",
                "choices": [
                    {"index": 0, "message": {"role": "assistant", "content": "metered"}, "finish_reason": "stop"}
                ],
                "usage": {"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4},
            }
        ).encode()
    )


def _config(directory: Path, upstream: Wire) -> Path:
    configuration: Final = yaml.safe_load(Path("tests/integration/proxy_config.yaml").read_text())
    params: Final = {"api_key": "sk-fixture", "api_base": f"{upstream.url}/v1"}
    configuration["model_list"] = [
        *configuration["model_list"],
        *(
            {
                "model_name": name,
                "litellm_params": {
                    "model": "openai/gpt-4o-mini",
                    "input_cost_per_token": 0.001,
                    "output_cost_per_token": 0.002,
                    **params,
                },
            }
            for name in (_GOOD, _LATENCY)
        ),
        {"model_name": _LIMITED, "litellm_params": {"model": "openai/429", **params}},
        {"model_name": _FAILING, "litellm_params": {"model": "openai/429", **params}},
    ]
    configuration["litellm_settings"]["callbacks"] = ["prometheus"]
    configuration["litellm_settings"]["disable_end_user_cost_tracking_prometheus_only"] = True
    configuration.setdefault("router_settings", {})["num_retries"] = 0
    path: Final = directory / "prometheus-request-metrics.yaml"
    path.write_text(yaml.safe_dump(configuration))
    return path


@pytest.fixture(scope="module")
def rig(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Rig]:
    directory: Final = tmp_path_factory.mktemp("prometheus-request-metrics")
    with (
        wire_server(_respond) as upstream,
        gateway_from_environment() as shared,
        owned_proxy(shared, directory, {}, config=_config(directory, upstream)) as owned,
    ):
        yield _Rig(owned, upstream)


def _ask(rig: _Rig, model: str, key: str | None = None, **extra: list[str] | str) -> httpx.Response:
    return rig.gateway.request(
        "POST",
        "/chat/completions",
        {"model": model, "messages": [{"role": "user", "content": f"metrics {uuid.uuid4().hex}"}], **extra},
        key=key,
    )


def _series(samples: Sequence[Sample], name: str, **labels: str) -> tuple[Sample, ...]:
    return tuple(
        sample
        for sample in samples
        if sample.name == name and all(sample.labels.get(label) == value for label, value in labels.items())
    )


def _until(rig: _Rig, name: str, **labels: str) -> tuple[Sample, ...]:
    return eventually(lambda: _series(scrape(rig.gateway), name, **labels), bool, seconds=30)


def test_a_rate_limited_call_counts_as_a_failed_and_a_429_total_request(rig: _Rig) -> None:
    response: Final = _ask(rig, _FAILING)
    assert response.status_code == 429, response.text
    assert len(rig.upstream.drain()) == 1
    failed: Final = _until(
        rig,
        "litellm_proxy_failed_requests_metric_total",
        api_key_alias="None",
        exception_class="Openai.RateLimitError",
        exception_status="429",
        hashed_api_key=LITELLM_PROXY_MASTER_KEY_ALIAS,
        requested_model=_FAILING,
        route="/chat/completions",
    )
    assert [sample.value for sample in failed] == [1.0]
    totals: Final = _until(
        rig,
        "litellm_proxy_total_requests_metric_total",
        hashed_api_key=LITELLM_PROXY_MASTER_KEY_ALIAS,
        requested_model=_FAILING,
        status_code="429",
    )
    assert [sample.value for sample in totals] == [1.0]


def test_a_good_call_exports_latency_histograms_on_the_shared_buckets_without_the_end_user(rig: _Rig) -> None:
    response: Final = _ask(rig, _LATENCY, user=_END_USER, tags=["teamB"])
    assert response.status_code == 200, response.text
    assert len(rig.upstream.drain()) == 1
    master: Final = {
        "api_key_alias": "None",
        "hashed_api_key": LITELLM_PROXY_MASTER_KEY_ALIAS,
        "requested_model": _LATENCY,
    }
    _until(rig, "litellm_request_total_latency_metric_bucket", le="0.005", **master)
    _until(rig, "litellm_llm_api_latency_metric_bucket", le="0.005", **master)
    for name in ("litellm_request_total_latency_metric_count", "litellm_llm_api_latency_metric_count"):
        assert [sample.value for sample in _until(rig, name, **master)] == [1.0], name
    samples: Final = scrape(rig.gateway)
    assert _END_USER not in label_values(samples)
    expected: Final = {str(bucket).replace("inf", "+Inf") for bucket in LATENCY_BUCKETS}
    for name in (
        "litellm_request_total_latency_metric_bucket",
        "litellm_llm_api_latency_metric_bucket",
        "litellm_overhead_latency_metric_bucket",
    ):
        assert {sample.labels["le"] for sample in _series(samples, name)} == expected, name


def test_client_side_fallbacks_count_one_success_and_one_failure(rig: _Rig) -> None:
    recovered: Final = _ask(rig, _LIMITED, fallbacks=[_GOOD])
    assert recovered.status_code == 200, recovered.text
    missing: Final = f"unknown-model-{uuid.uuid4().hex[:8]}"
    failed: Final = _ask(rig, _LIMITED, fallbacks=[missing])
    assert failed.status_code >= 400, failed.text
    rig.upstream.drain()
    shared: Final = {
        "api_key_alias": "None",
        "exception_class": "Openai.RateLimitError",
        "exception_status": "429",
        "hashed_api_key": LITELLM_PROXY_MASTER_KEY_ALIAS,
        "requested_model": _LIMITED,
    }
    succeeded: Final = _until(rig, "litellm_deployment_successful_fallbacks_total", fallback_model=_GOOD, **shared)
    assert [sample.value for sample in succeeded] == [1.0]
    lost: Final = _until(rig, "litellm_deployment_failed_fallbacks_total", fallback_model=missing, **shared)
    assert [sample.value for sample in lost] == [1.0]


@dataclass(frozen=True, slots=True)
class _Budget:
    remaining: float
    total: float
    hours: float


def _budget(samples: Sequence[Sample], scope: str, label: str, identity: str) -> _Budget | None:
    remaining: Final = _series(samples, f"litellm_remaining_{scope}_budget_metric", **{label: identity})
    total: Final = _series(samples, f"litellm_{scope}_max_budget_metric", **{label: identity})
    hours: Final = _series(samples, f"litellm_{scope}_budget_remaining_hours_metric", **{label: identity})
    if len(remaining) != 1 or len(total) != 1 or len(hours) != 1:
        return None
    return _Budget(remaining[0].value, total[0].value, hours[0].value)


def _reconciled(rig: _Rig, scope: str, label: str, identity: str, info: str, field: str, lookup: str) -> _Budget:
    def read() -> tuple[_Budget | None, float]:
        record: Final = object_value(rig.gateway.get(info, {field: lookup})[_INFO_FIELDS[info]])
        return _budget(scrape(rig.gateway), scope, label, identity), float(str(record["max_budget"])) - float(
            str(record["spend"])
        )

    budget, remaining = eventually(
        read,
        lambda state: (
            state[0] is not None and state[0].remaining < 10.0 and abs(state[1] - state[0].remaining) <= 0.001
        ),
        seconds=60,
    )
    assert budget is not None
    assert abs(remaining - budget.remaining) <= 0.001
    return budget


_INFO_FIELDS: Final = {"/team/info": "team_info", "/key/info": "info", "/user/info": "user_info"}


def test_a_team_call_exports_remaining_max_and_hours_gauges_matching_team_info(rig: _Rig) -> None:
    with rig.gateway.scenario() as scenario:
        team: Final = scenario.team(max_budget=10, budget_duration="7d")
        key: Final = scenario.key(team_id=team)
        assert _ask(rig, _GOOD, key).status_code == 200
        assert len(rig.upstream.drain()) == 1
        budget: Final = _reconciled(rig, "team", "team", team, "/team/info", "team_id", team)
        assert budget.total == 10.0
        assert 0 < budget.hours <= 168


def test_a_key_call_exports_remaining_max_and_hours_gauges_matching_key_info(rig: _Rig) -> None:
    with rig.gateway.scenario() as scenario:
        key: Final = scenario.key(max_budget=10, budget_duration="7d")
        assert _ask(rig, _GOOD, key).status_code == 200
        assert len(rig.upstream.drain()) == 1
        hashed: Final = sha256(key.encode()).hexdigest()
        budget: Final = _reconciled(rig, "api_key", "hashed_api_key", hashed, "/key/info", "key", key)
        assert budget.total == 10.0
        assert 0 <= budget.hours <= 168


def test_a_user_call_exports_remaining_max_and_hours_gauges_matching_user_info(rig: _Rig) -> None:
    with rig.gateway.scenario() as scenario:
        user: Final = f"prometheus-user-{uuid.uuid4().hex}"
        scenario.user(user_id=user, max_budget=10, budget_duration="7d")
        key: Final = scenario.key(user_id=user)
        assert _ask(rig, _GOOD, key).status_code == 200
        assert len(rig.upstream.drain()) == 1
        budget: Final = _reconciled(rig, "user", "user", user, "/user/info", "user_id", user)
        assert budget.total == 10.0
        assert 0 <= budget.hours <= 168


def test_a_user_email_labels_the_spend_and_failed_request_series_of_its_keys(rig: _Rig) -> None:
    with rig.gateway.scenario() as scenario:
        email: Final = f"prometheus-{uuid.uuid4().hex}@example.com"
        user: Final = f"prometheus-email-{uuid.uuid4().hex}"
        scenario.user(user_id=user, user_email=email)
        key: Final = scenario.key(user_id=user)
        assert _ask(rig, _GOOD, key).status_code == 200
        assert len(rig.upstream.drain()) == 1
        spend: Final = _until(rig, "litellm_spend_metric_total", user_email=email)
        assert [(sample.labels["user"], sample.value) for sample in spend] == [(user, pytest.approx(0.005))]
        assert email in label_values(scrape(rig.gateway))
        assert _ask(rig, _FAILING, key).status_code == 429
        assert len(rig.upstream.drain()) == 1
        failed: Final = _until(
            rig, "litellm_proxy_failed_requests_metric_total", user_email=email, requested_model=_FAILING
        )
        assert [(sample.labels["user"], sample.value) for sample in failed] == [(user, 1.0)]