FastAPI Backend Development Bootcamp · 课时

用量计量、配额与计费钩子

跟踪每个租户的用量，执行套餐配额，并为按用量计费的 SaaS 定价发布计费事件。

用量计量、配额与计费钩子 是 CoddyKit 上的免费 FastAPI Backend Development Bootcamp 课时。这是第 4 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 FastAPI Backend Development Bootcamp 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 FastAPI Backend Development Bootcamp 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Why Metering Matters in SaaS

In a multi-tenant SaaS, every tenant shares the same FastAPI app but pays based on what they actually use. To support plans like Free / Pro / Enterprise you need three cooperating subsystems:

Metering — count consumed resources (API calls, tokens, rows, GB).
Quotas — block or throttle once a tenant exceeds its plan limit.
Billing hooks — emit events that downstream billing (Stripe, etc.) turns into invoices.

The hard part is doing this per tenant, accurately, and without adding latency to every request. We will build each piece in this lesson.

Modeling Plans and Quotas

Start by modeling what each plan allows. A plan maps a metric (what we count) to a limit and a period (when the counter resets). Keeping this as plain data makes it easy to load from a DB or config.

Below, a metered plan tracks API requests and AI tokens with monthly limits. None means unlimited.

from dataclasses import dataclass

@dataclass(frozen=True)
class Quota:
    metric: str
    limit: int | None   # None = unlimited
    period: str          # 'day' or 'month'

PLANS = {
    "free": [Quota("api_calls", 1_000, "month"), Quota("ai_tokens", 50_000, "month")],
    "pro": [Quota("api_calls", 100_000, "month"), Quota("ai_tokens", 5_000_000, "month")],
    "enterprise": [Quota("api_calls", None, "month"), Quota("ai_tokens", None, "month")],
}

def limit_for(plan: str, metric: str) -> int | None:
    for q in PLANS[plan]:
        if q.metric == metric:
            return q.limit
    return 0  # metric not allowed on this plan

print(limit_for("free", "api_calls"))   # 1000
print(limit_for("enterprise", "ai_tokens"))  # None (unlimited)
print(limit_for("free", "unknown"))     # 0

An Atomic Usage Counter

The core of metering is a counter you can increment atomically per (tenant, metric, period). In production this is usually Redis (INCRBY + EXPIRE) or a Postgres UPSERT with an atomic +=.

Here is the concept modeled in pure Python so you can see the key and the increment logic. The period key (e.g. 2026-06) is what makes the counter reset every month automatically.

from collections import defaultdict
from datetime import datetime, timezone

usage = defaultdict(int)

def period_key(period: str, now: datetime) -> str:
    if period == "month":
        return now.strftime("%Y-%m")
    return now.strftime("%Y-%m-%d")

def incr(tenant: str, metric: str, amount: int, period: str, now: datetime) -> int:
    key = (tenant, metric, period_key(period, now))
    usage[key] += amount
    return usage[key]

now = datetime(2026, 6, 10, tzinfo=timezone.utc)
print(incr("acme", "api_calls", 1, "month", now))  # 1
print(incr("acme", "api_calls", 1, "month", now))  # 2
print(incr("acme", "ai_tokens", 1200, "month", now))  # 1200

Enforcing a Quota

Quota enforcement combines the plan limit with the current counter. The decision is simple but the order matters: you usually want to check-then-increment so that a request that would exceed the limit is rejected before doing the work.

For atomicity in real systems you increment first, then compare against the limit, and if you went over, you reject and optionally decrement (or simply let the counter sit over-limit, blocking further calls).

def check_quota(current: int, amount: int, limit: int | None) -> bool:
    """Return True if `amount` more units are allowed."""
    if limit is None:
        return True  # unlimited
    return current + amount <= limit

print(check_quota(995, 1, 1000))    # True  (996 <= 1000)
print(check_quota(1000, 1, 1000))   # False (1001 > 1000)
print(check_quota(9_000_000, 1, None))  # True (unlimited)

Wiring It Into FastAPI with a Dependency

In FastAPI the clean place to enforce quotas is a dependency. It resolves the tenant (from auth/JWT), looks up the plan, checks the quota, and raises HTTPException(429) when the limit is hit.

Returning 402 Payment Required is also common for hard plan caps, while 429 Too Many Requests fits rate-style limits. This is framework code, so it is not standalone-runnable.

from fastapi import Depends, HTTPException, Request

async def enforce_quota(metric: str, cost: int = 1):
    async def _dep(request: Request):
        tenant = request.state.tenant_id
        plan = await get_plan(tenant)
        limit = limit_for(plan, metric)
        current = await store.get(tenant, metric, "month")
        if not check_quota(current, cost, limit):
            raise HTTPException(
                status_code=429,
                detail=f"Quota exceeded for {metric}",
                headers={"X-Quota-Limit": str(limit)},
            )
        await store.incr(tenant, metric, cost, "month")
        return current + cost
    return _dep

@app.post("/v1/complete")
async def complete(used=Depends(enforce_quota("ai_tokens", cost=500))):
    return {"ok": True, "tokens_used_this_month": used}

Metering AFTER the Work for Variable Costs

Some costs are not known up front. AI token usage, bytes processed, or rows returned are only known after the handler runs. For these, do a cheap pre-check (is the tenant already over?) and record the real cost afterward.

A reliable pattern is a context manager that admits the request, runs the body, then reports actual usage — even on exceptions, so you don't lose metering data.

from contextlib import contextmanager

@contextmanager
def meter(record, tenant, metric):
    actual = {"amount": 0}
    try:
        yield actual            # caller sets actual['amount']
    finally:
        if actual["amount"]:
            record(tenant, metric, actual["amount"])

log = []
with meter(lambda t, m, a: log.append((t, m, a)), "acme", "ai_tokens") as use:
    # ... do real work, then we learn the true cost ...
    use["amount"] = 1342

print(log)  # [('acme', 'ai_tokens', 1342)]

Soft Limits, Hard Limits and Overage

Real billing rarely flips from 'works' to '403' at exactly the limit. Plans usually define:

Soft limit — allow but warn (email, X-Quota-Warning header) around ~80%.
Hard limit — block once exceeded (Free plans).
Overage — keep serving but bill each extra unit at a metered rate (Pro/Enterprise).

Encode this as a policy so the enforcement code stays declarative.

def classify(current: int, limit: int | None, overage_allowed: bool) -> str:
    if limit is None:
        return "ok"
    if current < 0.8 * limit:
        return "ok"
    if current < limit:
        return "warn"
    return "overage" if overage_allowed else "blocked"

print(classify(700, 1000, False))   # ok
print(classify(850, 1000, False))   # warn
print(classify(1200, 1000, True))   # overage
print(classify(1200, 1000, False))  # blocked

Emitting Billing Events

Metering produces a stream of usage events. Billing systems (Stripe Billing Meters, Orb, Metronome) consume these to compute invoices. The golden rules for billing events:

Include an idempotency key so retries don't double-charge.
Carry a timestamp so events land in the correct billing period.
Make them immutable — append-only, never edit.

Build the event payload as a plain, serializable record.

import uuid, json
from datetime import datetime, timezone

def billing_event(tenant: str, metric: str, qty: int, ts: datetime, request_id: str) -> dict:
    return {
        "id": str(uuid.uuid4()),
        "idempotency_key": f"{tenant}:{metric}:{request_id}",
        "tenant_id": tenant,
        "metric": metric,
        "quantity": qty,
        "occurred_at": ts.isoformat(),
    }

evt = billing_event("acme", "ai_tokens", 1342,
                    datetime(2026, 6, 10, tzinfo=timezone.utc), "req_9f3")
print(json.dumps(evt, indent=2))

Decoupling with a Background Queue

Never call your billing provider synchronously inside the request path — a slow or down Stripe API would break your API. Instead, write the usage event to a durable buffer and flush it asynchronously.

FastAPI's BackgroundTasks works for low volume; for real scale use an outbox table, Redis stream, or a worker (Celery/ARQ). The handler stays fast; a separate consumer ships events to billing with retries and idempotency.

from fastapi import BackgroundTasks

@app.post("/v1/complete")
async def complete(bg: BackgroundTasks, request: Request):
    tenant = request.state.tenant_id
    result, tokens = await run_completion(request)
    await store.incr(tenant, "ai_tokens", tokens, "month")
    evt = billing_event(tenant, "ai_tokens", tokens,
                        datetime.now(timezone.utc), request.state.request_id)
    bg.add_task(emit_to_billing, evt)  # fire-and-forget, off the hot path
    return result

Reconciliation: Trust but Verify

Counters drift. A crashed worker, a lost event, or a Redis flush can desync your metering from reality. Robust SaaS systems treat the append-only event log as the source of truth and periodically reconcile the fast counters against it.

A nightly job re-aggregates raw events per tenant/metric/period and corrects the cached counter. This also lets you regenerate billing if a provider rejected a batch.

from collections import defaultdict

def reconcile(events: list[dict]) -> dict:
    """Rebuild authoritative totals from the immutable event log."""
    totals = defaultdict(int)
    seen = set()
    for e in events:
        if e["idempotency_key"] in seen:
            continue  # dedupe replays
        seen.add(e["idempotency_key"])
        period = e["occurred_at"][:7]  # YYYY-MM
        totals[(e["tenant_id"], e["metric"], period)] += e["quantity"]
    return dict(totals)

events = [
    {"idempotency_key": "acme:t:1", "tenant_id": "acme", "metric": "t", "quantity": 100, "occurred_at": "2026-06-01T10:00:00"},
    {"idempotency_key": "acme:t:1", "tenant_id": "acme", "metric": "t", "quantity": 100, "occurred_at": "2026-06-01T10:00:00"},
    {"idempotency_key": "acme:t:2", "tenant_id": "acme", "metric": "t", "quantity": 50, "occurred_at": "2026-06-02T09:00:00"},
]
print(reconcile(events))  # {('acme','t','2026-06'): 150}

Tenant Isolation and Hot Keys

A few operational rules keep per-tenant metering correct and fast:

Always scope the counter key by tenant_id — never share a global counter, or one noisy tenant blocks others.
Watch for hot keys: a huge tenant hammering one Redis key can become a bottleneck; shard the key (tenant:metric:shard) and sum on read.
Make limit checks fail-open or fail-closed deliberately — if the counter store is down, decide whether to allow (availability) or block (revenue protection).
Reset boundaries must respect the tenant's billing-cycle timezone, not just UTC month.

Quick Check

Test your understanding of the metering decision flow.

Recap

You built a complete per-tenant metering and billing pipeline:

Plans & quotas as data: metric to limit to reset period, with None for unlimited.
Atomic counters keyed by (tenant, metric, period) so limits reset automatically.
Enforcement via a FastAPI dependency returning 429/402; check-then-increment for fixed costs, meter-after for variable costs.
Policies for soft limits, hard limits, and overage instead of a single hard cutoff.
Billing events that are immutable, timestamped, and idempotent, shipped asynchronously off the request path.
Reconciliation from an append-only event log as the source of truth, plus tenant-isolation and hot-key care.

With these pieces you can support usage-based pricing safely without slowing down your API.

常见问题解答

「用量计量、配额与计费钩子」课时是免费的吗？

是的 — 「用量计量、配额与计费钩子」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 FastAPI Backend Development Bootcamp 课程的其余内容，请升级到 CoddyKit PRO。 FastAPI Backend Development Bootcamp 课程共包含 4 节课。

「用量计量、配额与计费钩子」这节课中我会学到什么？

跟踪每个租户的用量，执行套餐配额，并为按用量计费的 SaaS 定价发布计费事件。你通过在浏览器中直接运行的动手代码来练习 FastAPI Backend Development Bootcamp，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 FastAPI Backend Development Bootcamp 需要有经验吗？

无需任何先前经验。CoddyKit 上的 FastAPI Backend Development Bootcamp 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 4 节课，共 4 节。

「用量计量、配额与计费钩子」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 FastAPI Backend Development Bootcamp 课中编写并运行代码吗？

能。每节 FastAPI Backend Development Bootcamp 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 FastAPI Backend Development Bootcamp