feat(sites): cross-process site rate limit via PG token bucket

Migration 0041_rate_limit adds rate_limit_bucket. PgTokenBucket (@umt/db)
refills and takes tokens under SELECT ... FOR UPDATE on the DB clock and
polls with jitter within a wait budget; penalize() carries Retry-After.
Site adapters get createSharedSiteRateLimiter (key site:<id>), shared by
api / worker / worker:agent, with an in-process serial queue in front and
a local fallback when the bucket is unavailable.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
This commit is contained in:
rootandClaude Opus 5.5 committed 2026-09-28 21:51:32 +08:00
1 parent 57346d438c
commit 61f585a707
12 files changed
+466 -5

No files matched your search

+30
View File
@@ -0,0 +1,30 @@
import type { Kysely } from "kysely";
import { dropTables, execDdl } from "../src/migration-helpers.ts";
// C2-A 跨进程站点限流:PG 令牌桶(api / worker / worker:agent 共用)。
// - key:`site:<站点 id>` 等;tokens 为上次结算时的令牌数(可为负,表示被 Retry-After 顺延);
// - capacity / refill_per_sec 由调用方每次扣令牌时带上(站点限流设置变化即时生效);
// - blocked_until:站点要求暂停(429 Retry-After)时在此之前不发放令牌。
// 扣令牌走 `select … for update`,补充按 DB 时钟(clock_timestamp)结算,多进程时钟不一致也不影响。
// biome-ignore lint/suspicious/noExplicitAny: 迁移不依赖应用表类型
export async function up(db: Kysely<any>): Promise<void> {
await execDdl(
db,
`
create table rate_limit_bucket (
key text primary key,
tokens double precision not null,
capacity double precision not null check (capacity > 0),
refill_per_sec double precision not null check (refill_per_sec > 0),
blocked_until timestamptz,
updated_at timestamptz not null default now()
);
`,
);
}
// biome-ignore lint/suspicious/noExplicitAny: 同上
export async function down(db: Kysely<any>): Promise<void> {
await execDdl(db, dropTables(["rate_limit_bucket"]));
}
+7
View File
@@ -2,4 +2,11 @@
export { type Kysely, sql, type Transaction } from "kysely";
export { type CreateDbOptions, createDb } from "./client.ts";
export { createMigrator, MIGRATIONS_DIR, type MigratorOptions, migrateToLatest, resetSchema } from "./migrator.ts";
export {
type AcquireOptions,
type AcquireResult,
type BucketSpec,
PgTokenBucket,
pgTokenBucketFor,
} from "./rate-limit.ts";
export type * from "./types.ts";
+127
View File
@@ -0,0 +1,127 @@
import { type Kysely, sql } from "kysely";
import type { DB } from "./types.ts";
/**
* 跨进程令牌桶(rate_limit_bucket,迁移 0041)。api / worker / worker:agent 三个进程对同一 key 共用一个桶:
* 每次扣令牌在事务里 `select … for update` 锁住该行,按 DB 时钟补充后扣减;拿不到则算出还差多久,
* 调用方在 waitMs 预算内睡到那时(带抖动)再试。进程内的突发由调用方的本地串行队列挡住(见 adapters 的站点限流器)。
*/
export interface BucketSpec {
/** 桶容量(突发上限),≥ 1 */
capacity: number;
/** 每秒补充的令牌数,> 0 */
refillPerSec: number;
}
export interface AcquireOptions extends BucketSpec {
/** 最多等多久(毫秒);0 = 只试一次。缺省 60s */
waitMs?: number;
signal?: AbortSignal;
}
export interface AcquireResult {
ok: boolean;
/** 本次累计等待(毫秒) */
waitedMs: number;
}
// biome-ignore lint/suspicious/noExplicitAny: 测试会传入隔离 schema 的实例
type AnyDb = Kysely<DB> | Kysely<any>;
function sleep(ms: number, signal?: AbortSignal): Promise<void> {
return new Promise((resolve, reject) => {
if (signal?.aborted) return reject(signal.reason);
const t = setTimeout(() => {
signal?.removeEventListener("abort", onAbort);
resolve();
}, ms);
const onAbort = () => {
clearTimeout(t);
reject(signal?.reason);
};
signal?.addEventListener("abort", onAbort, { once: true });
});
}
export class PgTokenBucket {
constructor(
private readonly db: AnyDb,
private readonly opts: { jitterMs?: number; random?: () => number } = {},
) {}
/**
* 试扣一次:成功返回 0,否则返回还需等待的毫秒数(被 penalize 顺延时为到期时间)。
* 行不存在时按满桶创建;容量 / 速率以本次传入为准(设置改了即时生效)。
*/
async tryTake(key: string, cost: number, spec: BucketSpec): Promise<number> {
const capacity = Math.max(1, spec.capacity);
const rate = Math.max(1e-6, spec.refillPerSec);
const need = Math.min(cost, capacity);
return await this.db.transaction().execute(async (trx) => {
await sql`insert into rate_limit_bucket (key, tokens, capacity, refill_per_sec)
values (${key}, ${capacity}, ${capacity}, ${rate}) on conflict (key) do nothing`.execute(trx);
const row = (
await sql<{ tokens: number; elapsed: number; blocked_ms: number | null }>`
select tokens,
greatest(0, extract(epoch from clock_timestamp() - updated_at))::float8 as elapsed,
(extract(epoch from blocked_until - clock_timestamp()) * 1000)::float8 as blocked_ms
from rate_limit_bucket where key = ${key} for update`.execute(trx)
).rows[0];
if (!row) return 1000;
let tokens = Math.min(capacity, Number(row.tokens) + Number(row.elapsed) * rate);
const blocked = row.blocked_ms !== null && Number(row.blocked_ms) > 0 ? Math.ceil(Number(row.blocked_ms)) : 0;
let wait = 0;
if (blocked > 0) wait = blocked;
else if (tokens >= need) tokens -= need;
else wait = Math.ceil(((need - tokens) / rate) * 1000);
await sql`update rate_limit_bucket set tokens = ${tokens}, capacity = ${capacity}, refill_per_sec = ${rate},
updated_at = clock_timestamp(),
blocked_until = case when blocked_until <= clock_timestamp() then null else blocked_until end
where key = ${key}`.execute(trx);
return wait;
});
}
/** 扣 cost 个令牌;waitMs 内拿不到返回 ok:false。signal 中止时抛出其 reason。 */
async acquire(key: string, cost: number, opts: AcquireOptions): Promise<AcquireResult> {
const budget = Math.max(0, opts.waitMs ?? 60_000);
const jitterMax = this.opts.jitterMs ?? 50;
const random = this.opts.random ?? Math.random;
const t0 = Date.now();
for (;;) {
opts.signal?.throwIfAborted();
const wait = await this.tryTake(key, cost, opts);
const spent = Date.now() - t0;
if (wait <= 0) return { ok: true, waitedMs: spent };
if (spent + wait > budget) return { ok: false, waitedMs: spent };
// 抖动:多个进程同时醒来时错开,避免一起撞锁再一起落空
const jitter = Math.floor(random() * Math.min(jitterMax, Math.max(1, wait * 0.2)));
await sleep(wait + jitter, opts.signal);
}
}
/** 站点要求暂停(Retry-After):ms 内不发放令牌,已有令牌清零 */
async penalize(key: string, ms: number, spec: BucketSpec): Promise<void> {
const capacity = Math.max(1, spec.capacity);
const rate = Math.max(1e-6, spec.refillPerSec);
const until = sql`clock_timestamp() + make_interval(secs => ${Math.max(0, ms) / 1000})`;
await sql`insert into rate_limit_bucket (key, tokens, capacity, refill_per_sec, blocked_until, updated_at)
values (${key}, 0, ${capacity}, ${rate}, ${until}, clock_timestamp())
on conflict (key) do update set tokens = least(rate_limit_bucket.tokens, 0),
blocked_until = greatest(coalesce(rate_limit_bucket.blocked_until, clock_timestamp()), ${until}),
updated_at = clock_timestamp()`.execute(this.db);
}
}
const buckets = new WeakMap<object, PgTokenBucket>();
/** 同一个 Kysely 实例共用一个 PgTokenBucket */
export function pgTokenBucketFor(db: AnyDb): PgTokenBucket {
let b = buckets.get(db);
if (!b) {
b = new PgTokenBucket(db);
buckets.set(db, b);
}
return b;
}
+13
View File
@@ -1156,6 +1156,15 @@ export interface ProcessHeartbeatSampleTable {
sse_connections: number | null;
}
export interface RateLimitBucketTable {
key: string;
tokens: number;
capacity: number;
refill_per_sec: number;
blocked_until: Timestamptz | null;
updated_at: GeneratedTimestamptz;
}
export interface ReconcileRunTable {
id: Generated<string>;
kind: ReconcileRunKind;
@@ -1653,6 +1662,7 @@ export interface DB {
preference_tag: PreferenceTagTable;
process_heartbeat: ProcessHeartbeatTable;
process_heartbeat_sample: ProcessHeartbeatSampleTable;
rate_limit_bucket: RateLimitBucketTable;
reconcile_run: ReconcileRunTable;
resolution_attempt: ResolutionAttemptTable;
rule_atom: RuleAtomTable;
@@ -1837,6 +1847,9 @@ export type ProcessHeartbeatUpdate = Updateable<ProcessHeartbeatTable>;
export type ProcessHeartbeatSampleRow = Selectable<ProcessHeartbeatSampleTable>;
export type NewProcessHeartbeatSample = Insertable<ProcessHeartbeatSampleTable>;
export type ProcessHeartbeatSampleUpdate = Updateable<ProcessHeartbeatSampleTable>;
export type RateLimitBucketRow = Selectable<RateLimitBucketTable>;
export type NewRateLimitBucket = Insertable<RateLimitBucketTable>;
export type RateLimitBucketUpdate = Updateable<RateLimitBucketTable>;
export type ReconcileRunRow = Selectable<ReconcileRunTable>;
export type NewReconcileRun = Insertable<ReconcileRunTable>;
export type ReconcileRunUpdate = Updateable<ReconcileRunTable>;
+21
View File
@@ -0,0 +1,21 @@
// 子进程:对同一 key 连续扣 N 个令牌,每次成功输出一行时间戳(毫秒)。由 rate-limit.test.ts 以 Deno 子进程启动。
import { Kysely, PostgresDialect } from "kysely";
import pg from "pg";
import { PgTokenBucket } from "../../src/rate-limit.ts";
const [url, schema, key, n, capacity, rate] = Deno.args as [string, string, string, string, string, string];
const db = new Kysely<unknown>({
dialect: new PostgresDialect({
pool: new pg.Pool({ connectionString: url, max: 1, options: `-c search_path=${schema},public` }),
}),
});
const bucket = new PgTokenBucket(db);
try {
for (let i = 0; i < Number(n); i++) {
const r = await bucket.acquire(key, 1, { capacity: Number(capacity), refillPerSec: Number(rate), waitMs: 30_000 });
if (!r.ok) throw new Error("等待超时");
console.log(Date.now());
}
} finally {
await db.destroy();
}
+98
View File
@@ -0,0 +1,98 @@
import { fileURLToPath } from "node:url";
import { Kysely, PostgresDialect, sql } from "kysely";
import pg from "pg";
import { afterAll, beforeAll, describe, expect, it } from "vitest";
import { type DB, migrateToLatest, PgTokenBucket } from "../src/index.ts";
// 跨进程令牌桶(0041):单进程语义 + 三个 Deno 子进程并发对同一 key 的总速率不超限。
const url = process.env.DATABASE_URL ?? "postgres://umt:umt@localhost:55433/umt";
const skip = process.env.UMT_SKIP_DB_TESTS === "1";
const schema = `umt_rl_${process.pid}_${Date.now().toString(36)}`;
describe.skipIf(skip)("PgTokenBucket(真实 PG)", () => {
let admin: Kysely<unknown>;
let db: Kysely<DB>;
beforeAll(async () => {
admin = new Kysely({ dialect: new PostgresDialect({ pool: new pg.Pool({ connectionString: url, max: 1 }) }) });
await sql.raw(`create schema ${schema}`).execute(admin);
db = new Kysely<DB>({
dialect: new PostgresDialect({
pool: new pg.Pool({ connectionString: url, max: 4, options: `-c search_path=${schema},public` }),
}),
});
await migrateToLatest(db, { migrationTableSchema: schema });
});
afterAll(async () => {
await db?.destroy();
await sql.raw(`drop schema if exists ${schema} cascade`).execute(admin);
await admin?.destroy();
});
it("容量内立即发放,之后按速率补充;waitMs=0 拿不到即返回", async () => {
const b = new PgTokenBucket(db);
const spec = { capacity: 2, refillPerSec: 10 };
expect(await b.tryTake("t:a", 1, spec)).toBe(0);
expect(await b.tryTake("t:a", 1, spec)).toBe(0);
const wait = await b.tryTake("t:a", 1, spec);
expect(wait).toBeGreaterThan(0);
expect(wait).toBeLessThanOrEqual(100);
expect((await b.acquire("t:a", 1, { ...spec, waitMs: 0 })).ok).toBe(false);
const r = await b.acquire("t:a", 1, { ...spec, waitMs: 2000 });
expect(r.ok).toBe(true);
});
it("penalize 顺延:期间不发放令牌", async () => {
const b = new PgTokenBucket(db);
const spec = { capacity: 5, refillPerSec: 100 };
await b.penalize("t:p", 400, spec);
const wait = await b.tryTake("t:p", 1, spec);
expect(wait).toBeGreaterThan(250);
const t0 = Date.now();
expect((await b.acquire("t:p", 1, { ...spec, waitMs: 3000 })).ok).toBe(true);
expect(Date.now() - t0).toBeGreaterThanOrEqual(200);
});
it("取消等待", async () => {
const b = new PgTokenBucket(db);
const spec = { capacity: 1, refillPerSec: 0.1 };
await b.tryTake("t:c", 1, spec);
const ac = new AbortController();
const p = b.acquire("t:c", 1, { ...spec, waitMs: 60_000, signal: ac.signal });
setTimeout(() => ac.abort(new Error("cancelled")), 50);
await expect(p).rejects.toThrow("cancelled");
});
it("三个子进程并发对同一 key:总速率不超过 容量 + 速率 × 时长", async () => {
const script = fileURLToPath(new URL("./fixtures/bucket-worker.ts", import.meta.url));
const capacity = 2;
const rate = 20;
const perProc = 12;
const procs = [0, 1, 2].map(() =>
new Deno.Command(Deno.execPath(), {
args: ["run", "-A", script, url, schema, "t:procs", String(perProc), String(capacity), String(rate)],
stdout: "piped",
stderr: "piped",
}).output(),
);
const outs = await Promise.all(procs);
const stamps: number[] = [];
for (const o of outs) {
expect(o.code, new TextDecoder().decode(o.stderr)).toBe(0);
for (const line of new TextDecoder().decode(o.stdout).split("\n")) if (line.trim()) stamps.push(Number(line));
}
stamps.sort((a, b) => a - b);
expect(stamps).toHaveLength(perProc * 3);
// 36 个令牌:首批 capacity 个立即发放,其余按 20/s → 至少 (36-2)/20 = 1.7s
const span = stamps.at(-1)! - stamps[0]!;
expect(span).toBeGreaterThanOrEqual(((perProc * 3 - capacity) / rate) * 1000 - 150);
// 任意 500ms 窗口内发放数 ≤ 容量 + 速率 × 0.5s(+1 容忍子进程打印时间戳的误差)
const win = 500;
for (let i = 0; i < stamps.length; i++) {
let j = i;
while (j < stamps.length && stamps[j]! - stamps[i]! < win) j++;
expect(j - i).toBeLessThanOrEqual(capacity + (rate * win) / 1000 + 1);
}
}, 60_000);
});
+4
View File
@@ -115,6 +115,8 @@ const EXPECTED_TABLES = [
"admin_recovery_code",
"admin_passkey",
"webauthn_challenge",
// C2-A 跨进程站点限流(0041)
"rate_limit_bucket",
].sort();
/** 不走 entityTable() 的表:追加型/引擎自有表,没有 uuid id + created_at/updated_at 三件套 */
@@ -139,6 +141,8 @@ const NON_ENTITY_TABLES = [
// 一次性凭据:只记创建 / 用掉的时间
"admin_recovery_code",
"webauthn_challenge",
// 限流桶:key 主键,只有 updated_at
"rate_limit_bucket",
];
describe.skipIf(skip)("schema migrations (real PG)", () => {