Catch slow queries before your users feel them.

queryd is a lightweight query observability layer for Node.js: driver-agnostic hooks (wrapQueryFn, wrapTaggedTemplate, SlowQueryDetector.executeQuery), optional Prisma integration via @olegkoval/queryd/prisma, sampling, optional per-request query budgets, optional EXPLAIN ANALYZE, and pluggable sinks.

• Add instrumentation in minutes, not days.
• Keep your current DB client and query style.
• Detect both single slow queries and request-level query storms.

Project site: oleg-koval.github.io/slow-query-detector.

npm install @olegkoval/queryd

Add @prisma/client only if you use the Prisma entry (see below).

import postgres from "postgres";
import {
  createSlowQueryDetector,
  wrapTaggedTemplate,
  createConsoleLogger,
  runWithDbContext,
} from "@olegkoval/queryd";

const sql = postgres(process.env.DATABASE_URL!);
const detector = createSlowQueryDetector(
  {
    warnThresholdMs: 200,
    dbName: "primary",
    requestBudget: { maxQueries: 80, maxTotalDurationMs: 2_000 },
  },
  { logger: createConsoleLogger() },
);
const db = wrapTaggedTemplate(sql, detector);

await runWithDbContext({ requestId: "req-home-1", userId: "u-42" }, async () => {
  await db`select ${1}::int`;
});

What you get right away:

• Structured query events you can route to logs/APM.
• Automatic request context propagation with runWithDbContext.
• Budget violation signal (db.request.budget) for high-query requests.

Use this command for a fast, reproducible local micro-benchmark:

npm run benchmark

Docker validation (recommended for CI/consistent environment):

docker run --rm -v "$PWD:/app" -w /app node:22 \
  bash -lc "npm ci --ignore-scripts && npm run benchmark"

Docker Compose validation:

npm run benchmark:docker

benchmark:docker installs dependencies with dev packages included, so tsx is available and no interactive install prompt appears.

E2E benchmark against Postgres in Docker (real DB round-trips):

npm run benchmark:e2e:docker

This mode starts a postgres:16 container and benchmarks wrapTaggedTemplate against a real select 1 query path. Use this for end-to-end signal; use npm run benchmark for isolated in-process wrapper overhead.

Dataset-backed E2E benchmark (seeded table + filtered aggregate query):

npm run benchmark:e2e:dataset:docker

Optional dataset/loop tuning:

BENCH_DATASET_ROWS=200000 BENCH_ITERATIONS=3000 BENCH_WARMUP=300 npm run benchmark:e2e:dataset:docker

This mode seeds bench_users in Postgres and runs a realistic indexed filter + aggregate query, so results include actual DB planner/cache/index behavior in Docker.

Multi-run series (mean / median / stdev / CV%) in one container — good for publishing stable numbers:

npm run benchmark:e2e:dataset:10:docker

CI compares each PR against bench/e2e-dataset-baseline.json (median ceilings + tolerance); see docs/benchmark-regression.md for how to turn enforcement on, refresh baselines after real improvements, and how to read the aggregate table and BENCH_JSON_RESULT output.

Timing is interleaved (all six orderings of the three scenarios rotate each iteration) so one path is not always measured first while Postgres buffers and JIT warm up — that avoids misleading rows where “overhead” looks negative.

If you see a negative “Added overhead” on the budget row, that is almost always measurement noise, not “instrumentation made Postgres faster”: same query and data, but order effects (the first scenario pays cold cache / planner / JIT; later rows look artificially fast). The current script uses interleaved timing to reduce this; treat small negative deltas as “≈0 within noise” if they still appear.

Example output from a sequential-only run (your parameters):

• Iterations: 3000
• Warmup: 300
• Dataset rows: 200000
• Node: v22.22.2
• Platform: linux arm64
• Database URL host: postgres

Override benchmark loop sizes (optional):

BENCH_ITERATIONS=200000 BENCH_WARMUP=10000 docker compose run --rm benchmark

Environment:

• Node v20.20.2
• Platform linux arm64
• Iterations 100000, warmup 5000

Interpretation:

• These numbers describe added latency overhead, not "speed increase."
• Example: +839% here means about +1.687 us/op over a tiny 0.201 us/op baseline.
• Large percentages are expected when the baseline operation is near-zero.

Scope (important):

• This micro-benchmark runs a noop async query function to isolate queryd wrapper cost.
• It does not execute real SQL and does not contact a database.
• Iterations: 100000 means 100k benchmark function calls (in-process), not 100k DB round-trips.

Why not benchmark only against DB in Docker?

• Real DB benchmarks mix many variables (network, kernel scheduler, DB cache, query planner, disk, container contention).
• For library overhead, we first measure a controlled in-process baseline so wrapper cost is visible.
• DB-inclusive benchmarks are still valuable, but they answer a different question: end-to-end system performance, not pure instrumentation overhead.

npm install enables a pre-commit hook (Husky) that runs npm run lint, npm run format:check, and npm test. To skip hooks for a one-off commit: HUSKY=0 git commit ….

SlowQueryDetector.executeQuery runs your callback and emits structured events (QueryEvent for each query, and optionally RequestBudgetViolationEvent — see Request budgets). Wrappers cover common shapes:

• wrapQueryFn — (sql: string, params?) => Promise<unknown> (raw clients, thin DB helpers).
• wrapTaggedTemplate — tagged template (strings, ...values) => Promise<unknown> (e.g. postgres.js sql, same literal shape as Prisma $queryRaw).
• extractQueryInfo — build $1…$n SQL + params from a TemplateStringsArray if you wire a custom executor.

createSlowQueryDetector appends a default LoggerSink with Array#push when your sinks array does not already include one, so the array you pass is mutated. Pass an array you own, or copy first (sinks: [...existing]), if immutability matters.

Use runWithDbContext so each HTTP request (or job) gets a stable requestId; createSlowQueryDetector defaults contextProvider to getDbContext(), so you usually do not pass contextProvider unless you merge ALS with your own source.

import postgres from "postgres";
import {
  createSlowQueryDetector,
  wrapTaggedTemplate,
  createConsoleLogger,
  runWithDbContext,
} from "@olegkoval/queryd";

const sql = postgres(process.env.DATABASE_URL!);
const detector = createSlowQueryDetector(
  {
    warnThresholdMs: 200,
    dbName: "primary",
    requestBudget: { maxQueries: 80, maxTotalDurationMs: 2_000 },
  },
  { logger: createConsoleLogger() },
);
const instrumentedSql = wrapTaggedTemplate(sql, detector);

await runWithDbContext({ requestId: "req-abc", userId: "user-42" }, async () => {
  await instrumentedSql`select ${1}::int`;
  await instrumentedSql`select ${2}::int`;
});

import {
  createSlowQueryDetector,
  wrapQueryFn,
  createConsoleLogger,
  runWithDbContext,
} from "@olegkoval/queryd";

const rawQuery = async (sql: string, params: unknown[]) => {
  /* your client */
  return [];
};
const detector = createSlowQueryDetector(
  { warnThresholdMs: 200, requestBudget: { maxQueries: 50 } },
  { logger: createConsoleLogger() },
);
const q = wrapQueryFn(rawQuery, detector);

await runWithDbContext({ requestId: "job-7" }, async () => {
  await q("select 1", []);
});

Background, practices, and sources: docs/budget.md.

Set requestBudget.maxQueries and/or requestBudget.maxTotalDurationMs to catch query storms (many fast queries) that stay under single-query latency thresholds. Counts and duration are summed per requestId for the lifetime of that id in the LRU map (in-process only). Negative or non-finite maxQueries values are ignored (no cap on query count); use 0 to mean “no queries allowed” before the first violation.

• Successful and failed executeQuery completions both increment the budget (every round-trip attempt counts).
• The first time a limit is exceeded for a requestId, sinks receive one db.request.budget event (LoggerSink → warn). A second violation for the same id is only possible after that id falls out of the LRU (e.g. many concurrent requests with unique ids).
• LRU eviction resets counters: when a requestId is evicted, its budget state is dropped. If that same id string appears again later, totals start from zero and another db.request.budget can fire for a new burst.
• requestBudget.maxTrackedRequests bounds memory (default 5000); non-finite or < 1 values are normalized.

Custom sinks: IEventSink.handle receives DetectorEvent (QueryEvent | RequestBudgetViolationEvent). Branch on event.event === "db.request.budget" before assuming sql / subtype exist.

Advanced: new SlowQueryDetector(config, undefined, sinks) does not install the default getDbContext() provider — use createSlowQueryDetector or pass contextProvider: { getContext: () => getDbContext() } if you construct the detector yourself.

npm install @olegkoval/queryd @prisma/client

import { PrismaClient } from "@prisma/client";
import { createSlowQueryDetector, createConsoleLogger, runWithDbContext } from "@olegkoval/queryd";
import { wrapPrismaClient } from "@olegkoval/queryd/prisma";

const base = new PrismaClient();
const detector = createSlowQueryDetector(
  {
    warnThresholdMs: 200,
    dbName: "primary",
    requestBudget: { maxQueries: 80, maxTotalDurationMs: 2_000 },
  },
  { logger: createConsoleLogger() },
);
export const prisma = wrapPrismaClient(base, detector);

await runWithDbContext({ requestId: "req-1", userId: "u-1" }, async () => {
  await prisma.$queryRaw`SELECT 1`;
});

wrapPrismaClient lives on @olegkoval/queryd/prisma so the core package stays free of a hard @prisma/client dependency. The Prisma path is the same requestBudget + runWithDbContext pattern as the driver-agnostic examples above.

Use ILogger and wire your adapter (e.g. @sentry/nextjs captureMessage / captureEvent) in the app; the package ships createNoopLogger and createConsoleLogger only.

• @queryd/core scoped publish (if namespace available)
• Companion packages: queryd-go, queryd-py (WIP)
• Raise test coverage thresholds back toward 100% (edge branches in ExplainThrottle / nested transaction paths)

See CONTRIBUTING.md. If adoption was awkward in your stack, open Integration feedback via New issue → choose a template (pick Integration feedback).

MIT — see LICENSE.