Why bqemulator?

Testing code against real BigQuery is slow (network + service latency), expensive (every query is billable), and dangerous (no rollback in shared environments). The alternatives — mocks, fakes, and shared sandboxes — drift from the real service the moment you stop chasing them.

bqemulator is a process you can run locally that speaks BigQuery's actual wire protocol (REST + gRPC), backs onto a real analytical SQL engine (DuckDB), and translates GoogleSQL → DuckDB SQL with a rule-based, ADR-grounded translator (SQLGlot). The official google-cloud-bigquery, @google-cloud/bigquery, cloud.google.com/go/bigquery, com.google.cloud:google-cloud-bigquery, and bq CLI clients all work against it unchanged — only the endpoint differs.

Three use cases, one binary:

• Ephemeral CI fixture — pytest plugin starts an in-process emulator on a random port; pip install bqemulator[testing] is all the wiring you need.
• Long-running local dev server — bqemulator start --data-dir ~/bqemu persists state across runs; works with the official bq CLI, dbt, Airflow, PySpark, Beam, Scio.
• Offline replica of a real project — bqemulator import --from-project <id> clones schema (and optionally data) from real BigQuery into a local data directory.

Highlights

• 🟢 Full REST + gRPC API parity — Datasets, Tables, Jobs, TableData, Routines, Row Access Policies, Authorized Views, plus Models CRUD metadata. Storage Read API (Arrow and Avro). Storage Write API (all four stream types — DEFAULT, COMMITTED, PENDING, BUFFERED — with both proto and Arrow row formats).
• ⚡ Real SQL — GoogleSQL translated to DuckDB SQL via 92 SQLGlot rules + 24 rewriters; covers date/time, string, array, struct, range, geography, JSON, approximate-aggregate, statistical, regex, civil-time, and bit operations.
• 🧠 Features goccy/bigquery-emulator doesn't have — JavaScript UDFs (embedded V8 via mini-racer), procedural scripting (DECLARE / BEGIN…END / IF / LOOP / EXCEPTION / BEGIN TRANSACTION), time travel (FOR SYSTEM_TIME AS OF), table snapshots, table clones, materialized views with refresh dispatch, GEOGRAPHY (planar via DuckDB-spatial + S2 helpers), RANGE, INTERVAL, authorized views, row-access policies, INFORMATION_SCHEMA.
• 🔌 Five-client e2e matrix — every release is exercised against the official Python, Node.js, Go, and Java BigQuery client libraries plus Google's bq CLI in a live Docker container.
• 🧪 7-tier test pyramid — unit + property + integration + conformance + e2e + perf + chaos, plus mutation / fuzz / differential siblings. Combined coverage is gated at ≥90% line + branch.
• 📐 Conformance corpus — 1,288 fixtures recorded against real BigQuery. Drift between the emulator and the real service surfaces as a failing test; documented divergences are pinned with ADR references.
• 🐍 Native pytest plugin — pip install bqemulator registers a pytest plugin; the bqemu_server fixture starts an ephemeral in-process emulator on random free ports and sets BIGQUERY_EMULATOR_HOST. No conftest.py wiring required.
• 🐳 Multi-arch container — ghcr.io/jjviscomi/bqemulator builds for linux/amd64 + linux/arm64, with cosign keyless signatures via GitHub OIDC.
• 🔭 Production-grade observability — structlog JSON logs, OpenTelemetry tracing (configurable OTLP exporter), Prometheus metrics endpoint.

Install

pip install bqemulator

Optional extras:

pip install "bqemulator[testing]"      # pytest, hypothesis, testcontainers, bigquery client
pip install "bqemulator[udf-js]"       # JavaScript UDF support (embedded V8)
pip install "bqemulator[orc]"          # ORC format for load jobs
pip install "bqemulator[compression]"  # zstd + snappy for load/extract jobs
pip install "bqemulator[import]"       # bqemulator import --from-project
pip install "bqemulator[all]"          # all runtime extras (no testing extras)

Docker:

docker run --rm -p 9050:9050 -p 9060:9060 ghcr.io/jjviscomi/bqemulator:latest

Both pip and the published image bundle the same emulator. The image exposes REST on 9050 and gRPC on 9060 by default — see configuration reference to change them.

Windows users: install Docker Desktop for Windows with the WSL2 backend (default since Docker Desktop 4.x); the published Linux image runs natively under WSL2 with no Windows-specific configuration. Native Windows-container variants of the image are explicitly out of scope for v1.0 — see docs/reference/out-of-scope.md#native-windows-containers for the rationale.

Quickstart

Python

import os
from google.cloud import bigquery

# Either set BIGQUERY_EMULATOR_HOST (picked up by every Google Cloud library)
# or pass api_endpoint explicitly to the Client. Both work.
os.environ["BIGQUERY_EMULATOR_HOST"] = "localhost:9050"

client = bigquery.Client(project="my-test-project")

client.create_dataset("sales")
client.create_table(
    bigquery.Table(
        "sales.orders",
        schema=[
            bigquery.SchemaField("id", "INT64"),
            bigquery.SchemaField("amount", "NUMERIC"),
            bigquery.SchemaField("placed_at", "TIMESTAMP"),
        ],
    )
)
client.insert_rows_json(
    "sales.orders",
    [{"id": 1, "amount": "12.50", "placed_at": "2026-05-21T00:00:00Z"}],
)

for row in client.query("SELECT COUNT(*) AS n FROM sales.orders").result():
    print(row.n)  # 1

pytest

bqemulator ships a pytest plugin via the pytest11 entry point. Installing the package is all the wiring you need — your conftest.py stays empty.

from google.cloud import bigquery

def test_orders_table(bqemu_client: bigquery.Client) -> None:
    bqemu_client.create_dataset("sales")
    # ... your test ...

The bqemu_server fixture is session-scoped (one emulator per test session); the bqemu_client fixture is function-scoped and returns a pre-configured bigquery.Client. See the pytest fixture guide and the python/pytest-integration example for a complete Flask app with integration tests.

Node.js

const { BigQuery } = require('@google-cloud/bigquery');

const bq = new BigQuery({
  projectId: 'my-test-project',
  apiEndpoint: 'http://localhost:9050',
  token: 'dummy',  // emulator accepts any token
});

await bq.createDataset('sales');

See the Node.js quickstart and the nodejs/nestjs-app example.

Go

client, _ := bigquery.NewClient(
    ctx, "my-test-project",
    option.WithEndpoint("http://localhost:9050"),
    option.WithoutAuthentication(),
)

See the Go quickstart and the go/beam-pipeline example.

Java

BigQuery bq = BigQueryOptions.newBuilder()
    .setProjectId("my-test-project")
    .setHost("http://localhost:9050")
    .setCredentials(NoCredentials.getInstance())
    .build()
    .getService();

See the Java quickstart and the java/spring-boot example.

bq CLI

bq --api=http://localhost:9050 \
   --project_id=my-test-project \
   query --use_legacy_sql=false 'SELECT 1 AS n'

See the bq CLI guide and the bq-cli-quickstart example.

docker-compose

services:
  bqemulator:
    image: ghcr.io/jjviscomi/bqemulator:latest
    ports: ["9050:9050", "9060:9060"]
    healthcheck:
      test: ["CMD", "curl", "-sf", "http://localhost:9050/healthz"]
      interval: 2s
      retries: 30

  app:
    build: .
    environment:
      BIGQUERY_EMULATOR_HOST: bqemulator:9050
    depends_on:
      bqemulator: { condition: service_healthy }

See the docker-compose/full-stack example for app + emulator + Prometheus + Grafana.

What works today

bqemulator is at v1.4.0 — fourth minor on the production-stable line. SemVer applies: breaking changes ship only in MAJOR, deprecations live ≥2 MINOR or 6 months. The compatibility matrix is auto-generated from the conformance corpus on every CI run; the conformance coverage matrix breaks down support by surface item.

Conformance corpus depth (the conformance coverage matrix carries the live, auto-generated breakdown):

Plus 10 non-deterministic items (RAND, CURRENT_*, SESSION_USER, GENERATE_UUID, TABLESAMPLE, FOR SYSTEM_TIME AS OF <expression>) that are excluded from the conformance corpus by ADR 0022 and exercised in unit / property / integration tiers instead — bringing the full inventory to 419 surface items across 20 categories, backed by a 1,288-fixture conformance corpus (1,213 SQL + 49 HTTP + 26 gRPC) under tests/conformance/.

We follow a no-deferral principle: features either ship complete or are excluded with documented rationale. There is no "TODO for v1.1." Scope boundaries are catalogued in docs/reference/out-of-scope.md.

Documentation

The full documentation lives at jjviscomi.github.io/bqemulator. Key entry points:

• Getting started — your first ten minutes.
• Per-language quickstarts — Python · Node.js · Go · Java · pytest · docker-compose · Testcontainers.
• Guides — loading data, querying, streaming inserts, Storage API, UDFs, scripting, partitioning, time travel, materialized views, row access policies, dbt, Airflow, Spark, the bq CLI, observability, and more.
• Reference — configuration, CLI, REST coverage, SQL function mapping, compatibility matrix, conformance coverage matrix, out-of-scope catalogue, troubleshooting.
• Architecture — hexagonal architecture, storage model, SQL translation, jobs lifecycle, Storage Read/Write API design, scripting, UDFs, versioning, row access, specialized types, observability, testing strategy, conformance tier.
• ADRs — 48 Architecture Decision Records documenting every non-obvious design choice.

Examples

Every example under docs/examples/ is a complete, runnable project with its own make test validated by CI:

Project status

bqemulator is at v1.4.0 — fourth minor on the production-stable line. SemVer applies: breaking changes ship only in MAJOR versions, preceded by ≥1 MINOR with deprecation warnings; deprecated APIs remain for ≥2 MINOR versions or 6 months.

Maturity signals:

• ✅ 48 Architecture Decision Records covering every non-obvious design choice (docs/adr/0001–0048).
• ✅ ≥90% line + branch coverage gated by CI (make verify).
• ✅ 7 test tiers passing (unit + property + integration + conformance + e2e + perf + chaos).
• ✅ 5-client e2e matrix (Python · Node.js · Go · Java · bq CLI).
• ✅ Mutation-tier (mutmut) pilot landed on pure-domain modules.
• ✅ Fuzz-tier (Atheris) harnesses on the SQL translator, dynamic-protobuf decoder, and Arrow bridge.
• ✅ Differential-tier row-order perturbation of the entire conformance corpus passes.
• ✅ Performance baselines committed for darwin-arm64, with regression gates (pytest-benchmark --benchmark-compare-fail=median:10%).
• ✅ PyPI publish via Trusted Publishing (sigstore-attested wheels) — pip install bqemulator==1.4.0 resolves from PyPI.
• ✅ GHCR publish with keyless cosign signatures — docker pull ghcr.io/jjviscomi/bqemulator:1.4.0 resolves and the image is cosign-verifiable.

See CHANGELOG.md for the complete release-by-release inventory.

Contributing

We welcome contributions of all sizes. Start with CONTRIBUTING.md for the mechanics; AGENTS.md captures the project's day-to-day conventions; and docs/architecture/overview.md is the canonical architectural reference.

Pull requests are squash-merged into main with a Conventional Commits subject; commits carry a DCO sign-off (git commit -s). The full review policy lives in GOVERNANCE.md.

Community

• 💬 GitHub Discussions — design questions, usage questions, and general help.
• 🐛 Issues — bug reports and feature requests. Please search existing issues first.
• 🔒 Security advisories — report vulnerabilities privately via the GitHub Security Advisory flow (see SECURITY.md for our disclosure policy).
• 📜 Code of Conduct — adapted from the Contributor Covenant 2.1.

License

bqemulator is released under the Apache License 2.0.

Acknowledgements

• goccy/bigquery-emulator for blazing the trail and providing a decade of issue reports that seeded our regression corpus.
• DuckDB, SQLGlot, FastAPI, Pydantic, Hatchling, and the Google Cloud client library teams whose work makes this project tractable.
• The Apache Beam, dbt, Airflow, PySpark, Spotify Scio, NestJS, and Spring Boot communities whose work the example projects compose with.