Changelog.
0.4.0 · 2026-11-19
Released automatically from c7507f6f on main, whose full CI run passed, compatibility shards included. Changes since v0.3.0 (diff). The curated entries for this work are under Unreleased in CHANGELOG.md at this tag; only a row of docs/compatibility.md marked Verified is a support claim.
Added
- #1855 (
feat/1855-artifact-registry-console, Landing 153) - #1858, #1859, #1893 (
feat/1858-1859-round-2, Landing 153) - #1895 (
feat/1895-custom-constraints-3, Landing 153) - #1898 (
feat/scheduled-release, Landing 152) - #1850, #1851 (
feat/1850-1851-secrets-kms-console, Landing 152) - #1848, #1849, #1858, #1859 (
feat/1858-bigtable-console, Landing 152) - #1845, #1846, #1880, #1881, #1883, #1884, #1885, #1888 (
feat/1888-datastore-operators, Landing 152) - #1861, #1864, #1879, #1886, #1887, #1889, #1890, #1891, #1892 (
feat/1891-org-policy-2, Landing 152) - #1524, #1860, #1875, #1876 (
feat/1860-sql-console-2, Landing 152) - #1874 (
feat/1874-pubsub-exports, Landing 152) - #1842, #1843, #1847, #1852, #1854, #1863, #1877, #1878, #1882 (
feat/1854-build-console-2, Landing 152) - #1839, #1840 (
feat/1840-hf-serve, Landing 152) - #1651 (
feat/1651-finish-3, Landing 152) - #1776 (
feat/1776-holidays-automl, Landing 152) - #1578 (
feat/1578-dropcolumn, Landing 152) - #1834 (
feat/1834-service-identity, Landing 152) - shared shell round 2 (
feat/1842-console-shell-2, feat/1847-pubsub-console) - #1799: Vision LABEL_DETECTION, OBJECT_LOCALIZATION and FACE_DETECTION on MediaPipe (
feat/1799-objects-faces, Landing 149) - #1799: ML.UNDERSTAND_TEXT, ML.TRANSLATE, ML.TRANSCRIBE, Vision image properties and text boxes (
feat/1799-nl, Landing 149) - #1762: copy jobs' destinationEncryptionConfiguration and layout rules (
feat/1762-copy-rules, Landing 149) - #1778: AI.SEARCH, autonomous embedding generation, VECTOR_SEARCH (
feat/1778-ai-search-rest, Landing 149) - #1819: jobs outlive their request, timeoutMs, RUNNING jobs, cancel (
feat/1819-bq-async-jobs, Landing 149) - #1810: AI.AGG's instruction as a query parameter (
feat/1810-aiagg-param, Landing 149)
Fixed
- main CI red after Landing 153 (
fix/main-153-race-bqml, the Landing 153 fix landing) - main CI red after Landing 153 (
fix/main-153-sql-served, the Landing 153 fix landing) - main CI red after Landing 153 (
fix/main-153-storage, the Landing 153 fix landing) - Cloud SQL version tests after #1524 (
fix/main-151-sql, Landing 153) - #1901 (
fix/1901-version-pins, Landing 153) - #1896, #1899 (
fix/main-149-red-2, Landing 152) - #1894 (
fix/dependabot-alerts, Landing 152) - #1838 (
fix/1838-terminal-loading, Landing 152) - #1824 (
fix/1824-bigtable-sessions, Landing 152) - #1837 (
fix/1837-offline, Landing 152) - #1791 (
fix/1791-anywherecache-pause, Landing 152) - #1833, #1813 (
fix/1833-1813-ci-red, Landing 152) - #1817, #1812, #1780: Cloud Identity groups, group and domain members, console groups (
fix/identity-membership-rules, Landing 149) - #1825: console Spanner, Bigtable and Storage clients never reach Google (
fix/no-google-egress, Landing 149) - #1816, with #1814: parameters keep their declared type (
fix/1816-null-array-params, Landing 149) - #1820: console Bigtable data clients on the classic data plane (
fix/1820-bigtable-sessions, Landing 149) - #1811, with #1808: Opus decoded by darui3018823/opus (
fix/1811-opus-memory, Landing 149)
Landings
0.3.0 · 2026-11-18
A fix release for v0.2.0, with one new BigQuery function. The generated coverage is unchanged: 1,133 of 1,227 methods Verified through official clients, across thirty API surfaces (docs/coverage). Only a row of docs/compatibility.md marked Verified is a support claim.
Added
BigQuery
AI.AGG, the aggregate generative AI function (#1777): one answer perGROUP BYgroup, as a STRING or anoutput_schemaSTRUCT, withDISTINCT,connection_id,endpointandmax_error_ratio. It runs on the instance's local Vertex AI endpoint, never Google, aggregating in batches that fit the local model's 4,096-token context.OVER, ObjectRef input and instructions that are not literals are 501.
Upgrades and docs
- The upgrade test starts from v0.2.0 (#1807): the nightly test downloads v0.2.0's published archive, checks it against the release's checksums, and upgrades it to this release, and can still start from v0.1.0. State that v0.2.0 wrote, in the state directory and in a
state savearchive, is read back in every CI run. docs/install.md names both releases in the upgrade promise. - Docs name the current release (#1806): install links, examples, the action's
version:line and the CLI flag table point at the newest release, and a test fails when a page pins a release older than this changelog's newest heading. - "What survives a restart" in docs/install.md (#1801) replaces "What is not persisted". A service is marked Yes only when a stop/up restart probe in CI proves it, and a test keeps the table and the probes in step.
- Signing and Gatekeeper (#605): docs/install.md says v0.2.0's macOS binaries are signed and notarized, and how to check one with
spctl -a -t open --context context:primary-signature.spctl -a -t execrejects any bare CLI as "not an app", so release.yml's check after notarization now uses the same command.
Fixed
Cloud Run
- CPU scaling works on small hosts. A container with a CPU limit now also requests 10m of CPU (or its limit, when that is smaller), instead of reserving the whole limit. In v0.2.0 each 1-CPU instance reserved a full core of the one kind node, so on a 4-CPU machine an automatically scaled service or worker pool could not grow past its first instance. The limit itself is unchanged (#1789, #1800, #1802).
gcloud beta run services updatewith--scaling-cpu-targetand--scaling-concurrency-targetworks: therun.googleapis.com/launch-stageannotation that gcloud beta sends is now accepted as the service's launch stage, where v0.2.0 answered 501. A value that is not a launch stage isINVALID_ARGUMENT(#1802).
BigQuery and Datastore
- Two tunnels no longer share one kubectl port. In v0.2.0, a port could be handed to the BigQuery and Datastore tunnels at once, so BigQuery answered 404 and Datastore failed with "frame too large". A tunnel is now never given a port already handed out, a kubectl that cannot listen fails its launch, and a guarded tunnel whose port is taken moves to a new one.
- Object tables' signed URLs name the real Cloud Storage port. With
--port-storage 0, the signed URLs ofEXTERNAL_OBJECT_TRANSFORM(..., ['SIGNED_URL'])namedhttp://127.0.0.1:0, the configured port, instead of the port the storage tunnel bound, so no client could fetch them (#1786). CREATE MODEL ... OPTIONS (MODEL_TYPE='TENSORFLOW')no longer times out on first use. The TensorFlow converter's image was built by the firstCREATE MODELand killed with it when the client's deadline passed, so a client that waits less than a cold build (a minute or more) never imported a model. The build now runs once, under its own 30-minute limit, and is never cancelled by a request waiting for it; nor is theCREATE MODELjob, which still finishes when its client gives up.cloudburrow prefetch --tensorflow-converterbuilds the 2.3 GB image ahead of time, for instances that will run offline (#1775). Its compat tests now give the statements after a cold build a deadline of their own:ML.PREDICTof the converted model was never slow, but on a CI runner it was sent with a deadline the build had already used up.
Console
- The console no longer contacts Google (#1825). Its Spanner data clients exported built-in metrics to Cloud Monitoring (
monitoring.googleapis.com) and probed the metadata server (169.254.169.254,metadata.google.internal) to find their location; its Bigtable clients' Direct Access check asked the metadata server too; and the client metrics of CloudBurrow's own Cloud Storage clients, turned on byGCP_STORAGE_GO_ENABLE_OTEL_METRICS, exported to Cloud Monitoring. Each is now off, whatever the environment, and an integration test fails on any dial or name lookup that leaves the host. - Bigtable pages no longer hang for 15 s (#1820). The console's Bigtable reads could wait out their deadline with "no sessions available" beside an idle session, a lost wake-up in the Go client's session pool. The console's Bigtable clients now use the classic data plane, which opens no session.
CLI
cloudburrow prefetchno longer caches a truncated image (#1805). Node images are exported to a file on the node and copied out withdocker cp, where v0.2.0 readdocker exec's output stream, which sometimes ended early and failedup --offlinewith "unexpected EOF". A copy shorter than its header says still fails and is never cached.
Every gRPC server
Stopwaits for every call's handler (#1804). This is v0.2.0's Resource Manager fix (#1803) applied to every server that serves gRPC through HTTP: the shared gRPC port, Secret Manager, Vertex AI, Cloud Storage, and the call fronts of Firestore, Datastore, Spanner and Pub/Sub. A call's trace span and request record are no longer cut off at shutdown, and Pub/Sub's front flushes its state only after the calls it cut off have finished.
0.2.0 · 2026-11-18
The second release. The generated coverage now tracks thirty API surfaces, up from nine in v0.1.0, and counts 1,133 of 1,227 methods Verified through official clients, up from 97 of 191 (docs/coverage). New since v0.1.0: Cloud Run's v1 (Knative) API, jobs and worker pools; Cloud Build, Artifact Registry, Secure Source Manager and Developer Connect; Cloud Functions v1 and v2; the Cloud SQL Admin API and Backup and DR; the Memorystore for Redis, Redis Cluster and Valkey admin APIs; IAM custom roles, predefined roles and service accounts; Service Usage; Firestore's Admin API and Firestore with MongoDB compatibility; Bigtable's instance admin API and SQL; the BigQuery Storage and Connection APIs, and BigQuery ML; and Resource Manager folders, organizations, tags and liens. CloudBurrow now builds its own Cloud Storage server and its own BigQuery emulator, and BigQuery serves every Resource Manager project. Local AI embeds text, images, audio, video and PDFs with EmbeddingGemma 2, OCR included, and serves Vertex AI's PredictionService over gRPC. IAM gains roles, service accounts and Conditions; policies are stored and still not enforced, with five named exceptions under Changed. The console has a Cloud Shell-style terminal and a screen for every service. Releases are smoke-tested before they become latest, and macOS signing and notarization are ready once the maintainer adds the Apple secrets. Only a row of docs/compatibility.md marked Verified is a support claim; docs/gaps.md lists what is still open.
Added
Cloud Storage
- CloudBurrow's own Cloud Storage server, built to the API, is now the only Cloud Storage backend (#485, #519). It serves buckets and objects over JSON, every upload type including resumable and XML multipart, ranged reads, every precondition, listing with prefixes, offsets, globs and pages, compose, copy, rewrite and move, the batch endpoint, versioning, soft delete and restore, retention policies, locks and holds, lifecycle rules, CORS, HMAC keys, notifications to Pub/Sub, the XML API, and signed URLs that fail closed (#489–#509, #1258). It runs as a locally built image in one Deployment, with reset, seed, state snapshots, persistence by
--mode, events, metrics and faults, and console pages (#510–#514, #518). - The gRPC API,
google.storage.v2, on the same port, sharing one store with JSON (#1321, #1345, #1443), including appendable objects and customer-supplied encryption keys. - Hierarchical namespace folders, managed folders and their updates, bucket long-running operations, relocation, Anywhere Cache and rapid caches,
objects.bulkRestore, object contexts, and Storage Intelligence configurations (#828, #1203, #1279, #1280, #1444, #1505, #1519, #1532, #1700). - Bucket and object ACLs, and bucket and managed-folder IAM policies, are stored and returned, never enforced (#504, #847, #1233).
softDeletePolicy.hardDeletePause, held to Google's rules (#1702, #1715).- Signed URLs verify against the instance's ADC key,
storage.signingCertsand the certificates of IAM service accounts (#577, #1787). - Verified through the official Python and Node.js clients,
gcloud storage,gsutiland Terraform, and compared against Google'sstorage-testbenchby a nightly oracle (#497, #515–#517, #589).
Pub/Sub
- Subscription expiration (#873, #891): a front in the Pub/Sub pod deletes a subscription idle for its
expirationPolicy.ttl(31 days by default), counts calls and open streaming pulls as activity, applies expiration updates that Google's emulator refuses, and works over gRPC and REST alike (#908). Exactly-once delivery is verified, and is refused with a push endpoint, as on Google (#880). - Pub/Sub in
state saveandstate load: schemas with their revisions, topics, subscriptions with every setting, and snapshots, but not messages (#880). Seed files take schemas, a topic'sschemaSettings, and a subscription's expiration, retention and exactly-once settings (#890, #899). - Topic and subscription label updates checked against Google's rules,
DetachSubscription,UpdateSnapshotand Protocol Buffer schema validation (#949, #962, #1406, #1446). - IAM policies on topics, subscriptions, snapshots and schemas, stored and not enforced (#1367).
- The front recreates every topic, subscription, schema and snapshot after the emulator restarts, so no client sees an empty emulator. Messages that were not yet acknowledged are still lost (#1608).
- Pub/Sub calls are recorded in
/metrics, the request log and traces, as the in-process services' calls are (#1266).
Cloud Tasks
UpdateQueueover gRPC and REST,RunTask,BatchCreateTasks,BatchDeleteTasks, and CMEK configuration, stored only (#692, #1215).- Rate limits are enforced:
max_dispatches_per_secondandmax_concurrent_dispatches, with queues dispatched concurrently anddispatch_deadlinehonoured per attempt (#579, #716).max_burst_sizeis CloudBurrow's own choice, because Google does not document it. oidcTokenandoauthTokenon HTTP tasks, minted locally and signed with the instance's key (#1243).- Queue-level
http_target,stackdriver_logging_config,ListQueuesfilters, and App Engine targets delivered to the server named by--appengine-url(#1222, #1255, #1256, #1282, #1447, #1506).
Cloud Run
- Jobs, executions and tasks run as Kubernetes batch Jobs (#582, #1223), and worker pools and instances (#1299, #1300, #1348, #1387).
- Scaling: service-level
scaling(automatic and manual, including manual zero instances), revisioncpuUtilizationandconcurrencyUtilization, and worker-pool automatic scaling over REST, which CloudBurrow drives on the pods' CPU utilization; the console edits service scaling (#1788, #1789, #1790, #1800, #1802). - Traffic splits, tags and pinned revisions, mapped to Knative (#1224).
- The service account, Secret Manager and in-memory volumes, Cloud SQL volumes through the Auth Proxy, NFS volumes, Cloud Storage FUSE volumes from the instance's own storage, the
h2cport name andK_SERVICE,K_CONFIGURATIONandK_REVISION(#1297, #1346, #1388, #1448, #1474, #1475, #1476). - IAM policies on services, jobs and worker pools, stored and not enforced (#1216).
- The v1 (Knative-shaped) Admin API that
gcloud runuses, including jobs, executions, domain mappings, instances and worker pools (#1293, #1349, #1379, #1420, #1533, #1546). gcloud run deploy --sourceandBuilds.SubmitBuild, and a REST surface for Cloud Run v2 that gcloud and Terraform reach (#591, #1268, #1347).- Container logs in Cloud Logging, with
logUri, and operations withListOperationsfilters andWaitOperation(#1229, #1263, #1298, #1333). - Two projects can each have a service, job or worker pool with the same ID (#1399, #1400).
Cloud Build, Artifact Registry, Secure Source Manager and Developer Connect
- Cloud Build: builds with docker build, push and pack steps; steps of any image when
--cloudbuild-host-stepsis set;secretEnv,script,allowFailureandallowExitCodes;RetryBuild, approvals and dynamic substitutions; manual, webhook and Pub/Sub triggers; private worker pools, stored; logs written to the default logs bucket (#1382, #1407, #1422, #1430, #1449). - Artifact Registry: a Docker registry that clients push to and pull from, with no authentication; Maven, npm, Python, Apt and Yum repositories; packages, versions, tags, files, rules, remote repositories over another repository, project configuration and Locations; and repository IAM, stored and not enforced (#1383, #1439, #1531, #1539, #1546, #1578).
- Secure Source Manager instances and repositories, git over HTTP, and push triggers into Cloud Build (#1521, #1527); Developer Connect connections and git repository links (#1546).
Cloud Functions
- The Cloud Functions v2 API (2nd gen, HTTP functions) and the v1 API (1st gen), with
gcloud functionsand Terraform (#1377, #1428). - v2
eventTriggerfor Pub/Sub and Cloud Storage events, delivered locally (#1414); secrets, KMS keys, worker pools and binary authorization on functions (#1464). - Functions in Node.js, Python, Java, Ruby, PHP and .NET, built with Google's buildpacks (#1303, #1319).
BigQuery
- CloudBurrow builds its own BigQuery emulator (#1061): goccy/bigquery-emulator v0.8.1, built from pinned sources with patches to its SQL engine (
third_party/bigquery-emulator) and embedded in the CLI. A validating front in the emulator's pod holds IDs, schemas, rows and statements to BigQuery's rules, and answers 501 for what the emulator cannot run (#861, #901, #902). - Every Resource Manager project is a BigQuery project (#1764, #1767, #1781): datasets, tables and jobs per project, queries, DML, loads, copies, clones and snapshots across projects, and
projects.listfrom Resource Manager. A name without a project is resolved in the job's project (#1784). - DDL:
ALTER TABLE(add, rename and drop columns,SET DATA TYPE, defaults, keys, options), views with column lists, materialized views,CREATE TABLE … LIKE,COPYandCLONE, snapshot tables, partitioning and clustering, column defaults,und:cicollation, and table, column and schemaOPTIONS(#1227, #1239, #1272, #1322, #1334, #1340, #1390, #1415, #1438, #1456, #1508). - Scripting:
IF,LOOP,WHILE,REPEAT,FOR,CASE,BEGIN … EXCEPTION,RAISE,ASSERT,EXECUTE IMMEDIATE, stored procedures, multi-statement transactions, and child jobs (#1037, #1201, #1217, #1253, #1353, #1459). INFORMATION_SCHEMA: dataset views and region-qualified views, includingJOBS,JOBS_BY_FOLDERandJOBS_BY_ORGANIZATION(#1274, #1275, #1286, #1328, #1329, #1351, #1355, #1392).- Time travel with
FOR SYSTEM_TIME AS OF, dataset and table undelete, and change history throughAPPENDSandCHANGES(#1276, #1418, #1460). - Loads: Avro, ORC and nested or repeated Parquet; CSV and JSON options, bad-record reporting and load statistics; autodetect,
columnNameCharacterMap, hive partitioning, manifest file sets and GeoJSON; and loads from the instance's Cloud Storage (#919, #944, #952, #960, #966, #988, #1004, #1324, #1326, #1339, #1457, #1547, #1713). - Extract jobs to the instance's Cloud Storage as CSV or newline-delimited JSON, with GZIP (#939, #957), and copy jobs that keep partitioning and clustering (#987, #1759).
- The Storage Read API (Arrow and Avro) and the Storage Write API (proto and Arrow rows, every stream type, missing values, RANGE and INTERVAL columns) (#1095, #1098, #1102, #1327, #1391, #1424).
- Routines (
routines.*), table functions, models (models.*),datasets.patch,updateandundelete,projects.getServiceAccountand row access policies (#1047, #1177, #1205, #1218, #1238). - Table and routine IAM policies, stored, with Terraform's
google_bigquery_table_iam_*(#1231, #1236). The per-project checks for minted tokens are under Changed. privacy_policy(aggregation threshold, join restriction and differential privacy), BigQuery tables for Apache Iceberg, and object tables over Cloud Storage with signed URLs (#1426, #1427, #1455, #1482, #1783, #1786).- BigQuery ML:
CREATE MODELwith linear and logistic regression, k-means, PCA, matrix factorization, boosted trees, random forests, DNNs, autoencoders, and ARIMA_PLUS with custom holidays, trained locally; imported ONNX, TensorFlow Lite and TensorFlow SavedModel models; and remote models over the local Vertex AI endpoint (ML.GENERATE_TEXT,ML.GENERATE_EMBEDDING) (#1465, #1481, #1518, #1548, #1562, #1600, #1711, #1760, #1765, #1775, #1776). compatibility.md lists the model types and functions served. - AI functions on the local Vertex AI endpoint, never Google:
AI.GENERATEwithoutput_schema,AI.GENERATE_BOOL,_INT,_DOUBLEand_TABLEthrough constrained decoding, andAI.GENERATE_EMBEDDING,AI.EMBED,AI.SIMILARITY,AI.CLASSIFY,AI.IFandAI.SCORE(#1771, #1772, #1779). Remote models honourmax_output_tokensand the sampling options (#1797). - Remote models over Cloud AI services:
REMOTE_SERVICE_TYPE = 'CLOUD_AI_VISION_V1'withML.ANNOTATE_IMAGEtext detection over an object table, read locally with Tesseract; other service types are 501 (#1773). - Image functions over object tables:
ML.DECODE_IMAGE,ML.RESIZE_IMAGE,ML.CONVERT_COLOR_SPACE,ML.CONVERT_IMAGE_TYPE, andML.PREDICTof an imported TensorFlow Lite or TensorFlow image model over them (#1798). - The BigQuery Connection API over REST and gRPC, with connection IAM. A deleted service account in a connection's policy reads as
deleted:and is dropped on purge, as in other services' policies (#1755, #1770, #1795). - Query parameters of every top-level type,
STRUCTandRANGEparameters, DML row counts and statement types, and thebqCLI'smk,insert,queryandrm(#696, #1008, #1082, #1108, #1111).
Firestore and Datastore
- The Firestore Admin API: databases, indexes, field overrides, backup schedules, backups, restore and clone, exports and imports, and user credentials (#1331, #1500).
- Firestore with MongoDB compatibility (#1703, #1738): CloudBurrow's own MongoDB wire-protocol server on
--port-firestore-mongodb, authenticated with the Admin API's user credentials or with an access token the instance minted. It serves CRUD, aggregation, indexes, transactions,explainand text search over the database's Firestore documents. - Firestore pipelines,
PartitionQuery, and DatastoreRunAggregationQueryandexplain_options(#1405, #1500, #1613). - Firestore keeps its documents across a pod restart and an emulator crash in persistent mode (#1615).
- Every Firestore and Datastore method coverage had marked Unknown is now verified through the official SDK (#1197).
Bigtable and Spanner
- Bigtable's instance admin API over gRPC and REST: instances, clusters, app profiles, logical views, memory layers, and IAM, stored (#1331, #1440, #1499).
- Bigtable authorized views, snapshots, backups, schema bundles,
UndeleteTableand change streams (#1499, #1511, #1523). - Bigtable SQL through
ExecuteQuery, including the documented functions and geography functions (#1499, #1511, #1704, #1720, #1728, #1735). - Spanner: the 55 methods coverage had marked Unknown, backups and backup schedules, instance configurations and partitions, database roles, and IAM, stored (#1199, #1486, #1498).
- Terraform
google_bigtable_*,google_spanner_*andgoogle_firestore_*resources (#1331, #1440, #1461).
Cloud SQL and Backup and DR
- The Cloud SQL Admin API (
sqladminv1 and v1beta4) on--port-sqladmin, over real PostgreSQL and MySQL servers (#1301): instances, databases and users; backups; read replicas, cascading and cross-region; high availability and failover; point-in-time recovery; import and export; SSL and server certificates; and the connector path that the Cloud SQL connectors and Auth Proxy use (#1330, #1380, #1408, #1425). - PostgreSQL 9.6 to 18 and MySQL 8.0 and 8.4, major version upgrades, clones to binary log coordinates, parallel imports and exports, external primaries and their syncs, DR replicas and switchover, blue-green deployments, CMEK and disk shrink (#1381, #1421, #1501, #1524, #1535, #1538, #1570, #1614, #1769).
- The Backup and DR Service API for Cloud SQL (
--services backupdr): backup vaults, plans, associations, data sources and restores, with gcloud and Terraform (#1433, #1462, #1536). - Terraform's
google_sql_*resources (#1301, #1304, #1330).
Memorystore
- The Memorystore for Redis admin API (
google.cloud.redis.v1, gRPC and REST) over real Valkey servers, one Deployment per instance (#1302): create, update, delete, export and import, RDB persistence, the Standard HA tier with read replicas and failover, upgrades, maintenance, CMEK and in-transit encryption (#1384, #1385, #1386, #1412, #1413, #1463, #1768). - Memorystore for Redis Cluster and Memorystore for Valkey: clusters and instances, token and IAM auth, ACL policies, backups, migration, cross-instance replication, Valkey 7.2 and 9.0, and CMEK (#1519, #1549, #1557, #1563, #1578, #1614).
gcloud redis,gcloud memorystore, Terraform and console screens for all three.
Secret Manager
- User-managed replica locations,
expire_timeandttl, rotation schedules,topicsnotifications to Pub/Sub,version_destroy_ttl, and list filters (#857, #1225, #1228, #1259, #1278). - Regional secrets and customer-managed encryption keys, including CMEK-protected versions mounted into Cloud Run (#1453, #1489, #1519).
Cloud KMS
- Asymmetric signing and decryption, MAC, raw symmetric and key encapsulation (ML-KEM) keys,
GenerateRandomBytes, key deletion and retired resources, and import jobs (#1212).EC_SIGN_SECP256K1_SHA256signing (#1719). - Automatic rotation (
rotation_period,next_rotation_time), editable in the console (#816). - EXTERNAL and EXTERNAL_VPC keys, EKM connections, Autokey, single-tenant HSM management, key access justifications and trusted key wrapping, without any real HSM or external key manager (#1452, #1469, #1506, #1521).
Locationsis served too (#1291). gcloud kmsandcloudburrow terraformfor key rings, keys, versions and IAM policies, which are stored and not enforced (#425–#431).
Cloud Scheduler
UpdateJob, a REST surface,Locations, OIDC and OAuth tokens minted locally, App Engine targets, and Terraform'sgoogle_cloud_scheduler_job(#591, #795, #1237, #1243, #1304, #1447).- State save, seeding, faults, tracing, the request log and metrics, as the other services have (#600).
Cloud Logging
- Log-based metrics, sinks, exclusions, log buckets and views, links, settings and CMEK settings (#1226, #1264, #1292, #1311).
- Sinks deliver to Cloud Storage, Pub/Sub, BigQuery and another project, and intercepting sinks (#1404, #1451, #1521).
- Folder, organization and billing-account parents, Log Analytics linked datasets that BigQuery can query,
CopyLogEntries,TailLogEntries, saved queries, log scopes and log view IAM (#1215, #1394, #1395, #1403, #1410, #1423, #1435, #1519). gcloud loggingand Terraform'sgoogle_logging_*resources (#1246, #1332, #1403).
Resource Manager, IAM and Service Usage
- Folders and the local organization,
MoveProject, tag keys, values, bindings and holds, and liens, over v1, v2 and v3, with Terraform (#1294, #1295, #1350). - Project, folder and organization IAM policies, stored and not enforced, with IAM Conditions stored and validated as Google documents them (#1368, #1742, #1745), and org policies on projects, folders and the organization (#1296, #1534).
UndeleteProject, v1projects.update, andPOST /admin/purge(#1216, #1265, #1537).- IAM: custom roles, Google's 2,423 predefined roles,
queryGrantableRolesandqueryTestablePermissions; service accounts and their keys;LintPolicy,QueryAuditableServicesand the deprecatedSignBlobandSignJwt; and the certificate URLs of service accounts, with gcloud, Terraform and console pages (#1743, #1746, #1747, #1750, #1751). - IAM Credentials:
generateIdToken,signJwtandsignBlob, besidegenerateAccessToken(#1747). - Service Usage (
serviceusage.googleapis.com) on--port-serviceusage, forgcloud servicesand Terraform'sgoogle_project_service. Disabling an API is recorded and not enforced (#1352, #1530).
Vertex AI and local AI
- Embeddings: a community ONNX conversion of EmbeddingGemma for text (#41), and one of EmbeddingGemma 2 for text, images, audio, video and PDFs, including
document_ocron scanned pages with Tesseract OCR (#1454, #1651, #1721–#1724, #1729, #1731–#1734, #1736, #1739). Both are labelled as community conversions, not Google artifacts. PredictionServiceover gRPC, forcloud.google.com/go/aiplatform:Predict,RawPredict,GenerateContent,StreamGenerateContent,EmbedContent,Explainwith Sampled Shapley, and the streaming and direct predict methods, raw ones relayed to a container'sgrpcPorts(#1193, #1244, #1257, #1284, #1285, #1454, #1475, #1477).- Local generation honours
temperature,topP,topK,seed,maxOutputTokensandstopSequenceson a newer LiteRT-LM, andresponseSchemathrough constrained decoding; the console playground offers them (#1779, #1797). - The local-AI runtime image is published with each release,
ghcr.io/cloudburrow/litert-lm:<tag>, and a release CLI uses it by digest, so local generation no longer needs a checkout (#602). - Console: model download, load and unload, an embedding view, and custom prediction requests from the playground (#869, #1305, #1306, #1396).
Console
- A terminal in the top bar, like Cloud Shell (#781, #834): a shell in a pod in the instance's cluster, never on this machine, from the pinned Cloud SDK image, with the instance's gcloud configuration and no credential. It has tabs, keeps its sessions across a reload, shows progress during the first image pull, and refuses any other origin (#824, #826).
- A screen for every service, and every feature reachable from it, checked by three parity audits (#782, #1552, #1591): edits, create options and IAM Permissions tabs across services (#784–#802, #852–#854, #1242); BigQuery, Firestore, Datastore, Bigtable, Spanner, Cloud SQL and Memorystore screens (#698, #868, #993–#999, #1207, #1343, #1358); fault injection, state save and load, reset, seed, and a Connect and About panel (#800–#802).
- Tested in headless Chrome in CI, at several viewport sizes and against a WCAG 2.1 AA audit (#594, #700, #1604).
CLI, tools and SDKs
--port-baseruns instances side by side, and--cloudsql-connectormoves the Cloud SQL connector listener (#584, #1364).cloudburrow seedandevents,env --format kubernetes, and endpoints injected into Cloud Run revisions; the metadata server and CLI-hosted APIs resolve from pods (#575, #576, #586, #707).cloudburrow prefetchandup --offlinefor air-gapped use (#604); see "Offline and air-gapped use" in docs/install.md.- Clusters are stamped with the CLI, node-image and Knative versions, and
upreconciles them (#601). State written by v0.1.0, in the state directory and in astate savearchive, loads in this release, Cloud Storage included; a nightly test upgrades a real v0.1.0 instance (#1605, #1606, #1714). gcloud-setupwrites the overrides for every API the instance serves, andcloudburrow terraformsets the matching custom endpoints. Terraform is verified for the resources compatibility.md lists, and OpenTofu for the modules its tests apply (#682, #719, #1304, #1578).--emulator-heap-miband--bigquery-heap-mibbound the emulators' memory (#1603, #1618).- The Python suite covers the opt-in emulators, Cloud KMS, BigQuery and impersonation, and a new Node.js suite covers Storage, Pub/Sub, Cloud Tasks, Secret Manager and Firestore (#589, #592).
Releases
- A release is built only from a commit whose
ci-greencheck passed, is published as a prerelease, and is marked latest and pushed to the Homebrew tap only after it smoke-installs on each platform's native runner (#596, #603, #680, #689). docs/install.md says how to withdraw a release. - The macOS binaries are signed with a Developer ID certificate and notarized once the maintainer adds the Apple secrets. Until then, releases are unsigned and the release notes say so (#605). docs/install.md explains how to check a binary.
- Release notes come from this changelog, and SECURITY.md says how to report a vulnerability (#603).
Changed
BigQuery
- BigQuery serves every ACTIVE Resource Manager project, and only those (#1764). v0.1.0 served one project. A project that Resource Manager does not have is now 404 "project … is not found", a project in
DELETE_REQUESTEDis 403, andprojects.listlists Resource Manager's projects. Create a project in Resource Manager, or seed one, before using it in BigQuery. - A name without a project now resolves in the job's project (#1784):
dataset.table, views, functions, procedures and models, and the default dataset'sprojectId. Before, the first project to create an object of that name won. - A request with an access token the instance minted is checked against that account's roles (#1767, #1774):
bigquery.jobs.createon the job's project, and data, dataset, table, routine, model and job permissions, with inherited roles, IAM Conditions and authorized views. Refusals are 403accessDenied. Requests with no bearer token, or with the ADC fixture's token, are served as before. - Remote models need their connection's permissions (#1765):
CREATE MODEL … REMOTE WITH CONNECTIONand every query of a remote model needbigquery.connections.useon the connection (403 otherwise), and need the connection's service account to holdroles/aiplatform.useron the model's project (400 otherwise).WITH CONNECTION DEFAULTuses the project's default connection, and creates it when the caller holdsbigquery.connections.create. Remote models are new in this release (#1760); this changes how they answered before #1765. - A
TIMESTAMPliteral with no time zone, and casts between dates, datetimes and timestamps, now use UTC as BigQuery does, not the emulator's America/Los_Angeles (#1188, #1189). tables.getand every job report times in milliseconds, not seconds (#971, #1310). A retriedjobs.insertof an existing job ID is 409 duplicate, not 400 (#1622). An unknown table or dataset in a query is 404notFound, notjobInternalError(#1582), and a failed query gives the same error throughjobs.queryandjobs.insert(#1581).- Query results go to a hidden
_cloudburrow_query_resultsdataset and expire after 24 hours or past 200 results, as BigQuery's anonymous datasets do.jobs.getQueryResultsof an expired result is 404 (#1017, #1059, #1086). - The emulator listens only on loopback in its pod, so pods reach it through the front's checks (#1114). Enabling BigQuery needs a CLI with the embedded emulator and storage server; a CLI built with a plain
go buildrefusesupwith BigQuery, anddoctorreports it (#902, #1061). The CLI is about 86 MB larger.
IAM, projects and credentials
- IAM policies are still not enforced, with five exceptions, each recorded in ADR-0006: BigQuery's per-project checks of minted tokens (above, #1767, #1774); BigQuery remote models' connection permissions (above, #1765); Memorystore for Redis Cluster and Valkey IAM auth (#1614); Firestore with MongoDB compatibility's token sign-in and command permissions (#1738); and IAM Credentials (
generateAccessToken,generateIdToken,signBlob,signJwt, and the IAM API's deprecatedSignBlobandSignJwt) on service accounts the IAM API created (#1747, #1751). All but the remote-model check accept only an access token this instance minted, for a principal holding the documented role. Calls as the instance's own account are checked no more than in v0.1.0. - One project number everywhere (#1211, #1544): the metadata server's
numeric-project-id, zone and machine type now carry the project's own number rather than000000000000, and Resource Manager v3 names a projectprojects/{number}, as Google does. A project recreated under a purged ID gets a new number. - Deleting a project keeps it in
DELETE_REQUESTEDfor 30 days, as on Google, instead of removing it at once. Its ID cannot be reused until it is purged, and the purge drops the project's state in every service (#1216, #1537, #1540–#1543, #1550). - Cloud Tasks and Cloud Scheduler accept
oidcTokenandoauthTokenand send a locally minted token, where v0.1.0 refused them (#1243).
Cloud Run
- Service accounts and Secret Manager, in-memory, Cloud SQL, NFS and Cloud Storage volumes are now mapped (#1297, #1346, #1448, #1474). VPC access, encryption keys, binary authorization,
executionEnvironmentandsessionAffinity, which v0.1.0 refused, are stored and returned but have no effect (#1346). - A revision that fails at startup is reported after 240 s rather than Knative's 600 s (#568).
Other services
- Secret Manager refuses the create fields it would otherwise drop, compares request etags, and verifies and returns payload CRC32C (#580). Each version is now a Kubernetes Secret of its own, so
AddSecretVersionno longer fails past about 900 versions (#1596). - Cloud Tasks keeps
dispatch_deadline, honoursmax_retry_duration, and refuses the fields it would drop (#578). - A Pub/Sub instance needs a CLI with the embedded Linux binaries (
make buildor a release), because its front runs from the storage image (#873). - Google's Java emulators (Pub/Sub, Firestore and Datastore) run with a 1 GiB heap by default, and the BigQuery emulator with a 512 MiB
GOMEMLIMIT. A Pub/Sub emulator whose heap fills exits and restarts (#1603, #1618). --mode ephemeralalso applies to Cloud KMS and Secret Manager state (#481, #483).- REST error bodies carry Google's
statusfield anddetails(#1470).
CLI
cloudburrow resetresets a running instance through its admin API (#586).upcreates the managed namespace whenever the cluster is up (#571), waits until Cloud SQL accepts a login and the cluster's DNS answers before it reports ready, and starts independent phases at the same time. A warm start takes about 29 s on macOS/arm64 (#1262, #1588, #1610, #1781).stop,envandstatusfind an instance by its PID and start time, not by process name (#820).envandstatusreport only the services and ports the named instance runs (#630, #652, #863).- A CLI built without the embedded Cloud Storage server is refused before
upcreates a cluster, anddoctorreports what is embedded (#686).doctornames the container engine and warns on unsupported ones (#712). - CI builds the embedded BigQuery emulator once per workflow run, or restores it from a cache (#1087).
make bigquery-binaries BQENGINE_FLAGS=-prebuiltchecks binaries built elsewhere.
Fixed
BigQuery
- Requests no longer slow down with every job: the emulator read every job it had run, with its results, on every request (#1086, #1103, #1133, #1661).
- The SQL engine no longer panics past 2 GiB of memory, and no longer grows by a whole builtin catalog at each DDL statement. An engine that does fail is restarted at once, and the front forgets the jobs it lost (#989, #1016, #1017, #1057, #1091).
- Table functions, nested RECORDs, NaN and ±Infinity values, non-UTF-8 BYTES,
-0, INTERVAL signs and comparisons, NULL arguments, GEOGRAPHY text andTO_JSON_STRINGnow give BigQuery's answers (#900, #1047, #1066, #1075–#1077, #1109, #1116, #1119–#1121, #1126–#1128). - CSV loads no longer drop the first row as a header, failed jobs no longer read back as succeeded,
IF NOT EXISTSof an existing table does nothing, and a failed script keeps the statements before the failing one, as BigQuery does (#931, #932, #934, #935, #955). - A table ID used in two datasets reads the right table in
tabledata.list, queries and the Storage Read API (#1015, #1032, #1046). - A
jobs.insertwith nojobReferenceno longer answers 500, andjobs.listhonours its paging and filters (#972, #973). - Storage Read API reads of REPEATED columns no longer crash the emulator, and Avro reads work in a project whose ID has a hyphen (#1095, #1098).
Pub/Sub
/admin/resetdeletes schemas too (#889).- Updated expiration policies survive a restart of the front, REST answers over 4 MiB pass through, and
GetSubscriptionno longer counts as activity (#898, #926, #1039). - The emulator's memory no longer rises for the length of a long run (#1603, #1619).
Cloud Run and Cloud Tasks
- A task deleted during its dispatch attempt is no longer recreated by the attempt (#656).
- Console actions on Cloud Run no longer dial a fixed port when the Run port is OS-assigned (#646).
Other services
- Bigtable answers
UNIMPLEMENTEDfor the 15 RPCs that crashed its emulator pod (#1206), and its pod no longer reachesdl.google.comon every start (#611). - Cloud KMS and Secret Manager no longer hold their lock across one Kubernetes read per record, so requests no longer time out on a large instance (#1232).
- Cloud SQL in ephemeral mode keeps a created instance's data across a server restart (#1740).
- Resource Manager's
Stopwaits for every call's handler, so a call's trace span is no longer cut short (#1803). upover a cluster made before the builtin Cloud Storage server removes the old server instead of crash-looping (#780).- The console loads back an archive that its own Save state wrote, over the upload limit (#1250), and long monospace text no longer pushes a 320 px page sideways (#1637).
CLI, networking and releases
- Port forwarding replaces a dead tunnel on evidence rather than on one client's failure, retries a launch that does not take, and notices a pod that restarted its containers (#526, #566, #572).
- The setup-cloudburrow action works with the CLI it installs (#679): it passes only the flags that CLI has, finds its own files outside composite actions, and is self-tested against the newest release. docs/ci.md has the table.
- Both Linux release archives are statically linked (#714): every release CLI is built with
CGO_ENABLED=0, and the release checks it. HTTP targets under.localhostare dialled on127.0.0.1without a DNS lookup. - With Cloud Run,
upchecks that the kind gateway can be bound, and otherwise names the engine and the workaround (#712). See docs/install.md.
Security
- Every
/adminroute needs the instance's admin token, because a cluster workload on Docker Desktop can reach the host's loopback ports (#553). - Every HTTP listener refuses a
Hostthat is not an IP address,localhost, a.localhostname or one of the instance's own names, answering 421. This is the DNS-rebinding defence, and it covers the tunnels to the HTTP emulators and the fronts in their pods (#676, #725, #1587). - Every listener now checks
Origin(#1586): a browser request from an origin that is not loopback, the listener's own, or one named with--cors-allow-originis 403 with no CORS headers. A cross-site form post could otherwise create resources. The Pub/Sub and Firestore emulators' permissive CORS is blocked at the tunnel. Cloud Storage's JSON API, which allowed any origin as Google's does, follows the same rule (#677). A request with noOrigin, as the SDKs andcurlsend, is unaffected. - A discovered
cloudburrow.jsoncan no longer setallowRemoteor a non-loopbackbindAddress, and a directory's config file and hooks run only once that directory is trusted (#598). uptightens an existing instance directory to mode 0700, and log redaction now covers quoted JSON tokens, bare bearer tokens and PEM key bodies (#1586). docs/security.md describes the model and each check.- The terminal's WebSocket is refused to any other origin, host or remote peer (#781).
- Releases build with Go 1.27.4, and dependencies with known reachable vulnerabilities were updated (#264, #1175, #1178).
Removed
- The community Cloud Storage emulator v0.1.0 ran, replaced by CloudBurrow's own server (#519).
- Google's product icons in the console, replaced by CloudBurrow's own marks (#684).
- About 59 MB of committed build output;
make buildand releases build the embedded servers (#585). - Java from the support matrix's planned languages: no Java suite is planned (#718).
0.1.0 · 2026-09-25
The first release: a local Google Cloud emulator that runs in a local Kubernetes cluster and is driven through the official Google Cloud SDKs. What is and is not supported is recorded per operation in docs/compatibility.md and summarised in docs/status.md.
Added
- A CLI and an owned local Kubernetes cluster:
up,status,stop,reset,delete,env,doctor,logs,wait,up --detach,status --format json,diagnose,gcloud-setupandgcloud-teardown, andcloudburrow terraform(#53, #54, #89, #293, #305, #325, #326, #327, #331). - Services started by default: Cloud Storage, with notifications to Pub/Sub (#95); Pub/Sub, on Google's emulator; Cloud Tasks with HTTP dispatch and retry (#66, #67, #322); Cloud Run v2 backed by Knative Serving, including revisions and updates (#68, #299, #300); and Secret Manager over gRPC and JSON, backed by Kubernetes Secrets (#93, #94).
- Opt-in services with
--services: Firestore, Datastore, Bigtable and Spanner on Google's emulators (#75, #371); BigQuery on a community emulator (#323); Cloud SQL for PostgreSQL and MySQL data planes (#122, #297, #311); Memorystore on Valkey (#296); Cloud Scheduler (#302); Cloud Logging write and read (#304); and Cloud KMS with symmetric keys, Encrypt and Decrypt, the version lifecycle and REST transcoding (#385–#424). Cloud KMS is not a security boundary. - Resource Manager v1 and v3 projects (#298, #301), local credentials and a GCE metadata server (#92), and IAM Credentials token generation from the local signer (#303).
- IAM policies stored, never enforced, on Secret Manager secrets and Cloud Tasks queues (ADR-0006, #308, #365, #366).
- Vertex AI on local runtimes: a
generateContentsubset verified through the officialgenaiSDK (#105, #106) and custom prediction containers (#90). - A local web console laid out after the Google Cloud console, with detail pages, actions where the backend supports them, SQL and query editors, Kubernetes views, metrics, a Logs Explorer and a request log (#97 onward).
- An admin API for reset, seed and event inspection (#70, #317, #319, #320, #334), state save and load (#338, #340), lifecycle hooks (#333), fault injection (#306), request logging (#314), request metrics (#341) and OpenTelemetry traces (#313, #379).
- The setup-cloudburrow GitHub Action (#330),
env --format terraformand--format docker-compose(#332). - SDK compatibility suites for Go and Python (#56, #335), and per-service API coverage generated from the proto surface (#337).
- Release packaging: archives for macOS and Linux on arm64 and amd64 with build attestations, one
checksums.txt, a Homebrew formula and an installer that refuses an archive it cannot verify (#71, #324).
Rendered from CHANGELOG.md at 0ea2594, synced 2026-11-19, and, for a release with no CHANGELOG.md section, its GitHub release notes.
Build against Google Cloud APIs, locally
Free and open source under Apache-2.0. No account, no sign-up, no Google Cloud bill.