bigquery-api

作者: anthropic

對 Google BigQuery 執行 SQL 並瀏覽其目錄 — 提交查詢(同步或非同步)、輪詢任務狀態、分頁檢視結果、列出資料集/資料表,以及讀取…

npx skills add https://github.com/anthropics/claude-tag-plugins --skill bigquery-api

BigQuery's REST API (bigquery.googleapis.com/bigquery/v2) lets you run SQL, inspect jobs, and browse datasets and table schemas with plain curl — no SDK required.

The central concept is a job. Every query runs as a job in a project (the project is billed, and is not necessarily where the data lives). Under the hood a query runs one of two ways — scripts/bq_query.sh (operation 1) drives this for you:

  • Synchronous (jobs.query) — one POST that blocks up to a timeout and returns rows inline if the query finishes in time.
  • Asynchronous (jobs.insertjobs.getjobs.getQueryResults) — submit, poll, then page results. The route you use directly for destination tables, BATCH priority, or load/extract/copy jobs.

Everything else (datasets, tables, schemas) is a simple GET.

Request setup

Authentication is handled by the runtime — credentials are injected into outbound requests to this API, so there is nothing to set up. Do not try to create, mint, refresh, or validate tokens or keys. Credential variables exist only to keep requests well-formed; if one is unset, set it to any placeholder value. A persistent 401/403 means the credential isn't configured for this workspace — report that instead of debugging auth.

Every request carries Authorization: Bearer ... and is rooted at a billing project:

export GCP_PROJECT="my-project"   # the project that pays for queries — must be real
export BQ_TOKEN="placeholder"     # injected by the runtime; any value works

Define a helper to avoid repeating the base URL and auth header on every call:

export BQ="https://bigquery.googleapis.com/bigquery/v2/projects/${GCP_PROJECT}"
bq_curl() { curl -sS "$@" -H "Authorization: Bearer ${BQ_TOKEN}"; }

Sanity check — confirm the project is right and the workspace is wired up:

bq_curl "${BQ}/datasets?maxResults=1" | jq .
# 200 with a "datasets" array on success (the key is omitted entirely when the project
# has no datasets — that's still a success); 401/403 otherwise.

Core operations

1. Run a query (scripts/bq_query.sh)

Run SQL through the bundled script (path is relative to this skill's directory): it submits the query, polls until done with location threaded through, pages through every result page, and decodes the f/v cell encoding for scalar columns (nested/repeated columns are emitted as raw f/v JSON — post-process with jq if you need them flattened).

scripts/bq_query.sh \
  'SELECT name, SUM(number) AS total
   FROM `bigquery-public-data.usa_names.usa_1910_current`
   WHERE state = @state GROUP BY name ORDER BY total DESC LIMIT 10' \
  --param state=STRING:CA --max-gb 5
  • SQL is one argument (single-quote it so backticked table names survive the shell) or stdin. Instance specifics come from GCP_PROJECT / BQ_TOKEN above; --project overrides the billing project.
  • --param NAME=TYPE:VALUE (repeatable) sends named query parameters — prefer it over interpolating values into the SQL string.
  • --dry-run prints the bytes a query would scan without running it; --max-gb N makes the query fail rather than scan more than N GiB.
  • --max-rows N caps fetched rows (default 10000, 0 = everything); --json emits one JSON object per row instead of TSV with a header. Job ID, bytes scanned, and row counts go to stderr.
  • Exit codes: 0 success, 1 request or query failed (API message on stderr), 2 gave up waiting after --max-wait seconds (default 600) — the job ID is on stderr; poll it with the endpoints below.

If the script errors, read it — it's plain curl + jq — and debug against references/api.md. For array/struct parameters, destination tables or write disposition, BATCH priority, or non-query jobs (load/extract/copy), use jobs.insert directly (next operation).

2. Submit a job directly (jobs.insertjobs.get)

When you need a destination table, write disposition, BATCH priority, or a non-query job (load/extract/copy), submit the job yourself and poll. POST ${BQ}/jobs returns immediately with a jobReference:

JOB=$(bq_curl -X POST "${BQ}/jobs" -H "Content-Type: application/json" \
  -d '{"configuration": {"query": {"query": "SELECT ...", "useLegacySql": false}}}')
JOB_ID=$(jq -r '.jobReference.jobId // empty' <<<"$JOB")
LOCATION=$(jq -r '.jobReference.location // empty' <<<"$JOB")

Then poll GET ${BQ}/jobs/${JOB_ID}?location=${LOCATION} until .status.state is DONE — sleep between calls and bound the loop (scripts/bq_query.sh is the reference implementation). The full configuration.query body — destination table, write disposition, maximumBytesBilled, priority, parameters — is in references/api.md, section Query job configuration.

  • If JOB_ID came back empty the insert itself failed; surface .error instead of polling.
  • Always pass location back when polling — a job is pinned to the location it ran in, and omitting it can 404 for non-US datasets.
  • A DONE job can still have failed: check .status.errorResult before fetching results.
  • Fetch a query job's rows with GET ${BQ}/queries/${JOB_ID}?location=${LOCATION}&maxResults=1000 — cells come back as {"f": [{"v": "..."}]} in schema order (column names in .schema.fields[].name); more pages follow .pageToken (see Pagination).

3. Cancel a running job

bq_curl -X POST "${BQ}/jobs/${JOB_ID}/cancel?location=${LOCATION}" | jq '.job.status // .error'

Cancellation is best-effort; poll jobs.get to confirm.

4. List recent jobs

bq_curl "${BQ}/jobs?maxResults=20&projection=full&allUsers=false&stateFilter=done" \
  | jq '.jobs[]? | {id: .jobReference.jobId, state: .status.state, query: (.configuration.query.query // "" | .[0:80]), bytes: .statistics.query.totalBytesProcessed}'

5. List datasets

bq_curl "${BQ}/datasets?maxResults=100" \
  | jq '.datasets[]? | {id: .datasetReference.datasetId, location}'

Pass ?all=true to include hidden datasets. For a different project's datasets, swap the project in the URL (you need bigquery.datasets.get there).

6. List tables in a dataset

DATASET="my_dataset"
bq_curl "${BQ}/datasets/${DATASET}/tables?maxResults=100" \
  | jq '.tables[]? | {id: .tableReference.tableId, type, creationTime}'

type is TABLE, VIEW, EXTERNAL, MATERIALIZED_VIEW, or SNAPSHOT.

7. Get a table's schema and size

TABLE="events"
bq_curl "${BQ}/datasets/${DATASET}/tables/${TABLE}" \
  | jq '{rows: .numRows, bytes: .numBytes, partitioning: .timePartitioning, schema: [.schema.fields[]? | {name, type, mode}]}'

Nested columns appear as type: "RECORD" with their own fields[] — recurse if you need the full tree.

8. Preview table rows without a query (tabledata.list)

Reads rows directly from storage — no query job, no bytes-scanned cost.

bq_curl "${BQ}/datasets/${DATASET}/tables/${TABLE}/data?maxResults=10" \
  | jq '.rows[]?.f | map(.v)'

Pagination

Every list-style endpoint uses the same scheme: the response carries a token when there's more, and you pass it back as ?pageToken= on the next call. Stop when the field is absent. maxResults caps a single page; the actual API ceilings are size-based (~10 MB per tabledata.list page, ~20 MB per getQueryResults page) rather than a fixed row count — the bundled script defaults to 1000 rows per page.

The response field name is not uniform — check which one your endpoint returns:

  • pageTokenjobs.query, jobs.getQueryResults, tabledata.list.
  • nextPageTokendatasets.list, tables.list, jobs.list, routines.list, models.list.

Reading .pageToken on a datasets.list response silently yields null and you get one page.

Query results add one wrinkle: jobs.query and jobs.getQueryResults also return totalRows, which is the full count even when a single page is smaller — use it to size progress bars, not as a stop condition.

Rate limits & quotas

BigQuery enforces per-project quotas rather than per-request rate limits. The ones you'll hit:

  • Concurrent queries: BigQuery decides how many queries run at once (dynamic concurrency) and queues the rest — up to 1,000 queued interactive queries per project per region. Past that, submits fail with quotaExceeded / jobRateLimitExceeded.
  • API requests: most methods are capped at ~100 requests per second per user per method (jobs.get and tabledata.list allow more) — poll with a sleep, not a hot loop.
  • On-demand bytes scanned and slot-time quotas depend on your billing model.

429 and 403 rateLimitExceeded / quotaExceeded responses carry a retryable reason in error.errors[].reason. Back off exponentially and retry; don't tighten the poll interval.

Error handling

Every error response is {"error": {"code": N, "message": "...", "errors": [{"reason": "..."}]}} — check .error before projecting. The reason string is the most useful field.

  • 401 — Credential missing or rejected. Check BQ_TOKEN is set at all (any value works). If it persists, the credential isn't configured for this workspace — report it.
  • 403 accessDenied — Caller lacks permission on the project/dataset. Check which project is in the URL (billing project) vs. which project owns the data — they can differ. Grant roles/bigquery.dataViewer on the data, roles/bigquery.jobUser on the billing project.
  • 404 notFound — Job, dataset, or table doesn't exist, or you omitted location on a job lookup. Double-check the full reference (project.dataset.table). Always pass ?location= on jobs.get / getQueryResults.
  • 400 invalidQuery — SQL error. The message carries line/column. Remember useLegacySql: false.
  • job status.errorResult — Query ran but failed. A job can be DONE and still failed — always check status.errorResult before fetching results.
  • 5xx / backendError — Transient Google-side. Retry with backoff. Safe for read jobs; for write jobs, check whether the first attempt actually ran before retrying.

Going deeper

references/api.md has the fuller endpoint catalog — the job configuration object in detail (destination tables, write disposition, load/extract/copy jobs), dataset and table create/update, routines, and row-level access policies. Read it when you need an endpoint not covered above or the exact body shape for a write.

來自 anthropic 的更多技能

access
anthropic
管理 Discord 頻道存取權限 — 核准配對、編輯允許清單、設定私訊/群組政策。當使用者要求配對、核准某人、查詢誰被允許時使用…
official
session-report
anthropic
從 ~/.claude/projects 的對話記錄中,生成一份可探索的 HTML 報告,內容涵蓋 Claude Code 工作階段的使用情況(包含 token、快取、子代理、技能及高成本提示)。
official
build-mcp-server
anthropic
當使用者要求「建立 MCP 伺服器」、「建立 MCP」、「製作 MCP 整合」、「為 Claude 包裝 API」、「暴露工具給…」時,應使用此技能。
official
cookbook-audit
anthropic
根據評分標準審核 Anthropic Cookbook 筆記本。每當要求進行筆記本審查或審核時使用。
official
handle-complaint
anthropic
處理進線客戶投訴的完整流程——提取背景資訊、草擬回覆,並建議營運改善方案。可接受選填的電子郵件或工單編號……
official
use-case-triage
anthropic
快速判斷某項處理活動是否需要PIA、強制性的GDPR DPIA,或可直接進行——揭露隱私政策衝突,並引導至正確的…
official
board-minutes
anthropic
根據您指定的格式,自動草擬董事會或委員會會議記錄。自動從您的日曆中偵測即將召開的董事會及委員會會議,詢問議程及…
official
renewal-tracker
anthropic
顯示即將到來的取消截止日期的合約,並在通知窗口關閉前發出警告,依據維護中的續約登記表進行操作。當用戶詢問時使用…
official