Agent Skills
› google/skills
› bigquery-bigframes
bigquery-bigframes
GitHub生成基于 BigQuery DataFrames (BigFrames) 的 Python 代码，适用于 Pandas/scikit-learn 风格的 DataFrame 操作及 ML 任务。指导在云端处理数据、使用 peek() 预览、避免本地物化及优先使用 DataFrame API 而非 SQL。
Trigger Scenarios
编写 BigFrames 代码
对 BigQuery 进行 Pandas 风格的数据框或机器学习工作
在笔记本环境中使用 BigQuery 数据处理
Install
npx skills add google/skills --skill bigquery-bigframes -g -y
SKILL.md
Frontmatter
{
"name": "bigquery-bigframes",
"metadata": {
"category": "BigDataAndAnalytics"
},
"description": "Generates Python code using BigQuery DataFrames (BigFrames), the pandas\/scikit-learn-style API over BigQuery. Use when writing BigFrames code or doing pandas-style dataframe\/ML work against BigQuery (e.g. in a notebook). Don't use for SQL-first workflows or the google-cloud-bigquery client library — use bigquery-basics."
}
BigFrames (BigQuery DataFrame) basics
BigFrames is a Python library that lets you take advantage of BigQuery data processing by using familiar Python APIs.
Dataframe API best practices
- Stay in the Cloud: Perform data cleaning, transformation, and analysis via BigFrames methods to leverage BigQuery's scale rather than downloading data.
- Prefer partial ordering mode: Enable partial ordering mode right after
importing BigFrames. This speeds up data processing significantly by relaxing
row-sequence constraints.
import bigframes.pandas as bpd bpd.options.bigquery.ordering_mode = 'partial' - Use
peek()for data preview: Usepeek(n)to preview data instead ofhead(n).peek(n)randomly samplesnrows and is significantly faster.head(n)returns rows in strict order and fails inpartialordering mode unless the DataFrame has been explicitly sorted. - Avoid materializing data locally: Methods like
to_pandas()download all data to client memory, bypassing BigQuery’s distributed computation and risking Out of Memory (OOM) errors. Do not materialize data locally unless:- The dataset is small enough to fit safely in memory.
- An error message explicitly requires local materialization.
- Prefer Dataframe API over SQL queries: Do not write raw SQL queries via
read_gbq()if a DataFrame/Series method achieves the same result, as it breaks the Pandas abstraction and prevents lazy query execution. - Accessors over UDFs/Lambdas:
- Use built-in accessors (e.g.,
df.col.str.*,df.col.dt.*) instead of remote User Defined Functions (UDFs). UDFs require extra resources and time to deploy. - Do not use lambdas with
Series.map()orDataFrame.apply(). These methods do not accept functions withoutudforremote_functiondecorators.
# Avoid: df["upper"] = df["name"].map(lambda x: x.upper()) # Prefer: df["upper"] = df["name"].str.upper() - Use built-in accessors (e.g.,
- Schema Verification: Do not assume the schema of intermediate outputs.
Proactively verify schemas using
.dtypesand inspect sample records usingdisplay()with.peek(). - Visualization: Plot directly from the BigFrames DataFrame/Series when
possible. BigFrames is compatible with Matplotlib and Seaborn. If direct
plotting fails, use the
.plotaccessor. If the dataset is too large to plot, aggregate or sample the data before calling.to_pandas()to plot locally.
Machine Learning
- Use
bigframes.bigquery.mlpackage: Do not use Scikit-learn or other ML libraries with BigQuery DataFrames. Standard Scikit-learn models require bringing data into local client memory, whereasbigframes.bigquery.mldelegates training directly to BigQuery's scalable ML engine. Import functions frombigframes.bigquery.ml.
Reference Directory
- Linear Regression: Train a linear regression model to predict numerical values.
- Logistic Regression: Train a logistic regression model to predict boolean values.
BigFrames ML (Legacy)
The BigFrames ML package (bigframes.ml) is a legacy package that mimics the
scikit-learn API but is no longer recommended for new projects. Only use this
package if the user explicitly requests BigFrames ML.
- Legacy Imports: When legacy BigFrames ML is requested, import tools and
classes from
bigframes.mlinstead ofbigframes.bigquery.ml. - DataFrame Return on Prediction: Unlike Scikit-learn, BigFrames'
predict()method always returns a DataFrame containing both predictions and features, rather than a single series of predictions. - No
random_state: Do not pass arandom_stateargument when instantiating BigFrames ML models, as this parameter is not supported in the BigFrames ML package. - Automatic Scaling: Do not use
OneHotEncoderorStandardScalerunless explicitly requested, as scaling is handled automatically. - Hyperparameter Tuning: Write custom loops for hyperparameter tuning, as
BigFrames lacks
GridSearchCVorRandomizedSearchCV. - ARIMA Plus (Forecasting):
- Import from
bigframes.ml.forecasting. - Sort data chronologically and split around a timepoint before training.
- Ensure the prediction horizon is less than or equal to the training horizon.
- Import from
- PCA: BigFrames' PCA class lacks a
transform()method. Usepredict()instead. - Model Persistence: To persist a model, use
model.to_gbq(). To load a persisted model, usebpd.read_gbq_model().
Version History
- 513a7a5 Current 2026-07-19 19:02
- aabe37a 2026-07-05 15:28
