APACレビュー税：日本発PRの初回レビューはなぜ約2.7倍待つのか

"""Fetch APAC PR review-latency from GH Archive via BigQuery. Requires GCP ADC."""
import sys
from google.cloud import bigquery
from google.api_core.exceptions import GoogleAPICallError, RetryError

SQL = """
WITH opens AS (
  SELECT repo.name AS repo, JSON_VALUE(payload,&
         created_at AS opened_at, EXTRACT(HOUR FROM created_at) AS open_hour_utc,
         CAST(JSON_VALUE(payload,&
         ARRAY_LENGTH(JSON_QUERY_ARRAY(payload,&
  FROM `githubarchive.day.20240*`
  WHERE type=&
),
reviews AS (
  SELECT repo.name AS repo,
         COALESCE(JSON_VALUE(payload,&
                  JSON_VALUE(payload,&
  FROM `githubarchive.day.20240*`
  WHERE type IN (&
  GROUP BY repo, pr
)
SELECT o.*, TIMESTAMP_DIFF(r.first_review_at, o.opened_at, MINUTE) AS latency_min,
       (o.open_hour_utc < 9) AS jst_window
FROM opens o JOIN reviews r USING (repo, pr) WHERE r.first_review_at > o.opened_at
"""

def main(out="apac_review_latency.parquet", page_size=10000):
    try:
        rows = bigquery.Client().query(SQL).result(page_size=page_size)  # server-side pagination
        rows.to_dataframe().to_parquet(out)  # iterates all pages
    except (GoogleAPICallError, RetryError) as e:
        print(f"BigQuery failed: {e}", file=sys.stderr)
        sys.exit(1)

if __name__ == "__main__":
    main()

import numpy as np
import pandas as pd
import statsmodels.formula.api as smf
from scipy.stats import mannwhitneyu

# Expected columns: author_apac(bool), first_reviewer_apac(bool), hours_to_first_review(float),
#                   lines_changed(int), repo(str), dow(int 0-6), overlap_hours_to_first_review(float)
df = pd.read_csv("prs.csv")
df = df[df["hours_to_first_review"] > 0].copy()

# H1: APAC penalty after controls
print("Median h2fr APAC vs non-APAC:",
      df.groupby("author_apac")["hours_to_first_review"].median().to_dict())
df["log_h2fr"] = np.log(df["hours_to_first_review"])
df["log_lines"] = np.log1p(df["lines_changed"])
df["weekend"] = df["dow"].isin([5, 6]).astype(int)
m = smf.ols("log_h2fr ~ author_apac + log_lines + C(repo) + weekend", data=df).fit()
b = m.params["author_apac[T.True]"]
print(f"H1 APAC coef={b:.3f} (x{np.exp(b):.2f}), p={m.pvalues[&


apac = df[df["author_apac"]]
same = apac.loc[apac["first_reviewer_apac"], "hours_to_first_review"]
diff = apac.loc[~apac["first_reviewer_apac"], "hours_to_first_review"]
u, p2 = mannwhitneyu(same, diff, alternative="less")
print(f"H2 median same-tz={same.median():.1f}h diff-tz={diff.median():.1f}h, MWU p={p2:.4g}")


raw_gap = (df.loc[df["author_apac"], "hours_to_first_review"].median()
           - df.loc[~df["author_apac"], "hours_to_first_review"].median())
overlap_gap = (df.loc[df["author_apac"], "overlap_hours_to_first_review"].median()
               - df.loc[~df["author_apac"], "overlap_hours_to_first_review"].median())
print(f"H3 raw gap={raw_gap:.1f}h -> overlap-only gap={overlap_gap:.1f}h")

git clone https://github.com/example/apac-review-tax.git
cd apac-review-tax
python3 -m venv .venv && source .venv/bin/activate
pip install google-cloud-bigquery pandas pyarrow statsmodels scipy db-dtypes
gcloud auth application-default login   # ADC for BigQuery
python fetch.py                          # -> apac_review_latency.parquet
python analyze.py                        # prints H1/H2/H3