JST深夜コミットは本当に2倍か — GH Archiveで検証する日本OSSの夜型仮説

import os
from google.cloud import bigquery
from google.api_core.exceptions import GoogleAPIError

PROJECT = os.environ.get("GCP_BILLING_PROJECT", "my-billing-project")

SQL = """
SELECT
  EXTRACT(YEAR FROM created_at) AS yr,
  actor.login AS actor,
  repo.name AS repo,
  MOD(EXTRACT(HOUR FROM created_at) + 9, 24) AS jst_hour,
  COUNT(*) AS events
FROM `githubarchive.year.2019`
WHERE type = &
GROUP BY yr, actor, repo, jst_hour
"""


def fetch(page_size=1000):
    client = bigquery.Client(project=PROJECT)
    try:
        job = client.query(SQL, location="US")
        # result() paginates server-side; iterating fetches pages lazily.
        for row in job.result(page_size=page_size):
            late = row.jst_hour >= 22 or row.jst_hour < 2
            yield row.yr, row.actor, row.jst_hour, row.events, "LATE" if late else ""
    except GoogleAPIError as e:
        raise SystemExit(f"BigQuery request failed: {e}")


if __name__ == "__main__":
    for rec in fetch():
        print(*rec)

import pandas as pd, numpy as np
from scipy import stats

# Expects commits.csv with: contributor, commit_time (UTC ISO), repo_owner_type ('org'/'user')
df = pd.read_csv("commits.csv", parse_dates=["commit_time"])

# Assign each contributor a home tz via modal UTC commit hour -> implied tz offset
utc_hour = df["commit_time"].dt.hour
df["utc_hour"] = utc_hour
modal = df.groupby("contributor")["utc_hour"].agg(lambda s: s.mode().iat[0])
# JST contributors: modal activity peaks at evening JST (UTC+9). Peak local ~21h -> UTC ~12h
df["modal_utc"] = df["contributor"].map(modal)
df["local_hour"] = (df["utc_hour"] + 9) % 24          # JST local hour
df["is_jst"] = df["modal_utc"].between(10, 14)         # modal UTC evening-JST band
df["late"] = df["local_hour"].between(0, 3)            # 00:00-04:00 JST

# H1: two-proportion z-test, JST vs pooled baseline
def zprop(x1, n1, x2, n2):
    p = (x1 + x2) / (n1 + n2)
    se = np.sqrt(p * (1 - p) * (1/n1 + 1/n2))
    z = (x1/n1 - x2/n2) / se
    return z, 2 * (1 - stats.norm.cdf(abs(z)))
jst = df[df["is_jst"]]
z1, p1 = zprop(jst["late"].sum(), len(jst), df["late"].sum(), len(df))
print("H1 JST late share %.3f vs base %.3f z=%.2f p=%.4g" %
      (jst["late"].mean(), df["late"].mean(), z1, p1))

# H2: within JST late-night commits, org vs user repo share
jl = jst[jst["late"]]
for t in ["user", "org"]:
    sub = jst[jst["repo_owner_type"] == t]
    print("H2 %s late share %.3f" % (t, sub["late"].mean()))

git clone https://github.com/example/jst-midnight-oss.git
cd jst-midnight-oss
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt   # google-cloud-bigquery pandas numpy scipy
export GCP_BILLING_PROJECT=your-billing-project
python fetch.py > commits_raw.tsv
python analyze.py