BigQuery Performance

Four Key Elements of Work

• I/O — how many bytes read? (읽기 바이트 수)
• Shuffle — bytes passed to next stage;
  • Grouping — bytes per group (그룹별 바이트)
• Materialization — bytes written to storage (쓰기 바이트 수)
• CPU work — UDFs & built-in functions (CPU 사용량)

Avoid I/O Waste

• Don’t SELECT * → select only needed columns.
• Push down filters early with WHERE.
• Avoid ORDER BY without LIMIT.

Prevent Data Hotspots

Shuffle wisely

• Filter early to avoid overloading workers on JOIN.
• Use Query Explanation map → compare Max vs Avg stage times to spot skew.
• BigQuery auto-reshuffles overloaded workers (자동 재셔플링).

GROUP BY

• Best when distinct groups are small.
• Bad: grouping by high-cardinality unique IDs.

Joins & Unions

• Know join key uniqueness → avoid accidental cross joins.