CONCEPT Cited by 1 source
Amdahl's law¶
Definition¶
Amdahl's law states that the theoretical maximum speedup of a program from parallelization is bounded by the fraction of work that must remain serial. If fraction s is serial, the speedup on N processors is at most:
As N → ∞, speedup caps at 1/s. This makes identifying and shrinking the serial fraction the dominant lever for strong scaling.
Implications for distributed query engines¶
- For scan-dominated aggregations (read Parquet from object storage, local partial-aggregate, merge), the serial fraction is small (planning + final merge) and strong scaling approaches N (Source: sources/2026-07-29-redpanda-single-query-scaling-in-redpanda-sql).
- For pruned queries where Iceberg statistics eliminate >99% of files at planning time, the remaining work is so small that adding nodes has minimal effect — the "serial fraction" is effectively the entire remaining query.
- Shuffle-heavy queries (joins, window functions) increase the serial/coordination fraction and reduce the benefit of scale-out.
Seen in¶
- sources/2026-07-29-redpanda-single-query-scaling-in-redpanda-sql: Explicitly cited; once pruning cuts a scan to ~1 GB, adding nodes gives negligible benefit (4.8 s → 3.3 s across 8× compute) because the workload is too small to parallelize.