How Do You Make Redshift Spectrum Queries on S3 the Fastest?

Why the Answer Is Correct

Redshift Spectrum is a scan-and-filter engine: the dominant cost is the number of bytes it must read from S3, so the fastest queries come from layouts that let it read less. A columnar format such as Parquet or ORC stores each column separately, so Spectrum projects only the columns the query references instead of reading entire rows. Partitioning the S3 prefixes on the columns used most often in WHERE clauses lets Spectrum prune whole partitions and never even list those objects. AWS's own Spectrum performance guidance pairs exactly these two recommendations — columnar formats and partitioning — as the highest-impact tuning steps, which is why B and C together are the answer.

Why the Other Options Are Wrong

Option A frames gzip compression around single 1–5 GB files, which is the wrong shape for Spectrum: gzip is not splittable for Spectrum workers, and very large individual objects limit the parallelism of the scan (AWS guidance favors roughly 100 MB–1 GB splittable files). Option D is the opposite failure mode — files under 10 KB create per-object overhead, heavy listing pressure, and terrible throughput. Option E is simply backwards: non-splittable formats prevent a large file from being read by multiple workers in parallel, so they slow queries down. Compression in general is helpful, but only when it is splittable and paired with sensible file sizing, which none of these distractors describe.

Community Comment Notes

GiorgioGss anchored the discussion by linking the AWS Redshift Spectrum external-table performance documentation, which is the same source this explanation relies on. rralucard_ and pypelyncar both walked through the mechanics: columnar formats like Parquet and ORC let Spectrum "scan only the needed columns," and partitioning lets it skip irrelevant prefixes entirely. andrologin summarized it as "Partioning helps filter the data and columnar storage is optimised" for OLAP workloads, and certplan quoted AWS guidance that partitioning should follow "the most common query predicates." The comment set is unanimous on B and C, and that consensus matches the vendor documentation rather than merely echoing the answer key.