Reader benchmark parity

Issue #470 compares the extracted reader with the frozen Delta Funnel source from issue #459. The baseline source is commit e2650427ff5e2e1a7a4e5ef9eaf30969b217fec3; its benchmark source blob is 52284e22a5deb0cddce4aa7257012468dc88f25e.

Only the 12 controlled reader cases move with this crate. Delta Funnel keeps its synthetic policy, host probe, write workflow, tracing, Python, and SQL Server benchmark modes.

Method

Both release-profile binaries used their locked dependency graphs on the same Linux x86_64 host, with the baseline and extracted runs alternated per case. Each case ran two warm-up repetitions in a separate invocation followed by five measured repetitions. Inputs retained the frozen seed, fixture, query, backend, storage model, scheduling profile, Parquet controls, and schema-22 CSV fields.

The current equivalent invocation is:

cargo bench --locked -p delta-arrow-reader --bench reader --all-features -- \
  --mode provider-exec --seed 0 \
  --provider-exec-storage-profile local \
  --provider-exec-workload provider_few_larger_files \
  --provider-exec-query full_rows \
  --provider-exec-backend direct_parquet \
  --provider-exec-scheduling-profile prefetch_2_ap_target_scan_3x \
  --provider-exec-parquet-metadata-size-hint-bytes 65536 \
  --provider-exec-parquet-full-file-read-threshold-bytes disabled \
  --provider-exec-repetitions 5 \
  --output target/delta-arrow-reader-benchmark.csv

The captured rows predate the backend naming cleanup and retain their original schema-22 native_async and official_kernel values. The current harness uses direct_parquet and delta_kernel and writes schema 30. The workloads and measurements are unchanged.

Change only the workload, query, backend, storage profile, and Parquet controls to reproduce the other frozen cases listed below. The harness rejects cases outside that matrix, validates each fixture fingerprint, and validates output schema and row count before accepting a sample.

Controlled result

Captured on 2026-08-10. Times are microseconds. The range is the minimum and maximum of the five measured repetitions. Source rows per second is the raw schema-22 value.

All deterministic CSV fields matched exactly, including fixture size and fingerprint, planned and completed work, produced rows and batches, deletion vector counters, Parquet request and byte counters, and configured bounds. The measured ratios range from 0.71x to 1.09x, with no material extraction regression outside the observed sample spread.

The redacted raw schema-22 rows are committed in benchmark-parity-results.csv. Its implementation and comparison_case prefix columns identify each baseline and extracted pair; the remaining 80 columns are the unmodified benchmark output. Each row records the five-repetition summary, including p50, p95, p99, minimum, and maximum timing observations.

Frozen test audit

The focused benchmark_harness integration target includes the applicable frozen assertions for the 12-case matrix, accepted and rejected options, fixture shapes and fingerprints, retained-fixture ownership, unequal-file pruning, delayed HTTP reads, percentile and optional-value edges, Linux memory parsing, CSV width, deletion-vector fields, and the Delta Kernel backend's unavailable Parquet I/O metrics.

The remaining tests from the 71-test frozen benchmark binary stay in Delta Funnel because they exercise code that did not move: synthetic partition policy, host probes and local-I/O probes, tracing and detailed profiling, large product-mimic schemas, phase-aligned write workflows, and the multi-mode runner's unrelated CLI and CSV fields.

The per-invocation working files remain ignored under target/issue-470-comparison/; the committed result file above is the durable comparison evidence.