Result buffer reuse and successful-call allocations

Measurements

Sassy 0.2.6, AVX2, one thread pinned to CPU 19 on an Intel Core i5-13500. Medians of seven runs after a warm-up, with inputs resident before timing. The adapter benchmark describes the timing boundaries and workloads: 262,144 DNA records and 4,096 dense ASCII records. The baseline includes cached DuckDB vector handles.

Workload	Path	Before (s)	After (s)	Elapsed time reduction (%)
ascii	ffi	0.137	0.134	2.3
ascii	sql	0.165	0.164	0.6
ascii	upstream	0.132	0.132	0.0
dna	ffi	0.755	0.751	0.5
dna	sql	0.811	0.773	4.7
dna	upstream	0.711	0.706	0.7

DNA SQL elapsed time falls from 0.811 s to 0.773 s (4.7%). The direct upstream median also falls by 0.7%; these sequential runs include run-to-run variation. The native FFI DNA change is small, and the dense ASCII SQL change is 1 ms. These measurements support a modest SQL improvement on the DNA workload; they do not establish a general speedup across inputs or separate the effects of buffer reuse from successful-call bookkeeping.

SQL takes about 9.5% longer than direct upstream calls for DNA and 24.1% longer for dense ASCII in the after measurements. SQL includes database execution, result materialization, aggregation and persistent CLI/R communication. The FFI path excludes the C runtime dispatcher. Both native paths use the same executable, dependency and compiler flags.

Allocation and ownership

Each searcher can retain one consumed result, including its hit and CIGAR buffer capacity. SQL returns each consumed result to its searcher; the native FFI benchmark uses the same sassy_c_result_recycle() operation. Searcher destruction releases that retained storage. Actual errors still allocate their diagnostic strings; successful calls use a static empty error string.

Tests count allocations on the calling thread and verify zero allocations for:

• 100 successful error guards and error-string reads, including clearing an error;
• 100 empty search/view/recycle cycles after warm-up.

These checks cover adapter overhead. Searches with hits still allocate in upstream Sassy and during per-hit CIGAR formatting. Packed CIGAR output remains the subject of issue #6.

Lifecycle tests check that recycling clears lengths and retains capacity, outstanding results survive other searches, and owned results remain valid after their searcher is destroyed. A recycled result and its borrowed views must no longer be used.

Equivalence

workload	records	bytes	hits	full_hit_multisets_equal
dna	262144	67108864	262676	TRUE
ascii	4096	1048576	524288	TRUE

Both native paths and SQL return identical full hit multisets, including coordinates, cost, strand and CIGAR. Timed runs also check count, cost sum and CIGAR bytes. The separate upstream CRISPR CLI oracle passes all 36 option profiles, covering 864 guide/record comparisons.

Reproduce

Build the AVX2 benchmark executable using the command in the adapter benchmark, then run from the repository root:

DUCKHTS_CACHE_DIR=/tmp/ducksassy-benchmark-cache R_LIBS=.deps/Rlib \
  Rscript benchmarks/adapter_overhead.R local
R_LIBS=.deps/Rlib Rscript tools/render_readme.R benchmarks/buffer_reuse.Rmd

Choose the cache directory for your installation. Raw before timings, after timings and their before receipt and after receipt retain artifact hashes, inputs and machine details. The baseline is commit ff0c363487d24eb13eb06d1cd3001d0cc49631d6; the optimized build was measured before commit and is identified by the artifact and source hashes in its receipt.