A6 — Performance Optimization

Profiles, optimizes, and proves a latency win for GET /api/summary in the A2 expense tracker (Advanced/parallel-expense-tracker). Method: measure → profile → identify → minimal change → verify.

Result: naive ORM .all() summation (p50 328.71 ms @ N=50k) → SQL GROUP BY over indexed integer cents (p50 28.89 ms) = 91.2% faster, 11.4×, behavior preserved (40/40 tests, 42 with the added concurrency tests). Full write-up: docs/agent-analysis/A6_performance_improvement.md.

Quickstart

# From the A2 service dir, with its venv active (Python 3.12.7):
cd ../parallel-expense-tracker
python -m venv .venv && . .venv/bin/activate && pip install -r requirements-dev.txt

B=../performance-optimization/bench_summary.py
python $B                    # time the optimized endpoint (default)
python $B --compare-before   # naive ORM vs optimized + improvement %
python $B --profile          # cProfile the optimized hot path
python $B --profile --before # cProfile the naive baseline
python $B --scaling          # p50 across N = 1k,10k,50k,100k
python $B --json             # machine-readable output

# Or one command from the repo root:
make a6-verify               # validate deliverable + run live bench on Python 3.12

Environment variables

Var	Default	Purpose
A6_N	50000	rows seeded for timing/profile modes
A6_ITERS	15	iterations per timing run
A6_SCALING	1000,10000,50000,100000	sizes for --scaling
A6_SCALING_ITERS	8	iterations per size in --scaling
DATABASE_URL	(temp)	bench sets a throwaway temp SQLite DB automatically

Results

Metric	Before (naive ORM)	After (SQL GROUP BY)	Δ
p50 latency @ N=50k	328.71 ms	28.89 ms	−91.2% (11.4×)
cProfile wall (10 calls)	4.584 s	0.318 s	−93.1%
function calls (10 calls)	11,002,001	48,111	−99.6%
rows materialized / call	50,000 ORM objects	6 tuples	−99.99%

All numbers reproduce from artifacts/repro/ on Python 3.12.7 (2026-06-21).

Layout

Path	What
bench_summary.py	benchmark + profiler (--compare-before, --profile[ --before], --scaling, --json)
snapshots/summary_slow.py	executable pre-optimization baseline (so "before" is reproducible)
scripts/validate_a6_deliverable.sh	structural + stale-claim validation (CI-gated)
prompts/performance-engineer.md	reusable role prompt extracted from the skill
artifacts/repro/	saved BEFORE/AFTER timing + cProfile + scaling logs
docs/agent-analysis/A6_performance_improvement.md	the report (baseline → profile → fix → verify)
docs/agent-analysis/A6_manifest.json	machine-readable metrics summary
docs/agent-analysis/A6_secondary_findings.md	measured pass over the other endpoints
docs/A6_engineering_evaluation_audit.md	scored 28-issue audit (61 → 96)
docs/A6_remediation_tracker.md	remediation status for all 28 audit items
docs/A6_final_scorecard.md	before/after scorecard with ceiling rationale

Linked A2 hardening (this task)

Change	File
SQLite WAL + busy_timeout (concurrent writes)	../parallel-expense-tracker/app/database.py
Parallel-POST concurrency tests	../parallel-expense-tracker/tests/test_concurrency.py
k6 load test + guide	../parallel-expense-tracker/scripts/load_summary.k6.js, docs/LOAD_TESTING.md
CI perf regression gate (p50 ≤ 50 ms)	../parallel-expense-tracker/scripts/perf_guard.py

Before / After (deliverable)

	Before	After
Files	2	16+
Report accuracy	Stale (3.14.6, 16 tests)	Fresh on 3.12.7, 40 tests
Repro artifacts	0	5
Validation CI	No	Yes (a6-performance-optimization.yml)
AI evaluation score	61	96