performance-analyst
/performance-analyst
Mierzy zanim zmieni: wyznacza wskaźnik, znajduje przyczynę w profilu lub śladzie i weryfikuje, czy poprawka działa dla wolnych zadań.
- Autor
- arijeetganguli ↗
- Repozytorium
- ai-kaarkhana ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 02.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 02.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Mierzy zanim zmieni: wyznacza wskaźnik, znajduje przyczynę w profilu lub śladzie i weryfikuje, czy poprawka działa dla wolnych zadań.
Oryginalny opis (EN)
Analyses performance problems with a measure-first approach: defines the metric, finds the bottleneck from profiles, queries or traces, proposes targeted changes and verifies the improvement. Use for slow endpoints, jobs or queries.
Jak działa
- Purpose
- When to invoke
- Required input
- Execution instructions
- Expected output
- Validation checklist
- Example usage
- Sample prompts
- Common mistakes
- Related skills
Podgląd SKILL.md
--- name: performance-analyst description: "Analyses performance problems with a measure-first approach: defines the metric, finds the bottleneck from profiles, queries or traces, proposes targeted changes and verifies the improvement. Use for slow endpoints, jobs or queries." license: MIT metadata: activity: performance difficulty: advanced frameworks: "any" version: "1.0" --- # Performance Analyst ## Purpose Prevent speculative optimisation. Tie every change to a measured bottleneck and a before and after comparison. ## When to invoke - A latency, throughput, memory or cost target is missed. - A review raises a performance concern without evidence. - A change adds loops, queries, serialisation or remote calls on a hot path. ## Required input - The metric and target (for example p95 latency under 300 ms at 50 requests per second). - Evidence: profile, trace, slow query log, benchmark or metrics. - The code path and data volumes. If a required input is missing, ask for it. Do not guess. ## Execution instructions 1. Define the symptom, metric, target and the load profile that reproduces it. 2. Establish a baseline measurement. If none exists, define how to take one before changing code. 3. Locate the bottleneck from evidence: where is time, memory or I/O spent. 4. Check typical causes: N+1 queries, missing indexes, unbounded result sets, synchronous remote calls in loops, repeated serialisation, lock contention, excessive allocation, missing caching with correct invalidation. 5. Propose the smallest change that addresses the measured bottleneck, with expected effect and trade-offs (complexity, staleness, memory). 6. Define the verification: same measurement, same load, compare to baseline. 7. Record regressions risks and add a guard (benchmark, query count test, alert) where practical. ## Expected output - Symptom, metric, target and measurement method. - Bottleneck analysis with evidence. - Ranked options with trade-offs. - Verification plan and before/after results (only if measured). ## Validation checklist - [ ] A baseline exists or its collection is the first step. - [ ] The bottleneck is supported by evidence. - [ ] No numbers are estimated or invented. - [ ] Trade-offs and correctness risks (cache invalidation, ordering) are stated. - [ ] A guard against regression is proposed. ## Example usage Input: > Endpoint `/orders` is slow. Trace shows 1 query for orders plus 1 query per order for customer. Illustrative output (abridged, hand-written for this guide, not captured from a model run): ```text Bottleneck: N+1 on `customer`. Evidence: trace spans (1 + N queries). Option A: join fetch (single query). Option B: batch load by ids. Choose A for page size 50. Verify: same request at the same data volume, compare query count and p95. Guard: test asserting query count <= 2. ``` ## Sample prompts ```text Act as the Performance Analyst. Symptom: [SYMPTOM]. Metric and target: [TARGET]. Evidence: [PROFILE/TRACE/QUERY PLAN]. Identify the bottleneck from the evidence only, rank options with trade-offs, and give a verification plan. Do not invent measurements. Code: [CODE] ``` ## Common mistakes - Optimising before measuring. - Adding a cache without an invalidation strategy. - Reporting improvements that were estimated instead of measured. - Benchmarking with unrepresentative data volumes. ## Related skills - [Debugging Specialist](../debugging-specialist/SKILL.md) - [Database Design Reviewer](../database-design-reviewer/SKILL.md) - [Architecture Reviewer](../architecture-reviewer/SKILL.md) ## Related resources Handbook: - [Chapter 7: Code Quality](../../handbook/07_code_quality/README.md) - [Chapter 11: Engineering Delivery and Operations](../../handbook/11_delivery_and_operations/README.md) Prompts: - [Performance Bottleneck Analysis](../../prompts/debugging/performance-bottleneck-analysis.md) - [Performance Review](../../prompts/code_review/performance-review.md) Workflows: - [Handle a production incident](../../workflows/08_production_incident.md) - [Refactor an existing service](../../workflows/03_service_refactoring.md)
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · arijeetganguli/ai-kaarkhana | 100 | MIT |
| arijeetganguli/ai-kaarkhana | 100 | MIT |
| cofin/flow | 100 | Apache-2.0 |
Migawka bazy z 13.11.2026 · wersja paczki 95dc7fc21b06 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100performance-profiling
Wskazuje metodą opartą na pomiarach prawdziwe wąskie gardło przy wolnych endpointach, dużych opóźnieniach, obciążeniu pamięci i CPU oraz zapytaniach N+1 i weryfikuje zysk. - 100performance-profiling
Znajduje i naprawia problemy wydajności backendu: mierzy p95 i p99, profiluje CPU, pamięć, I/O i blokady, a po poprawce dodaje test regresji. - 95performance-investigation
Mierzy i diagnozuje problemy wydajności na podstawie danych, gdy coś działa wolno lub regresuje czas, pamięć albo zapytania. - 100python-performance-optimization
Mierzy wąskie gardła w Pythonie narzędziami cProfile, line_profiler, memory_profiler i py-spy, zanim cokolwiek zmieni w kodzie. - 100performance-tuning
Znajduje i naprawia główny koszt wolnej usługi JVM: mierzy, profiluje JFR lub async-profilerem i sprawdza GC oraz blokady. - 100trove-perf
Diagnozuje problem wydajności względem punktu odniesienia: mierzy, poprawia, mierzy ponownie i podaje różnicę z artefaktem.