lift_skill

Overview

This skill delivers measurement-driven performance optimization focused on latency, throughput, memory/GC, and tail behavior while preserving correctness. It defines a clear performance contract, picks a proof workload, profiles to find real bottlenecks, runs tight experiments, and ships guarded changes with regression checks. The workflow emphasizes reproducible numbers, small reversible diffs, and risk-aware trade-offs.

How this skill works

Lift inspects the code and repository for runnable workloads, existing benchmarks, and hot paths, then builds or selects a proof harness. It captures environment and baseline metrics, profiles to classify the bottleneck (CPU/memory/I/O/lock/tail), and applies the highest-leverage levers (algorithm, data layout, concurrency, I/O) one change at a time. Every suggested change is measured, validated for correctness, and recorded with a regression guard and benchmark added.

When to use it

• You need lower latency (p95/p99) or tighter tail behavior
• You must increase throughput or QPS for a service or batch job
• High CPU usage, allocation rate, GC pauses, or memory blowups occur
• You suspect contention, lock hot paths, or concurrency limits
• You want a reproducible, measurement-first optimization process

Best practices

• Measure before and after every change with environment + command recorded
• Profile or trace first; optimize the true bottleneck, not guesswork
• Change one variable at a time; keep diffs minimal and reversible
• Avoid micro-optimizations until algorithmic and data-layout wins are exhausted
• Require passing correctness signals before accepting any perf regression

Example use cases

• Speed up a JSONL-heavy query pipeline to reduce p95 latency and lower GC pressure
• Increase QPS of a Python service by identifying and fixing a serialization hot path
• Diagnose and reduce tail latency for a request handler by addressing variance sources
• Reduce peak RSS and allocation rate in a data-processing job to avoid GC-induced pauses
• Create a minimal benchmark and guard to prevent future regressions after a perf patch