nateberkopec

Can you compile away Rails? An experiment with Roundhouse, YJIT, and Spinel

About a five-minute read.

koduki/example-rails-aot asks an interesting question: if we resolve a Rails application's framework machinery ahead of time, how much work can we remove from each request?

The answer is encouraging—and includes a classic performance trap. Generated Ruby can be dramatically faster and use less memory than Rails. But an inefficient association-loading algorithm can erase the speed advantage. Compiling that code to native machine code introduces another set of concerns: connection handling and operating-system limits.

Here's what the experiment actually compared, and what it found.

nateberkopec

The built-in code-review workflow in Claude Code 2.1.220

1. Result

Claude Code 2.1.220 contains a built-in JavaScript workflow named code-review.

The program registers this workflow with hidden: true. The workflow is inside the installed executable. It is not the public Markdown code-review plugin.

The workflow has five stages:

nateberkopec

#!/usr/bin/env ruby
# frozen_string_literal: true

require "bundler"
require "cgi"
require "json"
require "net/http"
require "optparse"
require "thread"
require "uri"

nateberkopec

Plan: Yabeda Monitoring System Adapter

Goal

Add an optional Yabeda adapter so that users who define metrics via Yabeda's DSL can have those metrics sent to CloudWatch through the existing Reporter infrastructure (async batching, StatisticSet aggregation, queue overflow protection, high-resolution support).

This gem does not collect metrics via Yabeda. The existing Puma/Sidekiq/Rack/ActiveJob

nateberkopec

#!/bin/bash
# After converting PNGs to JPEGs, update all references in the codebase.
# Parses git status for deleted .png / new .jpg pairs and does a global find-and-replace.
# Usage: ./rename_png_refs.sh [directory]

set -euo pipefail

DIR="${1:-.}"
DRY_RUN="${DRY_RUN:-false}"

nateberkopec

#!/bin/bash
# After converting PNGs to JPEGs, update all references in the codebase.
# Parses git status for deleted .png / new .jpg pairs and does a global find-and-replace.
# Usage: ./rename_png_refs.sh [directory]

set -euo pipefail

DIR="${1:-.}"
DRY_RUN="${DRY_RUN:-false}"

nateberkopec

#!/bin/bash
# Convert PNGs > 100KB to optimized JPEGs for Cloudflare Polish webp conversion
# Usage: ./png_to_jpeg.sh [directory]
# Requires: ImageMagick (convert), jpegoptim

set -euo pipefail

DIR="${1:-.}"
QUALITY=85
MIN_SIZE_KB=100

nateberkopec

#!/usr/bin/env bash
set -uo pipefail

usage() {
  echo "heroku-inventory: Catalog Heroku resources" >&2
  echo "" >&2
  echo "Usage:" >&2
  echo "  heroku-inventory --team <team-name>" >&2
  echo "  heroku-inventory --app <app-name>" >&2
  echo "" >&2

nateberkopec

Worker TERM/restart benchmark (Docker + k6)

This benchmark reproduces the scenario discussed in chat:

• 4 Puma workers
• open-model traffic at 12 req/s (k6 constant-arrival-rate)
• synthetic Rack app with ~250ms service time
• service alternates CPU and IO slices randomly every ~10ms
• compare threads=5 vs threads=25
• during traffic: kill one worker, wait for restart, let it serve ~5s, then kill a different worker (round-robin)

nateberkopec

Unslopper local run notes

This repo is just a scratchpad for running the Unslopper model locally with llama.cpp.

What worked

• Install llama.cpp via Homebrew.
• Use the GGUF Q8_0 quant (the bf16 repo is too large for typical Mac memory).
• Use llama-completion with a Qwen-style chat template prompt file and --no-conversation so it does a single completion and exits.
• Redirect logs away from stdout so the output file is clean.