DEV Community

zenlesscodes
zenlesscodes

Posted on

I Made My ZZZ Code Site 100x Faster by Removing Flask

#python #webdev #performance #beginners

Remember that ZZZ code aggregator I built? Well, I looked at my VPS metrics and realized something dumb: I was running Python on every single request for data that updates once per hour.

So I ripped out Flask entirely. Here's what happened.

The Problem

My original setup was standard Flask + Gunicorn behind Nginx:

Request → Nginx → Gunicorn → Flask → Response
Enter fullscreen mode Exit fullscreen mode

It worked fine. But the site was just... rendering the same JSON data every time. The codes only change when HoYoverse drops new ones (roughly weekly, sometimes monthly). Running a Python process for every visitor felt wasteful on a $3 VPS.

Memory usage hovered around 100MB. Not terrible, but not great either.

The Fix: Static File Generation

The new architecture is embarrassingly simple:

Request → Nginx → static file (HTML/JSON)
Enter fullscreen mode Exit fullscreen mode

A background daemon generates static files every hour. Nginx serves them directly. That's it.

The key insight: if your data doesn't change between requests, don't compute it between requests.

The Code That Made It Work

Atomic File Writes

The trickiest part was ensuring users never see a half-written file. Linux gives us atomic renames, so:

def atomic_write(filepath: Path, content: str) -> None:
    temp_path = filepath.with_suffix(".tmp")
    try:
        temp_path.write_text(content, encoding="utf-8")
        temp_path.rename(filepath)  # Atomic on Linux
    except Exception as e:
        if temp_path.exists():
            temp_path.unlink()
        raise e
Enter fullscreen mode Exit fullscreen mode

Write to a temp file, then rename. The rename is atomic - a reader either gets the old file or the new file, never a partial write.

Scheduler Config

APScheduler handles the hourly updates, but I needed to prevent overlapping runs if one takes too long:

scheduler.add_job(
    update_codes_task,
    "interval",
    minutes=60,
    max_instances=1,    # Prevent overlapping
    coalesce=True,      # Combine missed runs
)
Enter fullscreen mode Exit fullscreen mode

max_instances=1 ensures only one update runs at a time. coalesce=True means if the server was down and missed 3 runs, it only runs once when it comes back up (not 3 times in a row).

The Results

Metric Before (Flask) After (Static)
Memory ~100MB ~20MB
Requests/sec ~100-500 ~10,000+
Dependencies flask, gunicorn jinja2

That requests/sec number is from Nginx serving static files directly. Your mileage will vary based on file size and server specs, but the point is: it's way faster.

Bonus: Image Optimization

While I was at it, I converted all the reward icons from remote PNG URLs to self-hosted WebP files:

const ICON_MAP = {
  'Polychrome': '/static/e6ee639872c119aa6895758f3a755d3b.webp',
  'Denny': '/static/7db931d2138edcfb9e155907503f2fbe.webp',
  'Senior Investigator Log': '/static/ab0406e53b7f8c4afe08096a2f7aa587.webp',
  // ... 11 reward icons total
};
Enter fullscreen mode Exit fullscreen mode

The annoying part? I was loading these from game8.co on every page view. Now they're:

  • 25-34% smaller (WebP compression vs PNG)
  • Self-hosted (no external dependencies)
  • Content-hashed for cache-busting

I kept PNG fallbacks for older browsers, but modern browsers get the lighter WebP versions.

Unexpected Benefits

Crash-proof: If my Python daemon dies, the last-generated files keep getting served. Users see stale data (worst case: 1 hour old) instead of an error page.

Simpler deployment: No WSGI server to configure. Just run the Python script as a systemd service and point Nginx at a directory.

IndexNow integration: When codes actually change, I ping search engines. But it only happens when there's real new content, not on every request.

The Takeaway

The best optimization is often eliminating unnecessary work entirely.

Before: Run Python interpreter → load Flask → route request → fetch cached data → render template → return response.

After: Return file.

This pattern works whenever your data updates less frequently than your traffic. Blog posts, documentation sites, dashboards with hourly data - all good candidates.

Cost Breakdown

Still the same as before:

  • Domain: ~$10/year
  • VPS: ~$3/month
  • Cloudflare: Free tier

zenlesscodes.com - still aggregating ZZZ codes, now 100x faster at it.

Top comments (0)