All posts
Engineering

How We Think About Scale

Every plugin in our catalog needs to handle a startup's first 100 users and an enterprise's 10 million. Here's the architecture that makes that possible.

By Engineering Team

Every plugin in our catalog needs to do something difficult: work perfectly for a startup's first 100 users and an enterprise customer's 10 million — without any configuration changes in between.

Here's how we think about that.

Scale is a spectrum, not a threshold

The common mistake is treating "scale" as a binary. You're either "scaled" or you're not. In reality, the interesting engineering happens in the middle — at 10,000 users when you still care deeply about developer experience but latency is starting to matter.

We design each plugin to handle the full spectrum by default. That means:

  • Auto-scaling infrastructure that adjusts without operator intervention
  • Read replicas and caching layers that activate as traffic grows
  • Graceful degradation when upstream dependencies are slow or unavailable
  • Per-tenant rate limiting that protects all customers from noisy-neighbor effects

LingoLens as a case study

When we built LingoLens, the hardest problem wasn't the AI translation quality. It was delivery.

A startup with 50 users can tolerate a 200ms p99 on string fetches. A gaming platform with 2 million concurrent sessions cannot. Both need the same plugin.

Our solution was a three-layer delivery architecture:

  1. CDN edge cache — strings served from 200+ PoPs globally, sub-10ms for cached content
  2. Regional hot cache — Redis clusters per region for strings updated in the last 24 hours
  3. Origin with read replicas — PostgreSQL read replicas in four regions, writes always go to primary

The result: p99 latency under 15ms at any scale, with zero configuration required from the developer.

What we're working on next

We're currently investing heavily in our observability layer — every plugin now emits OpenTelemetry-compatible traces that you can pipe into your existing Datadog, Honeycomb, or Grafana stack.

We'll share more details on that in a future post.