WebSandbox Performance in 2026: What We Measured, What We Fixed

Performance is a feature. A slow IDE is a broken IDE. In Q1 2026, we ran a thorough performance audit of our proprietary WebSandbox platform using Lighthouse, Web Vitals, and custom RUM (Real User Monitoring) instrumentation deployed to 5% of production traffic.

Here's what we found and what we fixed — including the things we expected to be fast but weren't.

The Baseline Numbers

Before optimizations, our median measurements on a mid-tier laptop were:

• LCP (Largest Contentful Paint): 3.8 seconds
• INP (Interaction to Next Paint): 280ms
• CLS (Cumulative Layout Shift): 0.12
• WebContainer cold boot: 2.1 seconds
• First keystroke response in editor: 340ms

The 340ms first-keystroke latency was the most alarming. A user who opens a file and starts typing immediately experiences a noticeable freeze. This was caused by the TypeScript language server hydrating on the first edit.

Key Architectural Optimizations

• Asynchronous FileSystem Snapshots. We overhauled how we write to Dexie via bulk-operations.ts using a sophisticated withFileRepository queueing system. This prevents heavy file changes from locking the main UI thread during autosaves.
• Web Worker Global Search. Global search on thousands of files historically choked the browser. We moved this entirely to a dedicated Web Worker (use-global-search.ts). We explicitly tuned the SEARCH_WORKER_CHUNK_SIZE to process 50 files per chunk, yielding the best balance between search speed and main thread smoothness. We also strictly enforce worker termination and listener cleanup to eliminate memory leak-induced INP spikes.
• LSP prewarming. We now start the TypeScript Worker exactly when the workspace first loads rather than waiting for the first edit. First-keystroke latency dropped from 340ms to 18ms.
• Virtualized file tree. Replaced our flat list rendering with react-window, eliminating janky scrolling on projects with thousands of files, keeping a smooth 60fps framerate.
• Debounced LSP diagnostics. Throttled the TypeScript diagnostics cycle from per-keystroke to a 300ms debounce. INP improved from 280ms to 90ms.

Results After

After shipping these proprietary architectural changes, our median measurements became:

• LCP: 1.2 seconds (↓ 68%)
• INP: 90ms (↓ 68%)
• CLS: 0.004 (↓ 97%)
• WebContainer cold boot: 0.6s warm (↓ 71%)
• First keystroke: 18ms (↓ 95%)

What's Next

Our next focus is reducing memory usage for enterprise-scale projects. A workspace with 10,000 files currently uses ~180MB of RAM for the virtual filesystem mirror alone. We're exploring a lazy-load approach where only the directory tree is loaded upfront and file contents are fetched on demand.