Performance

What the viewer does for you

• One worker. Every document shares a single pdf.js worker (set configurePdf({ sharedWorker: false }) to opt out).
• Exact, cheap layout. All page shells exist from the start with their real size, so scroll height and jumps are exact; only pages within overscan viewport heights get a bitmap.
• Priority scheduler. Visible pages render first, nearby pages next, thumbnails last. Work for pages that scroll away is cancelled, and re-queuing a page replaces its pending job.
• Zoom without re-layout storms. A zoom frame changes one CSS variable (--pdf-scale); the stretched bitmap stays until the sharp render lands. Text layers follow through CSS and are re-measured once zoom settles.
• Sharp zoom-ins. Zoomed-out pages render at up to 2× (oversampling), and a page stretched past 1.35× during a long zoom is refreshed once.
• Bitmap cache. Pages that scroll away keep their bitmap in a shared LRU cache (configurePdf({ bitmapCacheBytes }), default 192 MB, 0 disables), so scrolling back or changing columns redraws instantly.
• Level of detail. Pages narrower than detailMinWidth (260 CSS px) — thumbnail-sized spreads when zoomed out — skip their text layer and text indexing, so four-per-row layouts stay smooth.
• Cached analysis. PaperState accepts a cache (KeyValueStore, e.g. indexedDbStore('papers')): analysis results are stored by document fingerprint and reopened papers skip analysis.
• Capped bitmaps. maxCanvasPixels (16M device pixels) bounds memory per page at high zoom.
• Shared text. Page text is extracted and indexed once, then shared by the text layer, find, selection, annotations and paper analysis.
• Thumbnails are rendered lazily, at low priority, and cached per document.

Tuning