0591 — the sandbox's Postgres is not prefetched
Edit this page
Context
The roadmap carried a PGlite wasm prefetch: a requestIdleCallback fetch of
the sandbox’s wasm after interactive, so its first statement is near-instant.
It was built, measured, and reverted.
The assets are four, not the one chunk the item named: pglite.wasm
(10.09 MB raw / 3.39 MB gzipped), pglite.data (6.30 MB / 1.86 MB),
initdb.wasm (395 KB / 145 KB), and the PGlite JS chunk (602 KB / 138 KB) —
17.4 MB raw, 5.53 MB over the wire, plus a 10 MB wasm compile.
Two findings from reading the code first:
sandbox.sveltealready boots the database inonMount, and the one sandbox on the site sits at the fold of the homepage underclient:visible. The observer fires on about the first frame, soclient:idlewould fire LATER — queued behind the page’s three other idle islands.- What actually delays the download is a serial module waterfall: the island
chunk and the Svelte runtime, then
database.ts, then the 602 KB PGlite chunk, each a round trip, and only then doesPGlite.create()ask for the 16 MB.
Decision
Nothing prefetches. The sandbox keeps booting in onMount off
client:visible, and PGlite keeps fetching and compiling its own files.
The version that was built and reverted, for anyone tempted to rebuild it: a
page-level <script> on the homepage that compileStreamings the three files
on idle, handed to PGlite.create() through its pgliteWasmModule,
initdbWasmModule and fsBundle options — which it does skip its own download
for. It worked. It was reverted on the ratio, not on a defect.
What it cost, against a saving of two round trips and a 602 KB parse on a 5.53 MB download:
- A module reaching into
node_modulesby relative path for its?urlimports, because the package’sexportsfield does not name the files and the package-specifier form fails to resolve. - The first page-level
<script>in the tree, load-bearing on being written above<BoardLayout>— written below it, Astro renders the tag at the end of the body instead of hoisting it into<head>. - A second place that knows how PGlite starts, holding a compiled 10 MB
WebAssembly.Moduleand a 6 MB Blob for a page the reader may never scroll into, with its own failure-and-retry rule so Reset sandbox still works.
Consequences
- The gain was never measured in a browser — it was reasoned from the chunk graph and the built output. Reopening this needs a real measurement first, not another estimate.
<link rel="prefetch">is not the cheaper way back in: the site sets no cache headers, and it would fetch everything twice underastro dev.- The sibling
transition:persistitem is the better lever and is untouched.ClientRouteris in the stack list but is not actually wired up, so PGlite’s module-level caches die on every navigation — keeping the live instance across pages would save a whole boot, where this saved part of one.