Making the Slow Thing Fast

A conference talk · The Midnight theme

Measure first

p99 latency, 10k requests, warm cache
Change	p50	p99	RPS
Baseline	18 ms	240 ms	4,100
Pooled connections	14 ms	121 ms	5,900
Batched writes	11 ms	58 ms	8,400
Both	9 ms	31 ms	11,200

Numeric columns get class="num" so they right-align and use tabular figures — the digits line up column-wise.

for (const id of ids) {
  // One round trip per id. This is the bug.
  const row = await db.get('users', id);
  results.push(shape(row));
}

Syntax colors come from .tok-* spans. Nothing highlights at runtime — you write the spans, or emit them from Shiki at author time.

const ids = req.query.ids.split(',');
const rows = await db.getMany('users', ids);
const results = rows.map(shape);

One call. N round trips become 1.

.line.hl calls out the line that matters; .line.dim pushes the context back. Both are optional — plain <pre><code> is still fine.

Regression tests, not vibes

assert.ok(
  p99 < 50,
  `p99 was ${p99}ms`
);

Premature optimization is the root of all evil.
— Donald Knuth, Structured Programming with go to Statements, 1974

Questions · Slides built with Dais