How we made search ten times faster

Shape: How it actually works
Plates: lede · magnitude · steps · note
Why: A deep dive fails by starting with the fix. The reader has no reason to care about your clever index until they know where the time was going, so `magnitude` comes first and does it in one figure — a list of durations in prose reads as four similar numbers. Then `steps` for what changed, in the order it happened. Keep the honest note at the end: a deep dive with no caveat reads as marketing.

Search took 2.4 seconds. It now takes 230 milliseconds, and almost none of that came from the part we spent the first fortnight on.

Where the 2.4 seconds actually went

Parsing the query

3 ms

Ranking the results

95 ms

Rendering

140 ms

Fetching the documents

2.16 s

10⁰10¹10²10³10⁴

Each mark is ten times the last

We had assumed the ranking was the cost. It was four per cent of it. The database was doing the same lookup nine hundred times.

What we changed

Measure

One trace, in production, on a slow query rather than a fast one

Batch

Nine hundred lookups became one

Cache

The documents that never change, kept

Re-measure

The ranking is now the slowest part, at 95ms

This is a sample. To author the same thing about your own subject, just name it. Your assistant can read this template, and will write you the same document about what you know.

Using the Ario tools, make me the "Technical deep dive"
template for the migration we finished last quarter.
Show me the shape before you build anything.

New here? How to connect one.

Other samples