Measure on your machine

The numbers on the overview come from the numbers example. Run it on your own machine from a clone of the repository:

cargo run --release --example numbers

It needs a release build: a debug build is many times slower and says nothing about what an app will see. It writes its databases under the system’s temporary folder, in pito-sync-numbers/, and removes them when it ends. This is one run on a busy Linux machine:

open, a new database               115.58 ms
open, an existing database         109.9 µs
Resync to its closing Changed      28.8 µs (p99 395.9 µs)
a pass of 10,000 rows              1.86 ms (5.37 M rows/s)
a UI read, writer idle             8.7 µs
a UI read, writer busy             36.7 µs (p99 147.3 µs, max 8.18 ms; 2530 reads, 670 passes)
  SQLite defaults, writer busy     147.7 µs (p99 3637.15 ms, max 3637.15 ms; 5 reads, 76 passes)

What each line times

Every figure is a median unless it says otherwise.

Line	What it does
open, a new database	open on a path that doesn’t exist yet: the folder, the file, WAL and the schema
open, an existing database	open on a file already built at the right version
Resync to its closing Changed	a Resync sent to a running source, until its pass’s closing Changed arrives
a pass of 10,000 rows	a commit that deletes and inserts 10,000 rows with the sync row
a UI read, writer idle	a page of 50 rows and the sync row through reader’s connection
a UI read, writer busy	the same read every millisecond for three seconds while a source commits 10,000-row passes back to back
SQLite defaults, writer busy	the same reads and writer on plain connections with SQLite’s defaults: rollback journal, synchronous=FULL

The last two lines are the chart in read from the UI. Your numbers depend on the CPU, the disk, the filesystem and what else the machine is doing; the run above and the overview’s were both made while other builds were running.

The steps compared

The chart is the overview’s run. The first open of a new database is the slow one, at over 100 ms in both runs, and it happens once per install or schema version; every open after that takes well under a millisecond.