880 thousand
rows in the test company’s biggest table, 0.6 GB in all
10 of 10
pages answer in under 300 ms at 5.25 million rows
129.2 s
to scan 4,000 tables, with no rows leaving the database
103.4 s
for an ordinary night at 5.25 million rows, with no model calls
Everyday pages
The same speed at both sizes.
Every page an owner opens, timed at 205 thousand rows and at 5.25 million rows. The slowest took 147 ms, against our budget of 300 ms.
205 thousand rows5.25 million rows
Every page, at both sizes
Median server time of each page, loaded 15 times, in milliseconds. Your browser and network add their own. Lower is better. The scan page at 4,000 tables, timed at 5.25 million rows only: 147 ms.
An ordinary night, step by step
The whole night took 28.2 s at 205 thousand rows and 103.4 s at 5.25 million rows, and replays to the same result, byte for byte.
Why it stays flat
Pages never touch your database.
Every page reads results the nightly run stored, so a page costs the same at any size. The work runs where your data is: your database computes the totals through a read-only role, and only totals come back. Each night reads what changed, plus the last 35 days again for late corrections.
Stored, not recomputed
A page opens from the run’s stored results. It never queries your systems.
Totals, not tables
The SQL runs on your database; only totals come back. Your tables are never copied, and personal columns are never read.
Only what changed
Each night reads new and changed rows, not the whole history again.
What a night costs your database
A night of reading, at 5.25 million rows.
Read-only, with a 120 s limit on every query. A role that can write is refused.
| One ordinary night | MySQL | PostgreSQL |
|---|---|---|
| Read-only queries | 92 | 59 |
| Rows that came back | 3,10,333 | 1,26,299 |
| The longest query | 8,519 ms | 681 ms |
How we measured
The sample company as seeded (205 thousand rows read a night) and copied 35 times (5.25 million rows), loaded into PostgreSQL and MySQL; the real nightly pipeline, then each page’s server time on a signed-in app. Measured 13 Nov 2026 on an AWS t3a.large in Mumbai, the type the live app runs on, by scripts/measure_site.py.
Othor AI is in a pilot: bigger databases, slower servers and many companies at once haven’t been tested yet.