Data architecture · 2026 · Confidential
The third orientation no one talks about. Data that lives at a perfect 45° angle to both your rows and your columns.
Context / the only two options
Nobody voted on this. There was no RFC. The right angle got in early and it has been coasting ever since.
The gap / missing middle
Every paper, every blog, every conference talk. Never 45°. We have checked. We have a spreadsheet, and the spreadsheet is rectangular, which is part of the problem.
The diagonal of your warehouse is empty air right now. You are paying for it. You are scanning past it. You have never once looked directly at it.
If it does not have a name, it does not exist. That is the actual mechanism by which whole categories of engineering stay invisible. We fixed the name first.
The solution / 45° or bust
Shard by key and time in a single elegant cut. The diagonal is the only line that moves through both axes at once, which is a fact about geometry that the industry has declined to act on.
Most of the matrix is empty by design. Not compressed — empty. Cache efficiency goes brrr. See why this is not a bug report.
Your data warehouse finally looks like modern art. This is listed third but it is, in the honest accounting, why the project exists.
Mechanics / implementation notes
"k">def diagonalize(matrix, bandwidth=1, angle=45):
"""Retain the band. Discard the rest."""
"k">for r, row "k">in enumerate(matrix):
"k">for c, value "k">in enumerate(row):
"k">if abs(c - r) <= bandwidth:
yield (r, c, value)
"k">else:
discard() "k">class="c"># the rest is just noiseControl the width of the band. Wider = more data, less pure. Deprecated in draft 4 for reasons we have documented.
You can rotate the diagonal. 30° and 60° are still under peer review.
Diagonal data compresses better because consecutive values are related. This is the one claim on this website that is straightforwardly true.
Applications / real world (ish)
Features that evolve over time but only near the present. A perfect diagonal band, and you built it by accident to satisfy a retention policy.
Local neighborhoods in ordered graphs. The diagonal is the only part that matters and the ordering was chosen to make that true.
Users who like item i also like item j when |i-j| is small. Coincidence? No.
Correlations decay with distance in instrument space. Diagonal risk matrices FTW. Every risk desk knows this and stores the full matrix anyway.
Nearby pixels. Diagonal in flattened space. You are already doing it. You have been doing it the whole time.
Finally something new to put on the slide after “vertical vs horizontal scaling.” Results have been mixed.
Next steps / the revolution starts now
The data was always there. We just needed the right angle.
Mailing list
Roughly monthly. Occasionally an incident report. We have eleven subscribers and we know all of your names.
Questions? · Criticisms? · Pull requests?