Public architecture graph of apache/airflow @ 37b6b04e — 769 files, 11 651 imports, no embeddings

[alexisromtz-lab]

I extracted a file-level import graph of apache/airflow at commit 37b6b04e7233ec47c193915d6cff04dac742d5fd and published it as a sanitized JSON. Every number below is fully checkable against the bytes.

Asset: public_graph_airflow_37b6b04e.json — 2,132,257 bytes — sha256 98c4a676fa825c55482b60fbdc78a88ee9ed46f425499bbd0839040b951df5c1
Release: https://github.com/alexisromtz-lab/platino-data-room/releases/tag/arch-index-20261029-batch2

What is in the file

Two arrays and nothing else:

• archivos: 769 entries of {rel, locus, sha256} — repo-relative path, a digest of the declared root, and the file digest.
• imports: 11,651 entries of {rel, modulo, nombre, modalidad} — which file, which module, which symbol, and how it was written (requerida 10,748 · type_checking 853 · except_import 50).
• n_ilegibles: 0. Every file in the ingested perimeter parsed successfully.

There are no embeddings in the file — no vectors, no model, no tokens. What ships is pure structural coupling.

What it costs

2,132,257 bytes for the entire perimeter. The 769 Python files it maps are 6,300,100 bytes at that commit. The graph is 33.84% of the bytes it describes (66.16% less), at roughly 2,772 bytes per file.

To put this in perspective, this is not a claim that the source can be deleted. It is the cost of the index that sits beside it. It is the number that decides whether a full architecture audit runs on a laptop or requires a cluster. And because there are no vectors, when the code moves, there is nothing to re-embed. The file is simply regenerated by parsing.

The perimeter, stated instead of implied

The graph maps 769 files, not the 8,486 .py files the repository has at that commit. The ingested root is airflow-core/src/airflow. The public projection deliberately omits the declared-roots list, so I am telling you exactly where the boundary sits rather than implying the graph covers the whole tree.

Two metrics you can verify

• Fan-in: 321 of the 769 files import airflow.api_fastapi.common.parameters; 235 import airflow.sdk.
• Fan-out: models/taskinstance.py pulls 196 distinct modules; jobs/scheduler_job_runner.py pulls 180.
  (These are in-degree and out-degree over declared bytes, not quality judgments).

Cycles
The published JSON includes the signed SCC layer (graph-input/1+scc/2). Its declared definition is a component with more than one node or a self-loop. The asset reports 2 cyclic SCCs, with 363 modules in the largest. The component list is complete for this asset (sccs_truncated: false).

Why post it here
I am building deterministic AST extractors. No ML, no LLM, no service. I am publishing the output for ten major open-source repositories so the people who own the code can attack the method.

Airflow is particularly interesting: 143 pyproject.toml files mean 143 internal distribution boundaries. A boundary is precisely what a text search cannot cross.

If the extraction is wrong somewhere — a missed import, a misread modality, a resolution that lands on the wrong file — I want the diff. I will re-issue the graph with the fix, and its new sha256, right here in this thread.

— Alexis Román, Founder, Platino Labs