Speed up Dag serialization by skipping redundant asset roundtrip

[shahar1]

Problem

_DependencyDetector.detect_task_dependencies builds the asset dependency_id
for every asset outlet of every task via:

serialized_asset = ensure_serialized_asset(obj)
dependency_id = SerializedAssetUniqueKey.from_asset(serialized_asset).to_str()

ensure_serialized_asset() runs a full decode_asset_like(encode_asset_like(obj))
encode→decode roundtrip — rebuilding the entire SerializedAsset (group, extra,
watchers, access_control) — purely to read back .name/.uri for the unique key.

Fix

encode_asset_like copies name/uri verbatim and _decode_asset reconstructs
them verbatim, so the roundtrip cannot change either field. The key is built
directly from the object instead:

dependency_id = SerializedAssetUniqueKey(name=obj.name, uri=obj.uri).to_str()

One full encode+decode per asset outlet is removed. The asset-alias branch is
left unchanged (rare path that genuinely needs the serialized object).

Correctness

• Output is byte-identical — URI normalization happens once at
  Asset.__init__, never inside the roundtrip, so obj.name/obj.uri already
  hold the normalized values.
• All 194 tests in
  airflow-core/tests/unit/serialization/test_dag_serialization.py pass.
• The benchmark's [0] section asserts identical keys across asset shapes
  (name+uri, uri-only, name-only, special characters, with group/extra).

Benchmark

Comparing two separate process runs proved unreliable on a loaded machine
(absolute timings swung ~2× from background load alone). The credible
measurement is a single-process interleaved A/B that patches in both
implementations and times them round-by-round under identical instantaneous
load (bench_asset_roundtrip_ab.py):

Scenario	OLD (roundtrip) min	NEW (direct) min	speedup
100 tasks × 1 outlet	10.73 ms	10.46 ms	1.03×
100 tasks × 5 outlets	15.28 ms	14.52 ms	1.05×
500 tasks × 5 outlets	77.24 ms	70.82 ms	1.09×
1000 tasks × 5 outlets	143.46 ms	137.79 ms	1.04×
200 tasks × 20 outlets	59.53 ms	54.79 ms	1.09×

Consistent ~3–9% end-to-end serialization speedup (largest on asset-heavy
Dags), and ~1.5–2× on the changed line in isolation.

Benchmark scripts: https://gist.github.com/shahar1/841592531adfd66def64fc67fcc3ea6c

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Opus 4.8)

Generated-by: Claude Code (Opus 4.8) following the guidelines