fix: handle param overlap between model and its dotted child

[ilongin]

Fixes #1883.

When .map(params=[...]) (or to_iter(...)) asked for both a model and one of its dotted children, e.g. params=["fr.name", "fr"], we hit IndexError when building the UDF input row.

Root cause: to_udf_spec dedupes shared columns (dict keyed by db-flat name), so the row has one copy of the shared leaf. row_to_objs walked a linear pos counter and expected two copies. row_to_features has the same shape - not user-visible today because db_signals doesn't dedupe, but same fragile pattern.

Both now build a {db_name -> row_index} map and look up each value's columns by name.

• row_to_objs: positions built from to_udf_spec().
• row_to_features: positions built from db_signals() (dupes still work - dict overwrite is a no-op since both indices hold the same column value).
• New helper _sub_row_for_model assembles the per-model sub-row, handling the _type_tag sentinel for nested Optional[Model].

Tests: unit tests for row_to_objs and row_to_features covering scalar/nested/optional overlap, plus e2e coverage via .map() and to_iter.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Copilot]

Pull request overview

Fixes parent/child parameter overlap during row hydration.

Changes:

• Replaces positional unpacking with DB-column name lookups.
• Adds shared model sub-row assembly with optional-model sentinel support.
• Adds unit and end-to-end overlap tests.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
src/datachain/lib/signal_schema.py	Implements name-based row hydration.
tests/unit/lib/test_signal_schema.py	Tests overlapping schema values.
tests/unit/lib/test_datachain.py	Tests .map() and to_iter() behavior.

Suppressed comments (1)

src/datachain/lib/signal_schema.py:765

• This recomputes db_signals() and rebuilds the positions dictionary for every yielded row, although to_iter() already computes the same DB signal list once before entering its row loop. On large iterations this adds a full schema traversal and dictionary allocation per record; compute the mapping once per schema/iterator and reuse it.

        positions: dict[str, int] = {
            str(name): i for i, name in enumerate(self.db_signals())
        }

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[cloudflare-workers-and-pages]

Deploying datachain with    Cloudflare Pages

Latest commit:	a63f6de
Status:	✅  Deploy successful!
Preview URL:	https://19390d81.datachain-2g6.pages.dev
Branch Preview URL:	https://ilongin-1883-fix-param-overl.datachain-2g6.pages.dev

View logs

[shcheklein]

AI generated as an attempt to simplify this:

_parse_row() already has a RowDict keyed by these flattened DB names, so converting it to a list and then reconstructing name→position caches seems avoidable. Could we pass that mapping directly to row_to_objs() and gather model subrows by their ordered DB column names? For sequence compatibility, row_to_objs() could normalize sequences with dict(zip(self.to_udf_spec(), row, strict=True)).
This would remove the four position caches and _model_row_positions(), while leaving row_to_features() and _hydrate_model() unchanged. Feature rows retain duplicate columns and already handle parent/child overlap correctly; the new to_iter/row_to_features overlap tests also pass on main.

[shcheklein]

Please check if we can simplify the implementation