[paddymul]

Problem

Opening a cache-worthy catalog entry (Aggregate/Join/Sort) in the viewer recomputes the entire expression DAG on every page interaction — once for the row count and again for each row window — instead of reading the materialised result.parquet the build already wrote.

BuckarooManager.ensure_session posts the entry's raw recipe build dir to /load_expr. That dir is the full recipe (from_catalog(...).group_by('plate').agg(...) over the multi-million-row source) with no read of result.parquet. So every expr.count().execute() and every LIMIT 300 window re-runs the whole DAG against datafusion. For an Aggregate, LIMIT 300 can't push down past the group-by, so even the first 300-row page pays the full aggregation. Per the profiling in #71, Aggregate/Join recipes cost 0.9–1.3s on each of count and window, tracking input size, not output.

Fix

In ensure_session, for a cache-worthy entry (cache_worthy(project, content_hash)), hand Buckaroo a build whose expression is a deferred_read_parquet of the materialised result rather than the raw recipe:

• New ensure_result_build (in result_cache.py) calls ensure_result(...) to guarantee the parquet is present, then builds a tiny portable deferred_read_parquet(result) build, marker-gated and lock-serialized like ensure_expanded_build.
• ensure_session expands that build into a dedicated .xorq_result_build_expanded dir and posts it. Cheap entries keep loading the recipe (their recompute ≈ a parquet read, so indirection buys nothing).
• If result-build prep fails for any reason, it falls back to the recipe — a cache hiccup must not take down the detail page.

Push-down still holds: a parquet scan pushes filters/limit/sort down too, so reading result.parquet keeps the push-down benefit while dropping the re-aggregation.

Tests

• test_unit_ensure_session_cache_worthy_serves_result_read — the posted build classifies as a cheap parquet read (classify_build worthy=False) while the recipe stays worthy=True. Seen failing on CI before the fix.
• The two expansion-mechanism tests now use a cheap entry so they continue to exercise the recipe-expansion path.

Closes #71.

🤖 Generated with Claude Code

[paddymul]

no longer relevant after 74