[frankmcsherry]

Since #1024 and #1025, a program reads every import as of its as-of time (e - 1, []), and runs at times at or beyond it. So every time a half join's arrangement presents is at or beyond the as-of time.

• Strict half joins: an update at exactly the as-of time meets only arranged rows at that time, and a strict half join meets none.
• Along a path: an update keeps its own time along a chain of half joins. Once every path from a lifted stream reaches a strict half join, its updates at the as-of time produce nothing.
• In a delta join, that is every path but the one that is non-strict throughout. Today those paths carry the imports' whole history, only to discard it.

This is a certain rewrite, so it is automatic. Install runs it after the projection step, with the program's as-of time (interactive/src/as_of.rs).

• Where it filters: it drops those updates where the stream is lifted, with a filter on the time lift writes. Any backend runs it with the operators it has.
• What may sit between the lift and the strict half join: only time-preserving stages: project, filter, negate, lift_iter, flat_map. Any other reader makes the stream live.
• Shared lifts: when other readers share a lift, as both paths of a self-join do, the lift keeps its rows, and the filter goes only on the edges to the readers it is dead for.
• Nested scopes: it applies there too, where the as-of time has zero coordinates added.
• Safety net: shapes are inferred again afterwards, and install falls back to the program as written if they disagree.

Tests

• Unit tests in as_of.rs:
  • a strict-only path filtered;
  • a three-way delta join with 2 of its 3 paths filtered;
  • a shared lift filtered on its strict edge only;
  • no filter for a non-strict path, a delay before the half join, or an exported lift;
  • nested scopes comparing every lane of the time.
• server_late_import.rs: a three-way delta join over three imports, installed at epochs 0, 7 and 12, with column and arranged imports, matches the plain join. If the rule is deliberately broken so non-strict paths drop too, this test and the existing self-join delta test both fail.
• Crate tests: cargo test -p interactive -p ddir-server: 211 passed. cargo check --all-targets: no warnings. Clippy: nothing new.
• Difftest: release, 1,000 cases with all nine oracles: all passed. delta,split,bind,reinstall, 2,000 seeds with late installs: all passed.

The saving

A three-way delta join over imports with eight epochs of history: three inputs, 50,000 rows each per epoch, with retractions, installed late. Install plus first tick, three alternated runs each:

keys	with the drop	without
100,000 (dense, about 1.15M join rows)	696–713 ms	783–795 ms
1,000,000 (sparse)	61 ms	124–125 ms

Steady ticks are unchanged, and the outputs are identical. The probe is kept out of the tree.

🤖 Generated with Claude Code

https://claude.ai/code/session_019XsakwFGwtESwZ1gSYwG4o