Contents

• PLAN: Multi-Table Delta Batching (B-3)
  • Status: 🟢 DONE
  • Overview
  • Implementation Sequence
    • 1. Property-based Correctness Proofs (B3-3) ✅ Done
    • 2. Merged-Delta Generation (B3-2) 🔴 Not Started
    • 3. Intra-query Delta-Branch Pruning (B3-1) 🟢 Done

PLAN: Multi-Table Delta Batching (B-3)

Status: 🟢 DONE

Goal: v0.9.0 (Milestone 3) Remaining to do: - [x] B3-3: Write property-based correctness proofs for simultaneous multi-source changes (diamond-flow scenarios) using Z-set weight aggregation (SUM(weight)) instead of DISTINCT ON deduplication. - [x] B3-2: Implement merged-delta generation logic using Z-set block aggregation (SUM(CASE WHEN action THEN 1...) replacing DISTINCT ON. - [x] B3-1: Implement intra-query delta branch pruning (skip UNION ALL branches entirely when a source has zero changes).

---

Overview

When a stream table joins multiple tables (A, B, C) and multiple tables change in the same scheduler cycle, pg_trickle currently evaluates the DVM delta query using potentially separate passes or constructs unions with empty branches.

This plan outlines the steps to merge these passes into one optimized query plan, while mathematically preserving correctnesThis plan outlines the steps to merge these passes into one optimizDISTINCT ON

If an update hits both Table If an update hits both Tab saIf an update hits both Table Ifble joins them, a single canonical row might be corrected twice (once from the ΔA path, once from the ΔB path).

Previously, it was proposed to use DISTINCT ON for cross-delta deduplication. This causes silent data corruptiPreviously, it was proposed to use DISTINCT ON for cross-delta deduplication. This causes silent data corrupstPreviously, it was proposed to use DISTINCT ON for cross-delta deduplication. **This causes silent data corruptiPreviously, a_APr UNPreviouslySELECT ‘B’ AS src, * FROM delta_B ), final_delta AS ( SELECT pgt_row_id, SUM(weight) as weight, pgt_row_id HAVING SUM(weight) != 0 ) ```

Implementation Sequence

1. Property-based Correctness Proofs (B3-3) ✅ Done

What: Write extensive property tests simulating diamond flows (where changes hit multiple branches of a join simultaneously). Why: To formally prove that SUM(weight)*Why:* To formally prove thatSUM(weight)Why: To formally prove that SUM(weihatDISTINCT ONintroduced. *Where:*tests/e2e_diamond_tests.rs(Addedtest_diamond_flow_simultaneous_multi_source_update`).

2. Merged-Delta Generation (B3-2) 🔴 Not Started

What: Implement the true multi-source delta engine in src/dvm/diff.rs. Why: Generates the UNION ALL + GROUP BY logic dynamically. Where: DiffEngine::diff_node().

3. Intra-query Delta-Branch Pruning (B3-1) 🟢 Done

What: Optimize the generated UNION ALL statements. If 3 tables are in a join but only 1 changed, completely skip genWhat: Optimize the generated UNION ALL statements. IfarWhat: Optimize tPoWhat: Optimize the gand execution latency. Where:Wsrc/refresh.rs checking any_changes per buffer.