# v0.84.0 — Vectorized Compute & Adaptive Engine

> **Status:** Planned  
> **Scope:** Large  
> **Driven by:** [Assessment 16](../plans/PLAN_OVERALL_ASSESSMENT_16.md) — MT-8, LT-7, LT-9, PERF-O4

## Theme

Introduce vectorized batch processing for high-throughput stream tables,
proper incremental algorithms for window functions (replacing partition-based
recomputation), and cost-based operator scheduling within delta queries. This
release targets a 5–10× throughput improvement for aggregate-heavy and
window-function-heavy workloads.

## Items

### MT-8: Vectorized Aggregate Path

For pure-aggregate STs (GROUP BY with no joins), process deltas using
Arrow-compatible columnar batches:

1. Read change buffer rows into columnar batches (1024-row pages)
2. Vectorized hash aggregation: group-by keys → aggregate accumulators
3. SIMD-accelerated SUM/COUNT/MIN/MAX operations on batch columns
4. Emit only changed groups as delta output
5. Apply via batched MERGE (INSERT/UPDATE per group)

Benefits:
- 10× throughput for SUM/COUNT/AVG aggregates (SIMD auto-vectorization)
- Cache-friendly memory access patterns (column-at-a-time processing)
- Reduced SPI overhead (bulk operations instead of per-row)

Implementation:
- New `VectorizedAggregateOperator` in `src/dvm/operators/vectorized_agg.rs`
- Uses `arrow-array` and `arrow-compute` crates for batch processing
- Activated when OpTree is a single Aggregate node over a Scan (no joins)
- Falls back to standard SQL-based path for complex trees

### LT-7: Incremental Window Function Computation

Replace the current partition-based recomputation strategy for window
functions with proper incremental algorithms:

| Function | Current Strategy | New Strategy |
|----------|-----------------|--------------|
| `ROW_NUMBER()` | Recompute partition | Maintain sorted B-tree index; insert/delete at position |
| `RANK()` / `DENSE_RANK()` | Recompute partition | Track rank counts per distinct value; adjust on delta |
| `LAG(col, N)` / `LEAD(col, N)` | Recompute partition | Ring buffer of N preceding/following values per partition |
| `SUM() OVER (...)` | Recompute partition | Running sum with prefix-sum tree for range frames |
| `COUNT() OVER (...)` | Recompute partition | Running count (algebraic) |
| `FIRST_VALUE()` / `LAST_VALUE()` | Recompute partition | Track min/max by sort key |
| `NTH_VALUE()` | Recompute partition | Indexed array per partition |

Fallback: Functions with `ROWS BETWEEN` or `RANGE BETWEEN` frames that are
not covered by the incremental algorithms fall back to partition recomputation.

New metadata in `pgt_stream_tables`: `window_strategy JSONB` storing per-window
function the chosen incremental strategy and auxiliary state location.

### LT-9: Cost-Based Operator Scheduling

Reorder operators within a delta query's OpTree based on estimated cardinality
to minimize intermediate result sizes:

1. **Selectivity estimation:** Use PostgreSQL's `pg_statistic` to estimate
   filter selectivity and join cardinality
2. **Operator reordering rules:**
   - Push high-selectivity filters before joins (existing P2-7, extended)
   - Order join inputs by ascending estimated cardinality (smaller relation
     as build side)
   - Place DISTINCT before expensive projections
3. **Cost model:** Estimate total intermediate tuples for each valid ordering;
   choose minimum-cost plan
4. **Safety:** Only reorder when semantically equivalent (respects NULL
   handling, outer join placement constraints)

New diagnostic: `EXPLAIN (FORMAT JSON) SELECT pgtrickle.explain_delta_plan(pgt_id)`
shows the estimated cost and chosen operator order.

### PERF-O4: Parallel Delta Computation Fan-Out

For joins with multiple independent source tables, compute each source's
delta contribution in parallel:

```
              Delta Query
              /    |    \
    Source A  Source B  Source C
    delta     delta     delta
       \       |       /
        \      |      /
         Join Merge
              |
         MERGE Apply
```

Implementation:
- Identify independent scan branches in the OpTree (no cross-dependencies)
- Spawn parallel SPI connections (one per branch, up to `parallel_fan_out_max`)
- Each branch computes its partial delta independently
- Results are joined/merged in the coordinator thread
- Activated when: (a) >1 source table changed, (b) ST is not in a diamond
  consistency group, (c) `pg_trickle.parallel_delta_fan_out = true`

New GUC: `pg_trickle.parallel_delta_fan_out = false` (opt-in, requires testing)