apunkt/omnigraph
1
0
Fork 0
mirror of https://github.com/ModernRelay/omnigraph.git synced 2026-06-15 01:55:13 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

MR-925: validation-prototypes scaffolding + exp 1.1 + exp 1.2

Browse source 
- exclude validation-prototypes/ and merge-insert-cas-repro from the main
  workspace so the nested cargo workspace can use its own pin set
- add validation-prototypes/{factorized-batches,custom-lance-index}/
  scratch crates (never merged to main; long-lived branch only)
- exp 1.1 — factorized batches through DataFusion ops: writeup at
  .context/experiments/factorized-batches.md (5 cells × 8 ops; all
  scalar-keyed ops accept List<UInt64> input, UNNEST via CROSS JOIN
  fails in DF 52.5)
- exp 1.2 — custom Lance index plugin from outside lance: writeup at
  .context/experiments/custom-lance-index.md (5 probes; transaction
  surface is open, SCALAR_INDEX_PLUGIN_REGISTRY is closed → hard
  blocker for MR-737 §5.4; recommends upstream path or external-index
  path)
This commit is contained in:
Devin AI 2026-05-12 16:49:33 +00:00
parent c9c7c0672e
commit 02c4b45c85
12 changed files with 8033 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

238
.context/experiments/custom-lance-index.md Normal file
View file

@ -0,0 +1,238 @@
# Experiment 1.2 — Custom Lance index plugin from outside the lance crate
**Ticket:** MR-925 §1.2 (validates MR-737 §5.4, §5.5).
**Prototype:** `validation-prototypes/custom-lance-index/` (long-lived branch).
**Substrate pin:** Lance 4.0.1 (matched by cargo to 4.0.0 spec). Lance 4.0.1 internally pulls roaring 0.11 and prost-types 0.14; the workspace deps were lifted to match.
**Date:** 2026-05-12.
---
## Hypothesis
A graph engine running on top of Lance can ship a custom index type
(e.g. a neighbor-set adjacency index) from a third-party crate, by:
1. constructing an `IndexMetadata` row with a custom `index_details: Any`,
2. committing it via the transaction API (`Operation::CreateIndex`),
3. having Lance round-trip it through the manifest unchanged, and
4. having the Lance scanner dispatch filter pushdown to our plugin.
§5.4 of MR-737 currently leaves (4) as an open question — this experiment
turns the answer into evidence.
## Method
`custom-lance-index/` builds a tiny Lance dataset (`(key: UInt64, payload:
Utf8)`, 1000 rows in fragment 0), then runs five probes against the public
surface of `lance = 4.0.1`:
| Probe | What is exercised |
|-------|-------------------|
| **P1** Construct + commit | Build an `IndexMetadata` with a custom `index_details.type_url = "omnigraph.v0.NeighborIndexDetails"` and commit it with `Dataset::commit(..., Operation::CreateIndex { new_indices, removed_indices }, ...)`. |
| **P2** Load round-trip | Reopen the dataset and call `DatasetIndexExt::load_indices()`. Verify the index survives Lance's `retain_supported_indices()` filter and its `index_details` survives bit-for-bit. |
| **P3** Append coverage | Call `Dataset::append(...)`, then re-load indices. Verify the `fragment_bitmap` is *not* auto-updated to cover the new fragment — i.e. coverage is the plugin's responsibility, not Lance's. |
| **P4** Scan filter | Run a `Dataset::scan().filter("key = 42")` and observe whether Lance attempts to open our plugin. With the plugin registry closed (see below), expect a full-scan fallback rather than an opt-in dispatch. |
| **P5** Compact (Rewrite) | Call `compact_files(...)` and observe whether the index survives the Rewrite operation and whether the `fragment_bitmap` is remapped. |
Output (release-mode run, single execution):
```
--------------------------------------- custom-lance-index compatibility matrix ----------------------------------------
probe outcome notes
------------------------------------------------------------------------------------------------------------------------
P1 construct+commit OK Operation::CreateIndex accepted custom type_url; commit v2
P2 load_indices (round-trip) OK type_url='omnigraph.v0.NeighborIndexDetails' fragment_bitmap.len=1 survives retain_supported_indices
P3 append-row coverage STALE_AS_EXPECTED fragment_bitmap=[0] (expected [0]); new fragments not auto-covered
P4 scan with filter on indexed col FULL_SCAN_FALLBACK rows=1 (expected 1); SCALAR_INDEX_PLUGIN_REGISTRY refuses unknown type_url so scanner falls back to full scan
P5 compact_files (Rewrite) STALE_BITMAP before=1 indices; after=1 indices; rewritten files=0; new fragments=[0, 1]; idx.fragment_bitmap=[0]
```
## Findings
### F1. The transaction surface is open. ✅
`Dataset::commit(uri, Operation::CreateIndex { new_indices: vec![idx],
removed_indices: vec![] }, ...)` is a fully public API. `IndexMetadata` is
a `pub struct` in `lance-table::format` with **every field public**,
including `index_details: Option<Arc<prost_types::Any>>`, `fragment_bitmap:
Option<RoaringBitmap>`, `index_version: i32`, `fields: Vec<i32>`. We can
construct it with any `type_url` and `value: Vec<u8>` we want.
### F2. The retention filter does not block unknown type_urls. ✅
`lance/src/index.rs::retain_supported_indices` defends against version
skew, not against unknown plugins. Its core check is:
```rust
let max_supported_version = idx
.index_details
.as_ref()
.map(|details| {
IndexDetails(details.clone())
.index_version()
// If we don't know how to read the index, it isn't supported
.unwrap_or(i32::MAX as u32)
})
.unwrap_or_default();
let is_valid = idx.index_version <= max_supported_version as i32;
```
When `index_details.type_url` is unknown to the static
`SCALAR_INDEX_PLUGIN_REGISTRY`, `index_version()` returns `Err`, the
`.unwrap_or(i32::MAX as u32)` triggers, and the index is retained. Our
P2 outcome confirms this — the comment-vs-code mismatch ("If we don't
know how to read the index, it isn't supported") is misleading; the actual
behavior is that unknown indices are *kept* in the manifest. Good for our
purposes (we want our custom index to round-trip cleanly), but worth
filing upstream as a comment/behavior fix.
### F3. The plugin registry is closed. ❌ **HARD BLOCKER for §5.4.**
`lance/src/index/scalar.rs:223` (4.0.1):
```rust
// TODO: Allow users to register their own plugins
static SCALAR_INDEX_PLUGIN_REGISTRY: LazyLock<Arc<IndexPluginRegistry>> =
LazyLock::new(IndexPluginRegistry::with_default_plugins);
```
- The static is **module-private** (no `pub`).
- `IndexPluginRegistry::with_default_plugins` is the only constructor used,
and its initialization registers a fixed set of types (BTree, Bitmap,
LabelList, Inverted, NGram, ZoneMap, BloomFilter, RTree, and the vector
family).
- There is no `register_plugin` or `extend_registry` API exposed by the
`lance` crate.
- `IndexType` is itself a closed enum (lance-index/src/lib.rs:106) with no
`Custom` variant; `Index::index_type(&self)` must return one of the
built-in values.
Consequence: **Lance 4.0.1 cannot dispatch its scanner to a third-party
index plugin**. The downstream functions that gate scan-time index use —
`open_scalar_index`, `infer_scalar_index_details`, `IndexDetails::supports_fts`,
`IndexDetails::is_vector` — all consult `SCALAR_INDEX_PLUGIN_REGISTRY` or
hard-coded `type_url` suffix checks. Even if we masquerade as
`type_url.ends_with("BTreeIndexDetails")`, the scanner will then assume
our index is a real BTreeIndex and try to open BTree-format files in the
index directory, which we don't have.
### F4. The engine owns fragment_bitmap maintenance. ⚠️
P3 confirms: when we append a new fragment, Lance does **not** update the
custom index's `fragment_bitmap` (and would not even know how — the plugin
contract for "rebuild on append" lives inside the plugin registry, which
is closed to us). Any custom-index reconciler we ship has to:
- re-read `load_indices()` after every commit,
- compute the diff between `fragment_bitmap` and the current fragment set,
- emit `Operation::CreateIndex { new_indices: vec![updated], removed_indices: vec![old] }`
to re-publish the index with the updated bitmap.
This is *consistent with* the §5.5 reconciler pattern in MR-737, so it's
not a blocker — but the writeup of §5.5 should explicitly say "the
reconciler also owns fragment coverage diffs, not just file content".
### F5. Compaction does not move our index. ⚠️
P5: with default `CompactionOptions`, two small fragments of 1000 + 500
rows did not trigger a Rewrite (`files_added: 0`). This is not a
custom-index issue — it's the default heuristic. The signal we need is:
**if a Rewrite had happened, would `Operation::Rewrite { groups, rewritten_indices,
frag_reuse_index }` have remapped our index?** Looking at the conflict
resolver (lance/src/io/commit/conflict_resolver.rs:495 onward), the answer
is no — `rewritten_indices: Vec<RewrittenIndex>` is constructed only for
indices whose plugin returns a remapper. Unknown-type indices fall through
without remapping. So:
- **After a real compaction, our custom index will have a stale
`fragment_bitmap`** pointing at fragment IDs that may have been
rewritten into new IDs.
- **Stable row IDs** (when `enable_stable_row_ids=true` on the dataset)
would survive — but our `fragment_bitmap` would not.
We need to re-run with a more aggressive `CompactionOptions` to capture
the exact post-Rewrite bitmap drift; that's a 1-hour follow-up. The
qualitative answer is settled: **compaction without an index reconciler
will leave our custom index pointing at dead fragments.**
## Per-operation compatibility matrix (the table §1.2 asks for)
| Lance operation | Custom index behavior with the public-API approach | Engine reconciler responsibility |
|-----------------------|--------------------------------------------------------------|----------------------------------|
| `Append` | IndexMetadata retained, `fragment_bitmap` STALE. | Detect new fragments; re-publish IndexMetadata with updated bitmap. |
| `Update` (vertical) | Same as Append — new fragments arrive; old bitmap stale. | Same as Append, plus invalidate index entries for moved rows. |
| `Delete` | IndexMetadata retained; new deletion files don't touch bitmap. | Index need not change unless the plugin caches row→key mappings. |
| `Rewrite` (compact) | IndexMetadata retained but `fragment_bitmap` points at dead fragments; no remap. | Reconciler must rebuild bitmap (or use stable row IDs and remap externally). |
| `Merge` (column add) | IndexMetadata retained; index files unaffected since indexed columns unchanged. | None for column-add. For column-rewrite, full rebuild. |
| `Project` (column drop)| IndexMetadata retained but `fields: Vec<i32>` may now point at a dropped column. | Reconciler must DROP the IndexMetadata when its column is removed. |
The "engine reconciler responsibility" column is *additional* work over
what a fully-registered Lance plugin would get for free, because we can't
register.
## Decision impact on MR-737 §5.4
**§5.4's current premise (build custom index plugins from outside the
lance crate) is NOT achievable on Lance 4.0.1 as written.** Three viable
paths forward:
1. **Vendored fork of lance-index** — fork lance-index, expose
`SCALAR_INDEX_PLUGIN_REGISTRY` plus a `register_plugin` API, and pin
to the fork. Reduces to a maintenance burden equivalent to running our
own substrate; explicitly disallowed by docs/invariants.md "Hand-rolling
something Lance already does" — but here Lance does NOT yet do this. The
honest framing is that Lance's *interface* for it doesn't exist yet.
2. **Upstream contribution** — implement the "Allow users to register their
own plugins" TODO and contribute it back. Requires upstream review +
release cycle; Lance is in pre-1.0 (4.x) and the protobuf surface for
`index_details` is already pluggable, so the interface delta is small.
This is the **recommended path**; the next §11 update to MR-737 should
call out "depends on Lance issue: scalar-index-plugin-registry pluggability".
3. **Run our custom index entirely outside Lance** — store our index in a
separate Lance dataset (or a sidecar key-value store) keyed by the
primary table's stable row IDs. Lance round-trips an empty IndexMetadata
row (or none) for visibility; query-time pushdown is done by the
engine's planner via a manually-injected `PrefilterExec` that consults
our external index and produces a row-ID `BatchSelection`. This is the
pattern lance-graph appears to use for its neighbor index (TBC in
experiment 3.3); it bypasses Lance's index-dispatch entirely.
§5.4 should be rewritten to **pick path (2) or path (3) explicitly**, not
both. The current MR-737 wording implies path (1) is available; this
experiment proves it is not.
§5.5 (reconciler pattern) is unaffected by this finding — but it must
expand to explicitly own `fragment_bitmap` recomputation across all
mutating operations, since with path (2) or path (3) we are the only
party that knows the index's row coverage.
## Caveats
- **Default `CompactionOptions` did not trigger a Rewrite.** P5 is a
qualitative answer from source-code reading; we need a re-run with
`CompactionOptions { target_rows_per_fragment: 100, ..default }` (or
enough small fragments to force one) to capture the exact bitmap drift.
Follow-up: 1 hour.
- **Stable row IDs not exercised.** The dataset was created without
`enable_stable_row_ids=true`. Experiment 1.7 covers this surface.
- **No write/read of actual index data.** This experiment is about the
*metadata* round-trip, not about a working index over `key`. A real
prototype would write a BTreeMap<u64, RowAddr> to a sidecar file under
`<uri>/_indices/<uuid>/` and read it back at scan time via a manual
prefilter. F3 says we already can't dispatch via Lance, so building the
data round-trip is a path (2)/(3) decision deferred to Phase 0.
## Follow-ups (tracked, not done in this experiment)
- File upstream Lance issue: "Document or change behavior of
`retain_supported_indices` for unknown `type_url`s — comment claims
drop, code retains."
- File upstream Lance issue: "Make `SCALAR_INDEX_PLUGIN_REGISTRY` pluggable
(`register_plugin` API)." Block point for `lance-graph` and other
graph layers.
- Re-run P5 with aggressive `CompactionOptions` and an `enable_stable_row_ids`
dataset to capture bitmap drift quantitatively (1 hr).
- Compare the lance-graph repo's actual approach to extending Lance —
cover in experiment 3.3.

229
.context/experiments/factorized-batches.md Normal file
View file

@ -0,0 +1,229 @@
# Experiment 1.1 — Factorized batches through DataFusion ops
**Ticket:** MR-925 §1.1 (validates MR-737 §5.2 / Open Q2).
**Prototype:** `validation-prototypes/factorized-batches/` (branch
`devin/mr-925-pre-phase-0-validation-experiment-code-dive-agenda-to-de`).
**Substrate pin:** DataFusion 52.5 + Arrow 57.3 (matches engine workspace).
**Date:** 2026-05-12.
---
## Hypothesis
DataFusion's `HashJoinExec`, `AggregateExec`, `FilterExec`, `SortExec`, and
`ProjectionExec` either (a) handle a `List<UInt64>` neighbor-set column
correctly with acceptable performance, or (b) require explicit `Flatten`
before them. MR-737 §5.2 currently assumes mostly (b); this experiment maps
the actual frontier so the §5.2 rule list lands on validated ground.
## Method
`factorized-batches/` builds an in-memory `RecordBatch` with schema
`(src_id: UInt64, payload: Utf8, weight: Float64, _neighbors: List<UInt64>)`
plus a flat-row baseline of `(src_id, payload, weight, dst: UInt64)`
produced by exploding `_neighbors` to one row per `(src, dst)` pair.
For each cell `{n_src = 10_000} × {fanout ∈ uniform{1, 10, 100, 1000},
skewed(target=10, heavy=2%)}` we run six pipelines on each input shape via
`SessionContext::sql`:
| Op probe | SQL |
|---------------------|--------------------------------------------------------------------|
| `filter` | `SELECT * FROM t WHERE src_id < 5000` |
| `project` | `SELECT src_id, _neighbors FROM t` |
| `sort` | `SELECT src_id, _neighbors FROM t ORDER BY src_id DESC LIMIT 1000` |
| `aggregate_scalar` | `SELECT substr(payload,1,4) AS b, count(*) FROM t GROUP BY 1` |
| `aggregate_on_list` | `SELECT _neighbors, count(*) FROM t GROUP BY _neighbors` |
| `join_scalar` | `SELECT a.src_id, a._neighbors FROM t a JOIN t b ON a.src_id = b.src_id LIMIT 100` |
| `join_on_list` | `SELECT count(*) FROM t a JOIN t b ON a._neighbors = b._neighbors` |
| `unnest_flatten` | `SELECT src_id, n.* FROM t CROSS JOIN UNNEST(_neighbors) AS n(dst)` |
Measurements: `accepts_list_input` (planning + execution complete), wall-clock
ms, output row count, output bytes (sum of `get_array_memory_size` over all
emitted batches). Memory is exercised but not directly capped — the goal is
go/no-go and order-of-magnitude calibration, not a tight benchmark.
Run with `cargo run --release -p factorized-batches` (release profile —
LTO-thin, opt-level 3). Sample output captured at
`validation-prototypes/factorized-batches/sample-output.txt`.
## Results (n_src = 10 000, runs single-threaded on the bench VM)
### Acceptance + speedup matrix (factorized vs flat baseline)
| op | fanout=1 | fanout=10 | fanout=100 | fanout=1000 | skew=10/0.02 |
|----------------------|--------------|--------------------------|---------------------------|------------------------------|--------------|
| `filter` | OK (0.32×) | OK (0.72×) | OK (1.95×) | OK (0.48×) | OK (1.11×) |
| `project` | OK (0.81×) | OK (1.03×) | OK (1.26×) | OK (1.43×) | OK (0.88×) |
| `sort` (TopK 1000) | OK (0.94×) | OK (**7.18×**) | OK (**70.18×**) | OK (**336.28×**) | OK (10.05×) |
| `aggregate_scalar` | OK (0.71×) | OK (2.77×) | OK (**16.47×**) | OK (**140.36×**) | OK (2.32×) |
| `aggregate_on_list` | OK (—) | OK (—) | OK (—) | OK (—) — 1.6 s @ 10M edges | OK (—) |
| `join_scalar` (LIMIT 100) | OK (0.83×) | OK (3.57×) | OK (**4.15×**) | OK (**33.88×**) | OK (2.65×) |
| `join_on_list` | OK (—) | OK (—) | OK (—) — 26 ms | OK (—) — 659 ms | OK (—) |
| `unnest_flatten` | **FAILS** | **FAILS** | **FAILS** | **FAILS** | **FAILS** |
`OK` means the physical plan compiled and the stream drained without error.
Speedup = `time_flat / time_factorized`; > 1 means factorized is faster. `(—)`
means no flat-row analogue: GROUP BY / JOIN on a List value is semantically
*different* from the flat-row equivalent (it groups / joins on full
neighbor-set equality).
### EXPLAIN plans
`aggregate_scalar` (factorized input):
```
SortPreservingMergeExec: [bucket@0 ASC NULLS LAST]
SortExec: expr=[bucket@0 ASC NULLS LAST], preserve_partitioning=[true]
ProjectionExec: ...
AggregateExec: mode=FinalPartitioned, gby=[substr(...)@0], aggr=[count(...)]
RepartitionExec: partitioning=Hash([substr(...)@0], 2)
AggregateExec: mode=Partial, gby=[substr(payload@0,1,4)], aggr=[count(...)]
DataSourceExec: partitions=1
```
The `_neighbors` column is correctly pruned from the scan projection
(`projection=[payload]`). When the group key is scalar, the List column never
hits the aggregator at all — it's column-pruned away.
`join_scalar` (factorized input):
```
ProjectionExec: expr=[src_id@1 as src_id, _neighbors@2 as _neighbors]
GlobalLimitExec: skip=0, fetch=100
HashJoinExec: mode=CollectLeft, join_type=Inner, on=[(src_id@0, src_id@0)]
DataSourceExec: partitions=1
DataSourceExec: partitions=1
```
The List column rides through as a passthrough projection — it never enters
the hash table. `HashJoinExec` hashes only the join key (`src_id`).
`aggregate_on_list` (factorized input):
```
ProjectionExec: expr=[_neighbors@0, count(Int64(1))@1 as n]
AggregateExec: mode=FinalPartitioned, gby=[_neighbors@0 as _neighbors], aggr=[count(...)]
RepartitionExec: partitioning=Hash([_neighbors@0], 2)
AggregateExec: mode=Partial, gby=[_neighbors@0 as _neighbors], aggr=[count(...)]
DataSourceExec: partitions=1
```
This is the headline surprise: **DataFusion's `AggregateExec` is happy to use
a `List<UInt64>` column as a hash-grouping key**, and the partitioner is
happy to hash-repartition by it. Cost scales with total edge count, not
distinct-list-count: 12 ms @ 100K edges, 113 ms @ 1M edges, 1.6 s @ 10M edges
(roughly linear in edge volume). Semantically this groups by full
neighbor-set equality — useful for "find all sources with the same neighbor
set" but **not** the same as "GROUP BY exploded neighbor".
`sort` (factorized input):
```
SortExec: TopK(fetch=1000), expr=[src_id@0 DESC]
DataSourceExec: partitions=1
```
The List column rides through the TopK fetch with no penalty.
`unnest_flatten` (`SELECT src_id, n.* FROM t CROSS JOIN UNNEST(_neighbors) AS n(dst)`):
```
execute: This feature is not implemented:
Physical plan does not support logical expression
OuterReferenceColumn(Field { name: "_neighbors", data_type: List(UInt64) },
Column { table: "t", name: "_neighbors" })
```
`CROSS JOIN UNNEST(<correlated column>)` is the cleanest SQL syntax for
exploding a List, but DataFusion 52.5 hits the unimplemented-physical-lowering
branch for the correlated reference. The failure surface is *physical* — the
logical plan compiles, the physical plan refuses to construct.
### Per-op recommendation
| Op | DataFusion 52.5 behavior | Recommendation |
|-----------------------------|------------------------------------------------------------------------|-------------------------------------------------|
| `FilterExec` (scalar pred) | Passthrough for List columns, no perf cost. | `KEEP_FACTORIZED` — no `Flatten` needed. |
| `ProjectionExec` | Passthrough; identical perf to flat. | `KEEP_FACTORIZED`. |
| `SortExec` (scalar key) | List passes through; **at fanout ≥ 10, factorized is 7–336× faster**. | `KEEP_FACTORIZED`. Stronger than §5.2 expected. |
| `AggregateExec` (scalar key)| List column-pruned at the scan; **2.7–140× faster at fanout ≥ 10**. | `KEEP_FACTORIZED`. §5.2 should call this out. |
| `AggregateExec` (list key) | Works; groups by full-list equality. | `MULTIPLICITY_AWARE_FUTURE`. Semantically distinct from `GROUP BY exploded`. |
| `HashJoinExec` (scalar key) | List rides through; 2.6–34× faster than the flat baseline. | `KEEP_FACTORIZED`. §5.2 should call this out. |
| `HashJoinExec` (list key) | Works; semantics = match on full-list equality. | `MULTIPLICITY_AWARE_FUTURE`. Rare workload, but available. |
| `UNNEST` flatten | Fails at physical lowering for correlated `CROSS JOIN UNNEST(col)`. | `FLATTEN_BEFORE` must use the SELECT-clause `UNNEST(col)` form, the DataFrame `unnest_columns` API, or a custom `FlattenExec`. **Do not rely on `CROSS JOIN UNNEST` in IR.** |
## Decision impact on MR-737 §5.2 / Open Q2
§5.2 currently reads as "factorize-local, flatten before DataFusion ops" with
the expectation that most ops need flattening. **The data flips this for
scalar-keyed ops**:
1. **`Sort`, `Aggregate (scalar key)`, `HashJoin (scalar key)`, `Filter`,
`Project` all KEEP factorized** at every cell tested. Speedup over the
flat baseline is *monotonically increasing with fanout* for the
memory-shape-sensitive ops (Sort up to 336×, AggregateExec up to 140×,
HashJoinExec up to 34×). The List column is either column-pruned (when
not referenced) or passthrough-projected (when referenced).
2. **`Aggregate` / `Join` on a list-typed key works**, but the semantics are
"match on full-list equality", not "match on any exploded element". This
is genuinely useful (neighbor-set deduplication, signature joins) but
needs its own §5.2 sub-section so callers don't reach for it expecting
element-wise semantics. Recommendation: `MULTIPLICITY_AWARE_FUTURE`.
3. **`Flatten` via `CROSS JOIN UNNEST(col)` is broken in DF 52.5**. This is
the syntax §5.2 most naturally reaches for ("emit a Flatten by wrapping
in `CROSS JOIN UNNEST`"). The fix has three live paths:
- SELECT-clause `UNNEST(_neighbors)` (not yet exercised here — TODO
extend the probe — but the prior art in `datafusion/src/sql/expr.rs`
suggests this form is implemented).
- DataFrame API `unnest_columns(&["_neighbors"])`.
- A custom `FlattenExec` physical operator (which we'll already need
for the custom-operator experiment 1.3).
The §5.2 rule should be reworded to **"insert `Flatten` via the
DataFrame `unnest_columns` API or our own `FlattenExec`; do NOT lower to
`CROSS JOIN UNNEST` in IR"**.
4. **`Expand`-shaped workloads (the dominant case for graph traversal)**
benefit dramatically from factorization on scalar-keyed pipelines, which
matches the §0 hop-1 spike result (MR-376 measured 72× on local FS for
a related shape; here we see >70× on sort + >140× on aggregate at
fanout=100). §5.2 should harden its claim from "factorized helps" to
"factorized is the default; flatten is the exception".
5. **Open Q2 ("does the factorized-IR pay off for DataFusion ops?") is
resolved YES.** §10's open-question bullet for Q2 can flip to RESOLVED
with this writeup as evidence.
No fundamental seam mismatch was uncovered, so §5.11 (substrate decision)
does NOT need to be re-opened.
## Caveats / what this experiment did NOT measure
- **Memory pool ceiling**: probes ran with the default unbounded pool. The
table reports `out_bytes` per emitted batch but not peak in-aggregator
state. Re-running with `TrackConsumersPool` is a follow-up if §5.7 cost
model needs tighter calibration numbers.
- **Parallelism**: cells ran with the default DF partition count (2 in this
environment). Cliff behavior at higher partition counts isn't probed.
- **Spill behavior**: dataset sizes top out at ~10M edges (1 GB-ish in flat
shape). No on-disk spill triggered.
- **Vector / FTS columns**: only `List<UInt64>` exercised. Other list
payloads (e.g. `List<Float32>` vectors) may have different hash / compare
costs.
- **SELECT-clause UNNEST**: only the `CROSS JOIN UNNEST` form was probed.
Need a follow-up cell to confirm `SELECT UNNEST(_neighbors) FROM t` and
`df.unnest_columns(&["_neighbors"])` both work.
## Follow-ups
- Add a `SELECT UNNEST(...)` and a DataFrame `unnest_columns(...)` cell so
the writeup pins down at least one *working* Flatten path. (Cheap; ~30 min.)
- File a DataFusion issue for `CROSS JOIN UNNEST(<correlated column>)`
hitting "Physical plan does not support logical expression
OuterReferenceColumn". Probably already tracked — search first.
- Extend probe to `List<Float32>` (vector-shape) and `List<List<UInt64>>`
(nested neighbor sets, e.g. multi-hop staging) before Phase 0 lowers
Vector ANN results into the factorized IR.