fix: Table-wide edge batching in the OpenGraph convert source - BED-9372

benchmarks/DESTINATION_MEMORY_REVIEW.md

@@ -0,0 +1,114 @@
+# BED-9372 destination memory & part-size review
+
+Acceptance criteria: peak memory and part sizes stay bounded independently of
+table cardinality, against the destination's 1,000-item batch, without
+unsupported upload artifacts.
+
+## Revision notes
+
+- An earlier revision called peak RSS "orthogonal to the batching fix." Wrong:
+  growth was linear (~1.2 KB per relationship) because each wrapper now carries
+  up to `batch_size` relationships. Root cause and fix below.
+- Causal chain: the batching fix is the feature; the memory fix was required to
+  keep it viable at scale (RSS scaled linearly with cardinality); and the memory
+  fix exposed a latent DLT 1.26.0 data-corruption defect, which required an
+  in-process correction — without it, delivery silently duplicates items.
+- The first mitigation attempt (buffer=333, DLT untouched) demonstrated that
+  defect: the faker BHE scheduler test ingested 2,875 nodes where 1,000 were
+  expected. Corrected in-process; all numbers below re-measured after.
+
+## Root cause
+
+DLT's writer buffer flushes on **item count** (`data_writer.buffer_max_items`,
+default 5,000), and each edge wrapper counts as one item regardless of its
+content. Post-batching that is up to
+
+    5,000 items x 150 edges = 750,000 relationships in RAM per flush.
+
+Measured: peak RSS 255 MB at 100k rows -> 1,257 MB at 1M rows (~1.2 KB per
+relationship); the `batch_size=1` baseline stayed flat at ~130 MB.
+
+## DLT 1.26.0 jsonl batching defect (corrected in-process)
+
+`DestinationJsonlLoadJob.get_batches` (`dlt/destinations/job_impl.py:240`)
+yields the accumulated batch at the end of every load-file line without
+resetting it, so multi-line files re-deliver earlier items as a growing prefix:
+
+    buffer=333, 1,200 rows -> 1,208 wrappers delivered as
+    333 + 666 + 999 + 1000 + 208 = 3,206 items
+
+Stock settings are only accidentally safe: 5,000 divides evenly by the
+destinations' `batch_size=1000`, so partial batches stay empty until each
+file's last line. Any other buffer value triggers duplicates. Upstream
+refactored this code in release 1.27.0 (`JsonlFileBatchIterator`, verified)
+with correct semantics; `openhound.core.dlt_jsonl_batching` installs those
+semantics process-wide for dlt 1.26.x, making any buffer value safe.
+
+## Shipped mitigation
+
+`writer_buffer_max_items()` scales the buffer so buffered *relationships* stay
+near a fixed budget; applied via `DATA_WRITER__BUFFER_MAX_ITEMS` (`setdefault`,
+user overrides win) in `Converter.pipeline` and the benchmark harness:
+
+    buffer_max_items = min(5000, max(1, 50_000 // batch_size))   # 333 @ 150
+
+More frequent flushes write to the same open file; wall time is unchanged
+within noise. The jsonl batching module above is what makes 333 safe.
+
+## Method
+
+`benchmarks/opengraph_batching_benchmark.py`, one-edge shape, 4 files, DLT
+1.26.0 with the batching module active, single load worker. Table-wide =
+`batch_size=150`; baseline = pre-fix `batch_size=1`. Untuned rows set
+`DATA_WRITER__BUFFER_MAX_ITEMS=5000` explicitly (delivery identical with and
+without the correction, since 5,000 aligns with the destination batch size).
+
+## Results
+
+Peak RSS (sampled at 50 ms):
+
+| Scale | Mode | buffer | Wrappers | Peak RSS | Wall |
+|------:|------|-------:|---------:|---------:|-----:|
+| 100k | baseline | 5,000 | 100,000 | 130 MB | 18.2s |
+| 100k | table-wide untuned | 5,000 | 667 | 255 MB | 6.5s |
+| 100k | table-wide tuned | 333 | 667 | 240 MB | 6.6s |
+| 1M | baseline | 5,000 | 1,000,000 | 132 MB | 111.8s |
+| 1M | table-wide untuned | 5,000 | 6,667 | 1,257 MB | 36.2s |
+| 1M | table-wide tuned | 333 | 6,667 | 499 MB | 38.9s |
+| 4M | table-wide tuned | 333 | 26,667 | 498 MB | 161.5s |
+
+Tuned RSS is **flat across cardinality** (499 MB @ 1M vs 498 MB @ 4M); the
+untuned trend extrapolates to multiple GB at customer scale (12.5M+ rows). The
+benchmark reports `peak_rss_per_edge` and enforces a guard band (300 MiB floor
++ 512 B/edge) — fires on untuned runs, silent on tuned ones.
+
+Destination callbacks / parts (one JSON part per callback):
+
+| Scale | Mode | Callbacks / Parts | Max rel/callback | Max bytes/callback |
+|------:|------|------------------:|-----------------:|-------------------:|
+| 100k | baseline | 100 / 100 | 1,000 | 178 KB |
+| 1M | baseline | 1,000 / 1,000 | 1,000 | 180 KB |
+| 1M | table-wide untuned | 7 / 7 | 150,000 | 27.0 MB |
+| 1M | table-wide tuned | 7 / 7 | 150,000 | 27.0 MB |
+| 4M | table-wide tuned | 27 / 27 | 150,000 | 27.3 MB |
+
+(An earlier revision listed 27 parts / ~9 MB at 1M and 107 parts at 4M —
+artifacts of the defect's fragmented callbacks.) All runs deliver exact item
+totals, `inner_relationships` = row count, and 0 warnings when tuned.
+
+## Bounds
+
+- Destination callback/part: `1,000 x 150 = 150,000` relationships (~27 MB),
+  independent of cardinality and of the extract buffer.
+- Process peak RSS: flat 1M -> 4M post-mitigation; guarded by the benchmark band.
+
+## Conclusion
+
+Both acceptance bounds hold, and no unsupported upload artifacts are created.
+Verified against release tags: only the 1.26 series carries the defect (fixed
+in 1.27.0; `buffer_max_items` semantics unchanged through 1.30.0, so the
+coordination applies on all versions). If the pin moves past 1.26,
+`ensure_dlt_jsonl_batching` becomes a no-op and upstream's corrected iterator
+takes over — worth reporting upstream with the minimal reproduction above.
+Future wrapper-width or buffering changes will surface via `peak_rss_per_edge`
+and the warnings list.

benchmarks/_bench_assets.py

@@ -0,0 +1,148 @@
+"""Generic synthetic assets and input generation for the BED-9372 benchmark.
+
+These assets are deliberately extension-agnostic (not Okta/SAML-specific) so the
+benchmark exercises the shared opengraph source, per the ticket requirement to
+use generic synthetic assets plus at least one non-Okta high-cardinality shape.
+"""
+
+from __future__ import annotations
+
+import gzip
+import json
+import math
+from dataclasses import dataclass, field
+from pathlib import Path
+
+from openhound.core.asset import BaseAsset
+from openhound.core.models.entries_dataclass import (
+    Edge,
+    EdgePath,
+    Node as DNode,
+    NodeProperties as DNodeProperties,
+)
+
+
+def _edge(idx: int, k: int = 0) -> Edge:
+    return Edge(
+        kind="BENCH_Relationship",
+        start=EdgePath(match_by="id", value=f"start-{idx}-{k}"),
+        end=EdgePath(match_by="id", value=f"end-{idx}-{k}"),
+    )
+
+
+class OneEdgeAsset(BaseAsset):
+    """High-cardinality zero-or-one-edge shape (e.g. membership/grant rows)."""
+
+    idx: int
+
+    @property
+    def as_node(self):
+        return None
+
+    @property
+    def edges(self):
+        return [_edge(self.idx)]
+
+
+class MultiEdgeAsset(BaseAsset):
+    """A row emitting several edges, to stress inner-relationship growth."""
+
+    idx: int
+    n: int
+
+    @property
+    def as_node(self):
+        return None
+
+    @property
+    def edges(self):
+        return [_edge(self.idx, k) for k in range(self.n)]
+
+
+@dataclass
+class _BenchNode(DNode):
+    id: str = field(default="")
+
+    def __post_init__(self):
+        self.id = f"node-{self.properties.name}"
+
+
+class NodeAndEdgeAsset(BaseAsset):
+    """Node-bearing row that also emits one containment/ownership edge."""
+
+    idx: int
+
+    @property
+    def as_node(self):
+        return _BenchNode(
+            kinds=["BENCH_Node"],
+            properties=DNodeProperties(
+                name=f"n{self.idx}",
+                displayname=f"Node {self.idx}",
+                environmentid="bench-env",
+            ),
+        )
+
+    @property
+    def edges(self):
+        return [
+            Edge(
+                kind="BENCH_Relationship",
+                start=EdgePath(match_by="id", value=f"node-n{self.idx}"),
+                end=EdgePath(match_by="id", value=f"end-{self.idx}-0"),
+            )
+        ]
+
+
+def _single_edge_row(idx: int, epr: int) -> dict:
+    """Row builder for shapes that emit exactly one edge per row.
+
+    These shapes cannot vary the edge count, so reject any edges_per_row other
+    than 1 rather than silently discarding it.
+    """
+    if epr != 1:
+        raise ValueError(
+            f"edges_per_row={epr} is unsupported for single-edge shapes; use 1"
+        )
+    return {"idx": idx}
+
+
+# Maps a shape name to (asset model, row builder). The row builder returns the
+# raw dict that read_jsonl will feed back into the model.
+ASSET_SHAPES: dict[str, tuple[type[BaseAsset], object]] = {
+    "one_edge": (OneEdgeAsset, _single_edge_row),
+    "multi_edge": (MultiEdgeAsset, lambda idx, epr: {"idx": idx, "n": epr}),
+    "node_and_edge": (NodeAndEdgeAsset, _single_edge_row),
+}
+
+
+def model_for_shape(shape: str) -> type[BaseAsset]:
+    return ASSET_SHAPES[shape][0]
+
+
+def _write_gz(path: Path, rows: list[dict]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with gzip.open(path, "wt", encoding="utf-8") as fh:
+        for row in rows:
+            fh.write(json.dumps(row) + "\n")
+
+
+def write_synthetic_input(
+    input_dir: Path, shape: str, rows: int, edges_per_row: int, files: int
+) -> str:
+    """Write `rows` synthetic rows for `shape` across `files` .jsonl.gz files.
+
+    Returns the table (subdirectory) name used by the opengraph file_glob.
+    """
+    model, row_builder = ASSET_SHAPES[shape]
+    table = model.__name__.lower()
+    per_file = math.ceil(rows / files)
+    written = 0
+    for f in range(files):
+        count = min(per_file, rows - written)
+        if count <= 0:
+            break
+        batch = [row_builder(written + i, edges_per_row) for i in range(count)]
+        written += count
+        _write_gz(input_dir / table / f"part-{f:04d}.jsonl.gz", batch)
+    return table