839224346 commented on code in PR #9315:
URL: https://github.com/apache/paimon/pull/9315#discussion_r3840420360
##########
paimon-python/pypaimon/table/source/vector_search_read.py:
##########
@@ -824,3 +924,222 @@ def _compute_score(query, stored, metric):
if metric == "inner_product":
return sum(float(q) * float(s) for q, s in zip(query, stored))
raise ValueError("Unknown vector search metric: %s" % metric)
+
+
+def _raw_search_vectorized(row_ids, vectors, query_vector, metric, limit,
+ score_candidates=None):
+ """Vectorized raw search using numpy for batch distance computation."""
+ import numpy as np
+
+ # Filter by score_candidates and null vectors.
+ if score_candidates is not None:
+ candidate_set = set(score_candidates)
+ filtered = [(rid, vec) for rid, vec in zip(row_ids, vectors)
+ if rid in candidate_set and vec is not None]
+ else:
+ filtered = [(rid, vec) for rid, vec in zip(row_ids, vectors)
+ if vec is not None]
+
+ if not filtered:
+ return DictBasedScoredIndexResult({})
+
+ filtered_ids, filtered_vecs = zip(*filtered)
+ row_id_array = np.array(filtered_ids, dtype=np.int64)
+ stored_matrix = np.array(
+ [_to_vector_list(v) for v in filtered_vecs], dtype=np.float32)
+ query_np = np.array(
+ _to_vector_list(query_vector) if not isinstance(query_vector,
np.ndarray)
+ else query_vector, dtype=np.float32)
+
+ return _numpy_topk(row_id_array, stored_matrix, query_np, metric, limit)
+
+
+def _raw_search_from_arrow(arrow_table, vector_column_name, query_vector,
+ metric, limit, score_candidates=None):
+ """Vectorized raw search directly from Arrow table (avoids Python list
intermediary)."""
+ import numpy as np
+ import pyarrow.compute as pc
+
+ row_ids_col = arrow_table.column(SpecialFields.ROW_ID.name)
+ vectors_col = arrow_table.column(vector_column_name)
+
+ # Filter out null vectors at the Arrow level before conversion.
+ valid_mask = pc.is_valid(vectors_col)
+ if not pc.all(valid_mask).as_py():
+ arrow_table = arrow_table.filter(valid_mask)
+ row_ids_col = arrow_table.column(SpecialFields.ROW_ID.name)
+ vectors_col = arrow_table.column(vector_column_name)
+
+ # Try fast path: fixed-size list → direct numpy reshape.
+ row_id_array = row_ids_col.to_numpy()
+ try:
+ # ChunkedArray has no .values; combine to a single array first.
+ if hasattr(vectors_col, 'combine_chunks'):
+ vectors_arr = vectors_col.combine_chunks()
+ else:
+ vectors_arr = vectors_col
+ flat = vectors_arr.values
+ dim = vectors_arr.type.list_size
+ if dim is not None and flat is not None:
+ stored_matrix = flat.to_numpy(zero_copy_only=False).reshape(-1,
dim).astype(
+ np.float32)
+ else:
+ stored_matrix = np.array(vectors_col.to_pylist(), dtype=np.float32)
+ except (AttributeError, TypeError, ValueError):
+ stored_matrix = np.array(vectors_col.to_pylist(), dtype=np.float32)
+
+ query_np = np.asarray(query_vector, dtype=np.float32)
+
+ if stored_matrix.shape[1] != query_np.shape[0]:
+ raise ValueError(
+ "Query vector dimension mismatch: expected %d, got %d"
+ % (stored_matrix.shape[1], query_np.shape[0]))
+
+ # Handle null vectors and score_candidates filtering.
+ if score_candidates is not None:
+ candidate_set = set(score_candidates)
+ mask = np.array([rid in candidate_set for rid in row_id_array],
dtype=bool)
+ # Also mask null vectors (check for any NaN row).
+ null_mask = ~np.isnan(stored_matrix).any(axis=1)
+ mask = mask & null_mask
+ row_id_array = row_id_array[mask]
+ stored_matrix = stored_matrix[mask]
+ else:
+ null_mask = ~np.isnan(stored_matrix).any(axis=1)
+ if not null_mask.all():
+ row_id_array = row_id_array[null_mask]
+ stored_matrix = stored_matrix[null_mask]
+
+ if len(row_id_array) == 0:
+ return DictBasedScoredIndexResult({})
+
+ return _numpy_topk(row_id_array, stored_matrix, query_np, metric, limit)
+
+
+def _numpy_topk(row_id_array, stored_matrix, query_np, metric, limit):
+ """Core numpy distance computation + topK selection."""
+ import numpy as np
+
+ if metric == "l2":
+ diffs = stored_matrix - query_np
+ dists = np.sum(diffs * diffs, axis=1)
+ scores = 1.0 / (1.0 + dists)
+ elif metric == "cosine":
+ dots = stored_matrix @ query_np
+ norms = np.linalg.norm(stored_matrix, axis=1) *
np.linalg.norm(query_np)
+ norms = np.where(norms == 0, 1.0, norms)
+ scores = dots / norms
+ elif metric == "inner_product":
+ scores = stored_matrix @ query_np
+ else:
+ raise ValueError("Unknown vector search metric: %s" % metric)
+
+ n = len(scores)
+ if n <= limit:
+ top_indices = np.argsort(-scores)
+ else:
+ top_indices = np.argpartition(-scores, limit)[:limit]
+ top_indices = top_indices[np.argsort(-scores[top_indices])]
+
+ return DictBasedScoredIndexResult(
+ {int(row_id_array[i]): float(scores[i]) for i in top_indices}
+ )
+
+
+def _raw_batch_search_from_arrow(arrow_table, vector_column_name,
query_vectors,
+ metric, limit, score_candidates=None):
+ """Batch raw search: multiple queries against the same Arrow table in one
SGEMM call."""
+ import numpy as np
+ import pyarrow.compute as pc
+
+ row_ids_col = arrow_table.column(SpecialFields.ROW_ID.name)
+ vectors_col = arrow_table.column(vector_column_name)
+
+ valid_mask = pc.is_valid(vectors_col)
+ if not pc.all(valid_mask).as_py():
+ arrow_table = arrow_table.filter(valid_mask)
+ row_ids_col = arrow_table.column(SpecialFields.ROW_ID.name)
+ vectors_col = arrow_table.column(vector_column_name)
+
+ row_id_array = row_ids_col.to_numpy()
+ try:
+ if hasattr(vectors_col, 'combine_chunks'):
+ vectors_arr = vectors_col.combine_chunks()
+ else:
+ vectors_arr = vectors_col
+ flat = vectors_arr.values
+ dim = vectors_arr.type.list_size
+ if dim is not None and flat is not None:
+ stored_matrix = flat.to_numpy(zero_copy_only=False).reshape(-1,
dim).astype(
+ np.float32)
+ else:
+ stored_matrix = np.array(vectors_col.to_pylist(), dtype=np.float32)
+ except (AttributeError, TypeError, ValueError):
+ stored_matrix = np.array(vectors_col.to_pylist(), dtype=np.float32)
+
+ query_matrix = np.array(
+ [q if isinstance(q, np.ndarray) else list(q) for q in query_vectors],
+ dtype=np.float32)
+
+ if stored_matrix.shape[1] != query_matrix.shape[1]:
+ raise ValueError(
+ "Query vector dimension mismatch: expected %d, got %d"
+ % (stored_matrix.shape[1], query_matrix.shape[1]))
+
+ if score_candidates is not None:
+ candidate_set = set(score_candidates)
+ mask = np.array([rid in candidate_set for rid in row_id_array],
dtype=bool)
+ null_mask = ~np.isnan(stored_matrix).any(axis=1)
+ mask = mask & null_mask
+ row_id_array = row_id_array[mask]
+ stored_matrix = stored_matrix[mask]
+ else:
+ null_mask = ~np.isnan(stored_matrix).any(axis=1)
+ if not null_mask.all():
+ row_id_array = row_id_array[null_mask]
+ stored_matrix = stored_matrix[null_mask]
+
+ if len(row_id_array) == 0:
+ return [DictBasedScoredIndexResult({}) for _ in
range(len(query_vectors))]
+
+ return _numpy_batch_topk(row_id_array, stored_matrix, query_matrix,
metric, limit)
+
+
+def _numpy_batch_topk(row_id_array, stored_matrix, query_matrix, metric,
limit):
+ """Batch SGEMM distance computation + per-query topK. Reuses stored
norms."""
+ import numpy as np
+
+ n_queries = query_matrix.shape[0]
+
+ if metric == "l2":
+ stored_sq = np.sum(stored_matrix * stored_matrix, axis=1,
keepdims=True)
+ query_sq = np.sum(query_matrix * query_matrix, axis=1, keepdims=True)
+ dots = stored_matrix @ query_matrix.T
+ dists = stored_sq + query_sq.T - 2 * dots
Review Comment:
Fixed. The L2 branch now uses direct subtraction (sum((a-b)²)) instead of
the norm-expansion formula.
Cosine and inner_product still use SGEMM — they don't suffer from this issue
because their computation structure has no large-magnitude subtraction:
- inner_product: pure dot product (a·b), only accumulation, no cancellation
risk.
- cosine: (a·b) / (||a|| × ||b||) is a ratio where numerator and denominator
are well-matched in scale — there's no "big minus big yields small" pattern
that destroys significant digits.
The L2 problem is specific to ||a||² + ||b||² - 2a·b where the first two
terms and the third are close in magnitude, causing catastrophic cancellation
in float32.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
