vecindex: implement C-SPANN search

Add the VectorIndex class, which will implement the C-SPANN algorithm, which
adapts Microsoft’s SPANN and SPFresh algorithms to work well with CockroachDB’s
unique distributed architecture. This PR implements K-means tree search with a
test implementation of bottom-up tree construction. Later PR's will include code
to incrementally build the tree.

Epic: CRDB-42943

Release note: None

pkg/sql/vecindex/BUILD.bazel

@@ -8,11 +8,16 @@ filegroup(
 
 go_library(
     name = "vecindex",
-    srcs = ["kmeans.go"],
+    srcs = [
+        "kmeans.go",
+        "vector_index.go",
+    ],
     importpath = "github.com/cockroachdb/cockroach/pkg/sql/vecindex",
     visibility = ["//visibility:public"],
     deps = [
         "//pkg/sql/vecindex/internal",
+        "//pkg/sql/vecindex/quantize",
+        "//pkg/sql/vecindex/vecstore",
         "//pkg/util/num32",
         "//pkg/util/vector",
         "@com_github_cockroachdb_errors//:errors",
@@ -21,14 +26,21 @@ go_library(
 
 go_test(
     name = "vecindex_test",
-    srcs = ["kmeans_test.go"],
+    srcs = [
+        "kmeans_test.go",
+        "vector_index_test.go",
+    ],
     data = glob(["testdata/**"]),
     embed = [":vecindex"],
     deps = [
         "//pkg/sql/vecindex/internal",
+        "//pkg/sql/vecindex/quantize",
         "//pkg/sql/vecindex/testutils",
+        "//pkg/sql/vecindex/vecstore",
         "//pkg/util/num32",
         "//pkg/util/vector",
+        "@com_github_cockroachdb_datadriven//:datadriven",
+        "@com_github_cockroachdb_errors//:errors",
         "@com_github_stretchr_testify//require",
         "@org_gonum_v1_gonum//floats/scalar",
         "@org_gonum_v1_gonum//stat",

pkg/sql/vecindex/quantize/rabitq_test.go

@@ -106,8 +106,10 @@ func TestRaBitQuantizerEdge(t *testing.T) {
 
 		vectors := vector.MakeSet(141)
 		vectors.AddUndefined(2)
+		zeros := vectors.At(0)
 		ones := vectors.At(1)
 		for i := 0; i < len(ones); i++ {
+			zeros[i] = 0
 			ones[i] = 1
 		}
 		quantizedSet := quantizer.Quantize(ctx, &vectors).(*RaBitQuantizedVectorSet)

pkg/sql/vecindex/testdata/delete.ddt

@@ -0,0 +1,90 @@
+# ----------
+# Test deleting vectors from primary index, but not from secondary index.
+# ----------
+new-index min-partition-size=1 max-partition-size=3 beam-size=2
+vec1: (1, 2)
+vec2: (7, 4)
+vec3: (4, 3)
+vec4: (5, 5)
+----
+• 1 (3.1667, 3)
+│
+├───• 2 (5.3333, 4)
+│   │
+│   ├───• vec2 (7, 4)
+│   ├───• vec3 (4, 3)
+│   └───• vec4 (5, 5)
+│
+└───• 3 (1, 2)
+    │
+    └───• vec1 (1, 2)
+
+# Delete vector from primary index, but not from secondary index.
+delete not-found
+vec3
+----
+• 1 (3.1667, 3)
+│
+├───• 2 (5.3333, 4)
+│   │
+│   ├───• vec2 (7, 4)
+│   ├───• vec3 (MISSING)
+│   └───• vec4 (5, 5)
+│
+└───• 3 (1, 2)
+    │
+    └───• vec1 (1, 2)
+
+# Ensure deleted vector is not returned by search. This should enqueue a fixup
+# that removes the vector from the index.
+search max-results=1
+(4, 3)
+----
+vec4: 5 (centroid=1.0541)
+4 leaf vectors, 6 vectors, 2 full vectors, 3 partitions
+
+# Again, with higher max results.
+search max-results=2
+(4, 3)
+----
+vec4: 5 (centroid=1.0541)
+vec2: 10 (centroid=1.6667)
+4 leaf vectors, 6 vectors, 4 full vectors, 3 partitions
+
+# Vector should now be gone from the index.
+# TODO(andyk): This will be true once fixups are added.
+format-tree
+----
+• 1 (3.1667, 3)
+│
+├───• 2 (5.3333, 4)
+│   │
+│   ├───• vec2 (7, 4)
+│   ├───• vec3 (MISSING)
+│   └───• vec4 (5, 5)
+│
+└───• 3 (1, 2)
+    │
+    └───• vec1 (1, 2)
+
+# Delete all vectors from one branch of the tree.
+delete not-found
+vec1
+----
+• 1 (3.1667, 3)
+│
+├───• 2 (5.3333, 4)
+│   │
+│   ├───• vec2 (7, 4)
+│   ├───• vec3 (MISSING)
+│   └───• vec4 (5, 5)
+│
+└───• 3 (1, 2)
+    │
+    └───• vec1 (MISSING)
+
+# Search the empty branch.
+search max-results=1 beam-size=1
+(1, 2)
+----
+1 leaf vectors, 3 vectors, 1 full vectors, 2 partitions

pkg/sql/vecindex/testdata/search-features.ddt

@@ -0,0 +1,72 @@
+# Load 500 512-dimension features and search them. Use small partition size to
+# ensure a deeper tree.
+
+new-index dims=512 min-partition-size=2 max-partition-size=8 quality-samples=4 beam-size=2 load-features=500 hide-tree
+----
+Created index with 500 vectors with 512 dimensions.
+
+# Start with 1 result and default beam size of 2.
+search max-results=1 use-feature=9999
+----
+vec441: 0.4646 (centroid=0.382)
+9 leaf vectors, 39 vectors, 2 full vectors, 5 partitions
+
+# Search for additional results.
+search max-results=3 use-feature=9999
+----
+vec441: 0.4646 (centroid=0.382)
+vec99: 0.6356 (centroid=0.382)
+vec296: 0.7638 (centroid=0.5962)
+9 leaf vectors, 39 vectors, 6 full vectors, 5 partitions
+
+# Use a larger beam size.
+search max-results=6 use-feature=9999 beam-size=8
+----
+vec74: 0.4155 (centroid=0.5092)
+vec195: 0.4359 (centroid=0.5127)
+vec441: 0.4646 (centroid=0.382)
+vec77: 0.4894 (centroid=0.4286)
+vec355: 0.5821 (centroid=0.4617)
+vec328: 0.6032 (centroid=0.5276)
+58 leaf vectors, 123 vectors, 14 full vectors, 15 partitions
+
+# Turn off re-ranking, which results in increased inaccuracy.
+search max-results=6 use-feature=9999 beam-size=8 skip-rerank
+----
+vec195: 0.4179 ±0.0264 (centroid=0.5127)
+vec74: 0.4322 ±0.0263 (centroid=0.5092)
+vec441: 0.4657 ±0.0215 (centroid=0.382)
+vec77: 0.4881 ±0.0221 (centroid=0.4286)
+vec355: 0.5658 ±0.0238 (centroid=0.4617)
+vec415: 0.6142 ±0.0302 (centroid=0.5306)
+58 leaf vectors, 123 vectors, 0 full vectors, 15 partitions
+
+# Return top 25 results.
+search max-results=25 use-feature=9999 beam-size=8
+----
+vec74: 0.4155 (centroid=0.5092)
+vec195: 0.4359 (centroid=0.5127)
+vec441: 0.4646 (centroid=0.382)
+vec77: 0.4894 (centroid=0.4286)
+vec355: 0.5821 (centroid=0.4617)
+vec328: 0.6032 (centroid=0.5276)
+vec389: 0.6183 (centroid=0.5267)
+vec415: 0.6298 (centroid=0.5306)
+vec99: 0.6356 (centroid=0.382)
+vec267: 0.6742 (centroid=0.526)
+vec6: 0.685 (centroid=0.6015)
+vec485: 0.6867 (centroid=0.362)
+vec236: 0.687 (centroid=0.5071)
+vec198: 0.6885 (centroid=0.5094)
+vec65: 0.6898 (centroid=0.4403)
+vec146: 0.6901 (centroid=0.5601)
+vec282: 0.7197 (centroid=0.4023)
+vec410: 0.728 (centroid=0.4261)
+vec356: 0.7341 (centroid=0.4352)
+vec439: 0.7428 (centroid=0.6023)
+vec116: 0.7462 (centroid=0.4643)
+vec273: 0.7555 (centroid=0.5226)
+vec453: 0.7735 (centroid=0.3571)
+vec233: 0.7737 (centroid=0.5502)
+vec331: 0.7793 (centroid=0.4871)
+58 leaf vectors, 123 vectors, 44 full vectors, 15 partitions