feat: add VADD_BATCH command for bulk vector inserts

README.md

@@ -159,6 +159,7 @@ cargo build --release --features vector
 | command | description |
 |---------|-------------|
 | `VADD key element f32 [f32 ...] [METRIC COSINE\|L2\|IP] [QUANT F32\|F16\|Q8] [M n] [EF n]` | add a vector to the set |
+| `VADD_BATCH key DIM n elem1 f32... elem2 f32... [METRIC\|QUANT\|M\|EF]` | add multiple vectors in one command |
 | `VSIM key f32 [f32 ...] COUNT k [EF n] [WITHSCORES]` | k nearest neighbors |
 | `VREM key element` | remove a vector |
 | `VGET key element` | retrieve stored vector values |
@@ -303,7 +304,7 @@ contributions welcome — see [CONTRIBUTING.md](CONTRIBUTING.md).
 | 4 | clustering (raft, gossip, slots, migration) | ✅ complete |
 | 5 | developer experience (observability, CLI, clients) | 🚧 in progress |
 
-**current**: 106 commands, 796+ tests, ~31k lines of code (excluding tests)
+**current**: 107 commands, 796+ tests, ~31k lines of code (excluding tests)
 
 ## security
 

bench/README.md

@@ -52,7 +52,7 @@ tested on GCP c2-standard-8 (8 vCPU Intel Xeon @ 3.10GHz), Ubuntu 22.04.
 
 dragonfly in particular offers features ember simply doesn't have:
 
-- full Redis API compatibility (200+ commands vs ember's ~106)
+- full Redis API compatibility (200+ commands vs ember's ~107)
 - sophisticated memory management (dashtable for ~25% of Redis memory usage)
 - transactional semantics (MULTI/EXEC, Lua scripting)
 - fork-free snapshotting
@@ -121,13 +121,13 @@ HNSW index: M=16, ef_construction=64 for all systems. tested on GCP c2-standard-
 
 | metric | ember (RESP) | ember (gRPC) | chromadb | pgvector | qdrant |
 |--------|-------------|-------------|----------|----------|--------|
-| insert (vectors/sec) | 963 | 1,009 | **3,891** | 1,617 | **7,747** |
-| query (queries/sec) | 1,264 | **1,462** | 390 | 831 | 596 |
-| query p50 (ms) | 0.79ms | **0.68ms** | 2.56ms | 1.18ms | 1.67ms |
-| query p99 (ms) | 0.93ms | **0.83ms** | 2.76ms | 1.56ms | 1.88ms |
-| memory (MB) | **36 MB** | — | 122 MB | 179 MB | 121 MB |
+| insert (vectors/sec) | 1,483 | 2,374 | 3,679 | 1,513 | **7,382** |
+| query (queries/sec) | 1,212 | **1,452** | 383 | 843 | 589 |
+| query p50 (ms) | 0.82ms | **0.68ms** | 2.60ms | 1.16ms | 1.69ms |
+| query p99 (ms) | 1.00ms | **0.86ms** | 2.82ms | 1.61ms | 1.93ms |
+| memory (MB) | **38 MB** | — | 123 MB | 179 MB | 122 MB |
 
-ember's query throughput is 3.2x chromadb, 1.5x pgvector, and 2.1x qdrant, with 3-5x lower memory usage. gRPC queries are 16% faster than RESP due to lower serialization overhead. insert throughput is lower due to per-vector protocol overhead — qdrant's batch API gives it a significant edge on ingestion.
+ember's query throughput is 3.2x chromadb, 1.4x pgvector, and 2.1x qdrant, with 3-5x lower memory usage. gRPC queries are 20% faster than RESP due to lower serialization overhead. insert throughput uses VADD_BATCH (batches of 500 vectors per command) — the gRPC path benefits most since packed floats avoid string parsing entirely.
 
 #### SIFT1M recall accuracy (128-dim, 1M vectors, 10k queries)
 

bench/bench-memory.sh

@@ -326,18 +326,22 @@ dim = int(sys.argv[3])
 
 r = redis.Redis(host="127.0.0.1", port=port, decode_responses=True)
 
-for i in range(count):
-    vec = [random.gauss(0, 1) for _ in range(dim)]
-    norm = sum(v * v for v in vec) ** 0.5
-    vec = [v / norm for v in vec]
-
-    args = ["VADD", "vectors", f"v{i}"] + [str(v) for v in vec]
-    if i == 0:
+batch_size = 500
+for start in range(0, count, batch_size):
+    end = min(start + batch_size, count)
+    args = ["vectors", "DIM", str(dim)]
+    for i in range(start, end):
+        vec = [random.gauss(0, 1) for _ in range(dim)]
+        norm = sum(v * v for v in vec) ** 0.5
+        vec = [v / norm for v in vec]
+        args.append(f"v{i}")
+        args.extend(str(v) for v in vec)
+    if start == 0:
         args += ["METRIC", "COSINE"]
-    r.execute_command(*args)
+    r.execute_command("VADD_BATCH", *args)
 
-    if (i + 1) % 10000 == 0:
-        print(f"    inserted {i + 1}/{count} vectors", file=sys.stderr)
+    if end % 10000 == 0 or end == count:
+        print(f"    inserted {end}/{count} vectors", file=sys.stderr)
 PYEOF
 
         # vector requires sharded mode

bench/bench-vector.py

@@ -75,14 +75,13 @@ def setup(self, dim: int, metric: str = "cosine"):
         self.conn.delete(self.key)
 
     def insert_batch(self, ids: list, vectors: np.ndarray):
-        pipe = self.conn.pipeline(transaction=False)
+        dim = vectors.shape[1]
+        args = [self.key, "DIM", str(dim)]
         for i, vid in enumerate(ids):
-            vec = vectors[i]
-            # VADD key element v1 v2 ... METRIC COSINE M 16 EF 64
-            args = [self.key, vid] + [str(float(v)) for v in vec]
-            args += ["METRIC", "COSINE", "M", "16", "EF", "64"]
-            pipe.execute_command("VADD", *args)
-        pipe.execute()
+            args.append(vid)
+            args.extend(str(float(v)) for v in vectors[i])
+        args += ["METRIC", "COSINE", "M", "16", "EF", "64"]
+        self.conn.execute_command("VADD_BATCH", *args)
 
     def query(self, vector: np.ndarray, k: int) -> list:
         args = [self.key] + [str(float(v)) for v in vector]
@@ -116,9 +115,8 @@ def setup(self, dim: int, metric: str = "cosine"):
         self.client.flushdb()
 
     def insert_batch(self, ids: list, vectors: np.ndarray):
-        for i, vid in enumerate(ids):
-            vec = vectors[i].tolist()
-            self.client.vadd(self.key, vid, vec, metric="cosine", m=16, ef=64)
+        entries = [(vid, vectors[i].tolist()) for i, vid in enumerate(ids)]
+        self.client.vadd_batch(self.key, entries, metric="cosine", m=16, ef=64)
 
     def query(self, vector: np.ndarray, k: int) -> list:
         results = self.client.vsim(self.key, vector.tolist(), count=k)

bench/bench-vector.sh

@@ -242,10 +242,10 @@ elif [[ "$QDRANT" == "true" ]] && ! python3 -c "import qdrant_client" 2>/dev/nul
     pip install --quiet $REQUIRED_DEPS
 fi
 
-# install ember-py for gRPC benchmarks
+# install ember-py for gRPC benchmarks (also install base deps into venv)
 if [[ "$EMBER_GRPC" == "true" ]]; then
     ensure_venv
-    pip install --quiet ./clients/ember-py
+    pip install --quiet $REQUIRED_DEPS ./clients/ember-py
 fi
 
 # build ember if needed