apexbase 1.25.0

High-performance HTAP embedded database with Rust core
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251

#!/usr/bin/env python3
"""
TopK distance benchmark: ApexBase vs SimSIMD (in-memory baseline).

SimSIMD is the fastest SIMD-accelerated distance library on CPU.
ApexBase reads from on-disk storage (binary-encoded float32 columns).

The purpose is to measure how close ApexBase's fused O(n log k) heap
gets to the theoretical in-memory SIMD ceiling.

Usage:
    python benchmarks/bench_simsimd.py [--rows N] [--dim D] [--k K] [--iters I]

Requires: simsimd  (pip install simsimd)
"""

import argparse
import gc
import tempfile
import time

import numpy as np
import simsimd


# ─────────────────────────────────────────────────────────────────────────────

def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument("--rows",  type=int, default=1_000_000)
    p.add_argument("--dim",   type=int, default=128)
    p.add_argument("--k",     type=int, default=10)
    p.add_argument("--iters", type=int, default=10)
    return p.parse_args()


def fmt(ms: float) -> str:
    return f"{ms:.3f} ms"


def median(vals):
    s = sorted(vals)
    n = len(s)
    return s[n // 2] if n % 2 else (s[n // 2 - 1] + s[n // 2]) / 2


# ─────────────────────────────────────────────────────────────────────────────
# SimSIMD topk helper
# ─────────────────────────────────────────────────────────────────────────────

def simsimd_topk(matrix: np.ndarray, query: np.ndarray, k: int, metric: str) -> np.ndarray:
    """Compute topk row indices by distance using SimSIMD + np.argpartition."""
    if metric == "l2":
        dists = np.asarray(simsimd.euclidean(query, matrix))
    elif metric == "l2_squared":
        dists = np.asarray(simsimd.sqeuclidean(query, matrix))
    elif metric == "cosine":
        dists = np.asarray(simsimd.cosine(query, matrix))
    elif metric == "dot":
        # inner product: larger = better, so negate for topk ascending
        dists = -np.asarray(simsimd.inner(query, matrix))
    else:
        raise ValueError(metric)

    k = min(k, len(dists))
    idx = np.argpartition(dists, k)[:k]
    idx = idx[np.argsort(dists[idx])]   # sort the k candidates
    return idx, dists[idx]


# ─────────────────────────────────────────────────────────────────────────────
# ApexBase setup & bench
# ─────────────────────────────────────────────────────────────────────────────

def setup_apex(vecs: np.ndarray, tmp_dir: str):
    from apexbase.client import ApexClient, encode_vector
    client = ApexClient(dirpath=tmp_dir, drop_if_exists=True)
    client.create_table("vecs")
    batch = 10_000
    n = len(vecs)
    for s in range(0, n, batch):
        client.store([{"vec": encode_vector(vecs[i])} for i in range(s, min(s + batch, n))])
    return client


def bench_apex(client, query: np.ndarray, k: int, metric: str, iters: int):
    # Apex metric names
    _m = {"l2": "l2", "l2_squared": "l2_squared", "cosine": "cosine_distance", "dot": "dot"}
    apex_metric = _m[metric]
    # One warmup to populate OS page cache (NOT timed)
    client.topk_distance("vec", query, k=k, metric=apex_metric)
    times = []
    for _ in range(iters):
        gc.collect()
        t0 = time.perf_counter()
        client.topk_distance("vec", query, k=k, metric=apex_metric)
        times.append((time.perf_counter() - t0) * 1000)
    return times


def bench_simsimd(matrix: np.ndarray, query: np.ndarray, k: int, metric: str, iters: int):
    times = []
    for _ in range(iters):
        gc.collect()
        t0 = time.perf_counter()
        simsimd_topk(matrix, query, k, metric)
        times.append((time.perf_counter() - t0) * 1000)
    return times


# ─────────────────────────────────────────────────────────────────────────────
# Correctness check
# ─────────────────────────────────────────────────────────────────────────────

def verify(client, matrix: np.ndarray, query: np.ndarray, k: int):
    """Check that ApexBase and SimSIMD return the same top-k set for L2."""
    from apexbase.client import encode_vector

    # ApexBase result (_id values)
    apex_rows = client.topk_distance("vec", query, k=k, metric="l2").to_dict()
    apex_ids = {r["_id"] for r in apex_rows}

    # SimSIMD result (row indices = _id since we inserted in order 0..n-1)
    ss_idx, _ = simsimd_topk(matrix, query, k, "l2")
    ss_ids = set(int(i) for i in ss_idx)

    if apex_ids == ss_ids:
        return True, None
    return False, (apex_ids, ss_ids)


# ─────────────────────────────────────────────────────────────────────────────
# Main
# ─────────────────────────────────────────────────────────────────────────────

def main():
    args = parse_args()
    rng  = np.random.default_rng(42)
    vecs = rng.random((args.rows, args.dim), dtype=np.float32)
    query = rng.random(args.dim, dtype=np.float32)

    caps = simsimd.get_capabilities()
    active = [k for k, v in caps.items() if v]

    print(f"\n{'='*72}")
    print(f"TopK Distance Benchmark: ApexBase (on-disk) vs SimSIMD (in-RAM)")
    print(f"  rows={args.rows:,}  dim={args.dim}  k={args.k}  iters={args.iters}")
    print(f"  SimSIMD capabilities: {', '.join(active)}")
    print(f"{'='*72}\n")

    with tempfile.TemporaryDirectory() as apex_dir:
        print("Setting up ApexBase (storing vectors to disk)...", end="", flush=True)
        client = setup_apex(vecs, apex_dir)
        print(" done.")

        print("SimSIMD uses the in-memory float32 matrix directly.\n")

        # OS page-cache warmup: read through all data once (not timed)
        print("Warming up OS page cache...", end="", flush=True)
        client.topk_distance("vec", query, k=args.k, metric="l2")
        print(" done.\n")

        # ── Correctness ──────────────────────────────────────────────────────
        ok, diff = verify(client, vecs, query, args.k)
        status = "✅ match" if ok else f"❌ MISMATCH apex={diff[0]} simsimd={diff[1]}"
        print(f"Correctness check (L2, k={args.k}): {status}\n")

        # ── Head-to-head ─────────────────────────────────────────────────────
        metrics = [
            ("l2",         "L2 / Euclidean",     "euclidean"),
            ("l2_squared", "L2 Squared",          "sqeuclidean"),
            ("cosine",     "Cosine distance",     "cosine"),
            ("dot",        "Dot / Inner product", "inner (neg)"),
        ]

        w = 22
        hdr = f"{'Metric':<{w}} {'ApexBase':>12} {'SimSIMD':>12} {'Ratio A/S':>12}  Note"
        print(hdr)
        print("─" * len(hdr))

        for metric, label, ss_label in metrics:
            apex_t  = bench_apex(client, query, args.k, metric, args.iters)
            simsimd_t = bench_simsimd(vecs, query, args.k, metric, args.iters)

            a_med = median(apex_t)
            s_med = median(simsimd_t)
            ratio = a_med / s_med
            flag  = "✅" if ratio < 2.0 else ("⚠️" if ratio < 5.0 else "❌")

            note = f"(SimSIMD: {ss_label})"
            print(f"{label:<{w}} {fmt(a_med):>12} {fmt(s_med):>12} {ratio:>10.2f}x {flag}  {note}")

        print()

        # ── Pipeline breakdown (L2) ──────────────────────────────────────────
        print("── Pipeline breakdown: what each step costs (L2) ──")

        def timed(fn, iters):
            results = []
            for _ in range(iters):
                gc.collect()
                t0 = time.perf_counter()
                fn()
                results.append((time.perf_counter() - t0) * 1000)
            return median(results)

        k = args.k
        n = args.rows

        # 1. SimSIMD: distance-only (no topk)
        t_simd_dist = timed(lambda: simsimd.euclidean(query, vecs), args.iters)

        # 2. SimSIMD: distance + argpartition topk
        t_simd_topk = timed(lambda: simsimd_topk(vecs, query, k, "l2"), args.iters)

        # 3. Numpy baseline: L2 norm via numpy + argpartition
        t_np_topk = timed(
            lambda: (lambda d: d[np.argpartition(d, k)[:k]])(
                np.linalg.norm(vecs - query, axis=1).astype(np.float32)
            ),
            args.iters,
        )

        # 4. ApexBase full pipeline (storage decode + heap)
        t_apex = timed(lambda: client.topk_distance("vec", query, k=k, metric="l2"), args.iters)

        print(f"  {'Step':<42} {'Time':>10}  {'M vec/s':>10}")
        print(f"  {'-'*64}")
        print(f"  {'SimSIMD euclidean() — pure distance, no topk':<42} {fmt(t_simd_dist):>10}  {n/(t_simd_dist/1000)/1e6:>9.1f}")
        print(f"  {'SimSIMD + argpartition topk':<42} {fmt(t_simd_topk):>10}  {n/(t_simd_topk/1000)/1e6:>9.1f}")
        print(f"  {'NumPy linalg.norm + argpartition topk':<42} {fmt(t_np_topk):>10}  {n/(t_np_topk/1000)/1e6:>9.1f}")
        print(f"  {'ApexBase topk_distance (storage+decode+heap)':<42} {fmt(t_apex):>10}  {n/(t_apex/1000)/1e6:>9.1f}")
        print()

        decode_overhead = max(t_apex - t_simd_topk, 0.0)
        print(f"  Storage/decode overhead vs SimSIMD+topk: {fmt(decode_overhead)} ({decode_overhead/t_apex*100:.0f}% of ApexBase time)")
        print()

        # ── min/med/max detail ────────────────────────────────────────────────
        print("── Per-run detail (min / med / max) ──")
        for metric, label, _ in metrics:
            at = bench_apex(client, query, args.k, metric, args.iters)
            st = bench_simsimd(vecs, query, args.k, metric, args.iters)
            print(f"  {label:<22}  ApexBase {fmt(min(at))}/{fmt(median(at))}/{fmt(max(at))}"
                  f"   SimSIMD {fmt(min(st))}/{fmt(median(st))}/{fmt(max(st))}")

        client.close()


if __name__ == "__main__":
    main()