|
| 1 | +""" |
| 2 | +Test Qdrant's accuracy in scenarios of continuous updates of real data. |
| 3 | +
|
| 4 | +
|
| 5 | +This script will: |
| 6 | +
|
| 7 | +- Create a Qdrant collection, and make initial upload of all available vectors from `data/dataset1` |
| 8 | +- Measure the accuracy of the search |
| 9 | +- Start replacing vectors of collection by removing points and replacing them with new ones from `data/dataset2` |
| 10 | +- Once finished, measure the accuracy of the search |
| 11 | +
|
| 12 | +""" |
| 13 | + |
| 14 | +import json |
| 15 | +import os |
| 16 | +import random |
| 17 | +import sys |
| 18 | +import time |
| 19 | +from datetime import datetime |
| 20 | +from pathlib import Path |
| 21 | + |
| 22 | +import numpy as np |
| 23 | +import tqdm |
| 24 | +from qdrant_client import QdrantClient, models |
| 25 | + |
| 26 | +QDRANT_COLLECTION_NAME = "benchmark" |
| 27 | +OUTPUT_FILENAME = os.getenv("OUTPUT_FILENAME", "output.json") |
| 28 | +DATASET_DIM = int(os.getenv("DATASET_DIM", 512)) |
| 29 | +DATASET_NAME = os.getenv("DATASET_NAME", "laion-small-clip-no-filters-1") |
| 30 | +DATASET_NAME_2 = os.getenv("DATASET_NAME_2", "laion-small-clip-no-filters-2") |
| 31 | +DATA_DIR = Path(__file__).parent / "data" / DATASET_NAME |
| 32 | +DATA_DIR_2 = Path(__file__).parent / "data" / DATASET_NAME_2 |
| 33 | + |
| 34 | +VECTORS_FILE_2 = DATA_DIR_2 / "vectors.npy" |
| 35 | +VECTORS_FILE_1 = DATA_DIR / "vectors.npy" |
| 36 | + |
| 37 | +TEST_DATA_FILE_2 = DATA_DIR_2 / "tests.jsonl" |
| 38 | +TEST_DATA_FILE_1 = DATA_DIR / "tests.jsonl" |
| 39 | + |
| 40 | +TOTAL_VECTORS = 100_000 |
| 41 | +BATCH_SIZE = 500 |
| 42 | + |
| 43 | + |
| 44 | +def read_test_data(file: Path, limit: int = 1000): |
| 45 | + """ |
| 46 | + { |
| 47 | + "query": [ |
| 48 | + 0.022043373435735703, |
| 49 | + -0.022230295464396477, |
| 50 | + .... |
| 51 | + ], |
| 52 | + "closest_ids": [ |
| 53 | + 43749, |
| 54 | + 43756, |
| 55 | + .... |
| 56 | + ] |
| 57 | + } |
| 58 | + """ |
| 59 | + with open(file, "r") as f: |
| 60 | + for idx, line in enumerate(f): |
| 61 | + if idx >= limit: |
| 62 | + break |
| 63 | + |
| 64 | + yield json.loads(line) |
| 65 | + |
| 66 | + |
| 67 | +class QdrantBenchmark: |
| 68 | + |
| 69 | + def __init__(self, url): |
| 70 | + |
| 71 | + client = QdrantClient(url=url, prefer_grpc=True) |
| 72 | + self.client = client |
| 73 | + |
| 74 | + self.client.delete_collection(QDRANT_COLLECTION_NAME) |
| 75 | + |
| 76 | + self.collection = self.client.create_collection( |
| 77 | + QDRANT_COLLECTION_NAME, |
| 78 | + vectors_config=models.VectorParams( |
| 79 | + size=DATASET_DIM, |
| 80 | + distance=models.Distance.COSINE, |
| 81 | + ), |
| 82 | + optimizers_config=models.OptimizersConfigDiff( |
| 83 | + deleted_threshold=0.001, |
| 84 | + vacuum_min_vector_number=100, |
| 85 | + ), |
| 86 | + ) |
| 87 | + |
| 88 | + def initial_upload(self, vectors: np.ndarray): |
| 89 | + self.client.upload_collection( |
| 90 | + collection_name=QDRANT_COLLECTION_NAME, |
| 91 | + vectors=vectors, |
| 92 | + ids=range(len(vectors)), |
| 93 | + ) |
| 94 | + |
| 95 | + def upload_points(self, vectors: np.ndarray, ids: list[int]): |
| 96 | + points = [ |
| 97 | + models.PointStruct(id=idx, vector=vectors[idx].tolist()) for idx in ids |
| 98 | + ] |
| 99 | + |
| 100 | + self.client.upsert( |
| 101 | + collection_name=QDRANT_COLLECTION_NAME, |
| 102 | + points=points, |
| 103 | + ) |
| 104 | + |
| 105 | + def validate_test_data(self, file: Path) -> float: |
| 106 | + total_results = 0 |
| 107 | + matched_results = 0 |
| 108 | + for test in tqdm.tqdm(read_test_data(file), desc="Validating test data"): |
| 109 | + query = test["query"] |
| 110 | + closest_ids = set(test["closest_ids"]) |
| 111 | + |
| 112 | + results = self.client.query_points( |
| 113 | + collection_name=QDRANT_COLLECTION_NAME, |
| 114 | + query=query, |
| 115 | + limit=len(closest_ids), |
| 116 | + ) |
| 117 | + |
| 118 | + results_idx = set(obj.id for obj in results.points) |
| 119 | + |
| 120 | + matched_results += len(closest_ids & results_idx) |
| 121 | + total_results += len(closest_ids) |
| 122 | + |
| 123 | + return matched_results / total_results |
| 124 | + |
| 125 | + def delete_points(self, points_to_delete: set): |
| 126 | + self.client.delete( |
| 127 | + collection_name=QDRANT_COLLECTION_NAME, |
| 128 | + points_selector=models.PointIdsList( |
| 129 | + points=[idx for idx in points_to_delete] |
| 130 | + ), |
| 131 | + ) |
| 132 | + |
| 133 | + def wait_ready(self) -> float: |
| 134 | + wait_interval = 0.2 |
| 135 | + confirmations_required = 2 |
| 136 | + |
| 137 | + start_time = time.time() |
| 138 | + confirmations = 0 |
| 139 | + first_green_time: float | None = None |
| 140 | + |
| 141 | + while True: |
| 142 | + collection_info = self.client.get_collection(QDRANT_COLLECTION_NAME) |
| 143 | + if collection_info.status == models.CollectionStatus.GREEN: |
| 144 | + confirmations += 1 |
| 145 | + first_green_time = first_green_time or time.time() |
| 146 | + if confirmations == confirmations_required: |
| 147 | + return first_green_time - start_time |
| 148 | + else: |
| 149 | + confirmations = 0 |
| 150 | + first_green_time = None |
| 151 | + time.sleep(wait_interval) |
| 152 | + |
| 153 | + def __del__(self): |
| 154 | + self.client.close() |
| 155 | + |
| 156 | + |
| 157 | +def store_to_file(data_dict): |
| 158 | + timestamped_dict = data_dict.copy() |
| 159 | + timestamped_dict["timestamp"] = datetime.now().isoformat() |
| 160 | + |
| 161 | + with open(OUTPUT_FILENAME, "w", encoding="utf-8") as f: |
| 162 | + json.dump(timestamped_dict, f, ensure_ascii=False) |
| 163 | + |
| 164 | + |
| 165 | +def main(): |
| 166 | + result = {} |
| 167 | + vectors_1 = np.load(VECTORS_FILE_1) |
| 168 | + vectors_2 = np.load(VECTORS_FILE_2) |
| 169 | + |
| 170 | + benchmark = QdrantBenchmark("http://localhost:6333") |
| 171 | + benchmark.initial_upload(vectors_1) |
| 172 | + benchmark.wait_ready() |
| 173 | + |
| 174 | + initial_precision = benchmark.validate_test_data(TEST_DATA_FILE_1) |
| 175 | + print("Precision dataset1: ", initial_precision) |
| 176 | + result["initial_precision"] = initial_precision |
| 177 | + result["precision_before_iteration"] = initial_precision |
| 178 | + |
| 179 | + points_to_migrate = list(range(TOTAL_VECTORS)) |
| 180 | + |
| 181 | + random.shuffle(points_to_migrate) |
| 182 | + |
| 183 | + total_indexing_time = 0 |
| 184 | + for i in tqdm.tqdm(range(0, len(points_to_migrate), BATCH_SIZE), desc="Iterating"): |
| 185 | + batch = points_to_migrate[i : i + BATCH_SIZE] |
| 186 | + |
| 187 | + benchmark.delete_points(set(batch)) |
| 188 | + |
| 189 | + benchmark.upload_points(vectors_2, batch) |
| 190 | + |
| 191 | + total_indexing_time += benchmark.wait_ready() |
| 192 | + |
| 193 | + print(f"Indexing: {total_indexing_time}") |
| 194 | + result["indexing_total_time_s"] = total_indexing_time |
| 195 | + |
| 196 | + precision_after_iteration = benchmark.validate_test_data(TEST_DATA_FILE_2) |
| 197 | + print(f"Precision dataset2: {precision_after_iteration}") |
| 198 | + result["precision_after_iteration"] = precision_after_iteration |
| 199 | + |
| 200 | + store_to_file(result) |
| 201 | + |
| 202 | + |
| 203 | +if __name__ == "__main__": |
| 204 | + sys.stdout.reconfigure(line_buffering=True) |
| 205 | + sys.stderr.reconfigure(line_buffering=True) |
| 206 | + |
| 207 | + main() |
| 208 | + |
| 209 | + sys.stdout.flush() |
| 210 | + sys.stderr.flush() |
![pre-commit-ci[bot]](https://avatars.githubusercontent.com/in/68672?v=4&size=40){kind=avatar}
0 commit comments