class Ingester:
def __init__(self,
host,
port,
collection,
collection_param,
partition=None,
drop=False,
batch_size=100,
dtype=np.float32):
self.collection = collection
self.partition = partition
self.client = Milvus(host, port)
self.dtype = dtype
self.batch_size = batch_size
if drop and self.client.has_collection(collection):
self.client.drop_collection(collection)
if not self.client.has_collection(collection):
self.client.create_collection(collection, collection_param)
if partition and not self.client.has_partition(collection, partition):
self.client.create_partition(collection, partition)
def ingest(self, entities, ids):
if self.partition:
return self.client.insert(self.collection,
entities,
ids=ids,
partition_tag=self.partition)
else:
return self.client.insert(self.collection, entities, ids=ids)
def _add_partition(self, collection: str, partition: str, milvus: mv.Milvus) -> None:
"""
创建场景类
@param {str} collection - 问题分类
@param {str} partition - 场景类
@param {mv.Milvus} milvus - Milvus连接对象
"""
_status, _exists = milvus.has_partition(collection, partition)
self.confirm_milvus_status(_status, 'has_partition')
if not _exists:
# 创建场景
self.confirm_milvus_status(
milvus.create_partition(collection, partition), 'create_partition'
)
class MilvusClient(object):
def __init__(self,
collection_name=None,
host=None,
port=None,
timeout=180):
self._collection_name = collection_name
start_time = time.time()
if not host:
host = SERVER_HOST_DEFAULT
if not port:
port = SERVER_PORT_DEFAULT
logger.debug(host)
logger.debug(port)
# retry connect remote server
i = 0
while time.time() < start_time + timeout:
try:
self._milvus = Milvus(host=host,
port=port,
try_connect=False,
pre_ping=False)
break
except Exception as e:
logger.error(str(e))
logger.error("Milvus connect failed: %d times" % i)
i = i + 1
time.sleep(i)
if time.time() > start_time + timeout:
raise Exception("Server connect timeout")
# self._metric_type = None
def __str__(self):
return 'Milvus collection %s' % self._collection_name
def check_status(self, status):
if not status.OK():
logger.error(status.message)
logger.error(self._milvus.server_status())
logger.error(self.count())
raise Exception("Status not ok")
def check_result_ids(self, result):
for index, item in enumerate(result):
if item[0].distance >= epsilon:
logger.error(index)
logger.error(item[0].distance)
raise Exception("Distance wrong")
# only support the given field name
def create_collection(self,
dimension,
data_type=DataType.FLOAT_VECTOR,
auto_id=False,
collection_name=None,
other_fields=None):
self._dimension = dimension
if not collection_name:
collection_name = self._collection_name
vec_field_name = utils.get_default_field_name(data_type)
fields = [{
"name": vec_field_name,
"type": data_type,
"params": {
"dim": dimension
}
}]
if other_fields:
other_fields = other_fields.split(",")
if "int" in other_fields:
fields.append({
"name": utils.DEFAULT_INT_FIELD_NAME,
"type": DataType.INT64
})
if "float" in other_fields:
fields.append({
"name": utils.DEFAULT_FLOAT_FIELD_NAME,
"type": DataType.FLOAT
})
create_param = {"fields": fields, "auto_id": auto_id}
try:
self._milvus.create_collection(collection_name, create_param)
logger.info("Create collection: <%s> successfully" %
collection_name)
except Exception as e:
logger.error(str(e))
raise
def create_partition(self, tag, collection_name=None):
if not collection_name:
collection_name = self._collection_name
self._milvus.create_partition(collection_name, tag)
def generate_values(self, data_type, vectors, ids):
values = None
if data_type in [DataType.INT32, DataType.INT64]:
values = ids
elif data_type in [DataType.FLOAT, DataType.DOUBLE]:
values = [(i + 0.0) for i in ids]
elif data_type in [DataType.FLOAT_VECTOR, DataType.BINARY_VECTOR]:
values = vectors
return values
def generate_entities(self, vectors, ids=None, collection_name=None):
entities = []
if collection_name is None:
collection_name = self._collection_name
info = self.get_info(collection_name)
for field in info["fields"]:
field_type = field["type"]
entities.append({
"name":
field["name"],
"type":
field_type,
"values":
self.generate_values(field_type, vectors, ids)
})
return entities
@time_wrapper
def insert(self, entities, ids=None, collection_name=None):
tmp_collection_name = self._collection_name if collection_name is None else collection_name
try:
insert_ids = self._milvus.insert(tmp_collection_name,
entities,
ids=ids)
return insert_ids
except Exception as e:
logger.error(str(e))
def get_dimension(self):
info = self.get_info()
for field in info["fields"]:
if field["type"] in [
DataType.FLOAT_VECTOR, DataType.BINARY_VECTOR
]:
return field["params"]["dim"]
def get_rand_ids(self, length):
segment_ids = []
while True:
stats = self.get_stats()
segments = stats["partitions"][0]["segments"]
# random choice one segment
segment = random.choice(segments)
try:
segment_ids = self._milvus.list_id_in_segment(
self._collection_name, segment["id"])
except Exception as e:
logger.error(str(e))
if not len(segment_ids):
continue
elif len(segment_ids) > length:
return random.sample(segment_ids, length)
else:
logger.debug("Reset length: %d" % len(segment_ids))
return segment_ids
# def get_rand_ids_each_segment(self, length):
# res = []
# status, stats = self._milvus.get_collection_stats(self._collection_name)
# self.check_status(status)
# segments = stats["partitions"][0]["segments"]
# segments_num = len(segments)
# # random choice from each segment
# for segment in segments:
# status, segment_ids = self._milvus.list_id_in_segment(self._collection_name, segment["name"])
# self.check_status(status)
# res.extend(segment_ids[:length])
# return segments_num, res
# def get_rand_entities(self, length):
# ids = self.get_rand_ids(length)
# status, get_res = self._milvus.get_entity_by_id(self._collection_name, ids)
# self.check_status(status)
# return ids, get_res
def get(self):
get_ids = random.randint(1, 1000000)
self._milvus.get_entity_by_id(self._collection_name, [get_ids])
@time_wrapper
def get_entities(self, get_ids):
get_res = self._milvus.get_entity_by_id(self._collection_name, get_ids)
return get_res
@time_wrapper
def delete(self, ids, collection_name=None):
tmp_collection_name = self._collection_name if collection_name is None else collection_name
self._milvus.delete_entity_by_id(tmp_collection_name, ids)
def delete_rand(self):
delete_id_length = random.randint(1, 100)
count_before = self.count()
logger.debug("%s: length to delete: %d" %
(self._collection_name, delete_id_length))
delete_ids = self.get_rand_ids(delete_id_length)
self.delete(delete_ids)
self.flush()
logger.info("%s: count after delete: %d" %
(self._collection_name, self.count()))
get_res = self._milvus.get_entity_by_id(self._collection_name,
delete_ids)
for item in get_res:
assert not item
# if count_before - len(delete_ids) < self.count():
# logger.error(delete_ids)
# raise Exception("Error occured")
@time_wrapper
def flush(self, _async=False, collection_name=None):
tmp_collection_name = self._collection_name if collection_name is None else collection_name
self._milvus.flush([tmp_collection_name], _async=_async)
@time_wrapper
def compact(self, collection_name=None):
tmp_collection_name = self._collection_name if collection_name is None else collection_name
status = self._milvus.compact(tmp_collection_name)
self.check_status(status)
@time_wrapper
def create_index(self,
field_name,
index_type,
metric_type,
_async=False,
index_param=None):
index_type = INDEX_MAP[index_type]
metric_type = utils.metric_type_trans(metric_type)
logger.info(
"Building index start, collection_name: %s, index_type: %s, metric_type: %s"
% (self._collection_name, index_type, metric_type))
if index_param:
logger.info(index_param)
index_params = {
"index_type": index_type,
"metric_type": metric_type,
"params": index_param
}
self._milvus.create_index(self._collection_name,
field_name,
index_params,
_async=_async)
# TODO: need to check
def describe_index(self, field_name):
# stats = self.get_stats()
info = self._milvus.describe_index(self._collection_name, field_name)
index_info = {"index_type": "flat", "index_param": None}
for field in info["fields"]:
for index in field['indexes']:
if not index or "index_type" not in index:
continue
else:
for k, v in INDEX_MAP.items():
if index['index_type'] == v:
index_info['index_type'] = k
index_info['index_param'] = index['params']
return index_info
return index_info
def drop_index(self, field_name):
logger.info("Drop index: %s" % self._collection_name)
return self._milvus.drop_index(self._collection_name, field_name)
@time_wrapper
def query(self, vector_query, filter_query=None, collection_name=None):
tmp_collection_name = self._collection_name if collection_name is None else collection_name
must_params = [vector_query]
if filter_query:
must_params.extend(filter_query)
query = {"bool": {"must": must_params}}
result = self._milvus.search(tmp_collection_name, query)
return result
@time_wrapper
def load_and_query(self,
vector_query,
filter_query=None,
collection_name=None):
tmp_collection_name = self._collection_name if collection_name is None else collection_name
must_params = [vector_query]
if filter_query:
must_params.extend(filter_query)
query = {"bool": {"must": must_params}}
self.load_collection(tmp_collection_name)
result = self._milvus.search(tmp_collection_name, query)
return result
def get_ids(self, result):
idss = result._entities.ids
ids = []
len_idss = len(idss)
len_r = len(result)
top_k = len_idss // len_r
for offset in range(0, len_idss, top_k):
ids.append(idss[offset:min(offset + top_k, len_idss)])
return ids
def query_rand(self, nq_max=100):
# for ivf search
dimension = 128
top_k = random.randint(1, 100)
nq = random.randint(1, nq_max)
nprobe = random.randint(1, 100)
search_param = {"nprobe": nprobe}
query_vectors = [[random.random() for _ in range(dimension)]
for _ in range(nq)]
metric_type = random.choice(["l2", "ip"])
logger.info("%s, Search nq: %d, top_k: %d, nprobe: %d" %
(self._collection_name, nq, top_k, nprobe))
vec_field_name = utils.get_default_field_name()
vector_query = {
"vector": {
vec_field_name: {
"topk": top_k,
"query": query_vectors,
"metric_type": utils.metric_type_trans(metric_type),
"params": search_param
}
}
}
self.query(vector_query)
def load_query_rand(self, nq_max=100):
# for ivf search
dimension = 128
top_k = random.randint(1, 100)
nq = random.randint(1, nq_max)
nprobe = random.randint(1, 100)
search_param = {"nprobe": nprobe}
query_vectors = [[random.random() for _ in range(dimension)]
for _ in range(nq)]
metric_type = random.choice(["l2", "ip"])
logger.info("%s, Search nq: %d, top_k: %d, nprobe: %d" %
(self._collection_name, nq, top_k, nprobe))
vec_field_name = utils.get_default_field_name()
vector_query = {
"vector": {
vec_field_name: {
"topk": top_k,
"query": query_vectors,
"metric_type": utils.metric_type_trans(metric_type),
"params": search_param
}
}
}
self.load_and_query(vector_query)
# TODO: need to check
def count(self, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
row_count = self._milvus.get_collection_stats(
collection_name)["row_count"]
logger.debug("Row count: %d in collection: <%s>" %
(row_count, collection_name))
return row_count
def drop(self, timeout=120, collection_name=None):
timeout = int(timeout)
if collection_name is None:
collection_name = self._collection_name
logger.info("Start delete collection: %s" % collection_name)
self._milvus.drop_collection(collection_name)
i = 0
while i < timeout:
try:
row_count = self.count(collection_name=collection_name)
if row_count:
time.sleep(1)
i = i + 1
continue
else:
break
except Exception as e:
logger.debug(str(e))
break
if i >= timeout:
logger.error("Delete collection timeout")
def get_stats(self):
return self._milvus.get_collection_stats(self._collection_name)
def get_info(self, collection_name=None):
# pdb.set_trace()
if collection_name is None:
collection_name = self._collection_name
return self._milvus.get_collection_info(collection_name)
def show_collections(self):
return self._milvus.list_collections()
def exists_collection(self, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
res = self._milvus.has_collection(collection_name)
return res
def clean_db(self):
collection_names = self.show_collections()
for name in collection_names:
self.drop(collection_name=name)
@time_wrapper
def load_collection(self, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
return self._milvus.load_collection(collection_name, timeout=3000)
@time_wrapper
def release_collection(self, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
return self._milvus.release_collection(collection_name, timeout=3000)
@time_wrapper
def load_partitions(self, tag_names, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
return self._milvus.load_partitions(collection_name,
tag_names,
timeout=3000)
@time_wrapper
def release_partitions(self, tag_names, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
return self._milvus.release_partitions(collection_name,
tag_names,
timeout=3000)
milvus = Milvus(host='localhost', port='19530')
#创建集合
param = {
'collection_name': 'test01',
'dimension': 256,
'index_file_size': 1024,
'metric_type': MetricType.L2
}
#milvus.create_collection(param)
#删除集合
milvus.drop_collection(collection_name="test01")
print("collection:", milvus.list_collections())
milvus.create_collection(param)
#print(milvus.list_partitions("test01"))
#创建分区
milvus.create_partition('test01', 'tag01')
#print(milvus.list_partitions("test01"))
#删除分区
#milvus.drop_partition('test01','tag01')
print("partition:", milvus.list_partitions("test01"))
import time
import random
import numpy as np
#vectors = [[random.random() for _ in range(256)] for _ in range(3)]
#print(np.shape(np.array(vectors)))
#vector_ids = [1,2,3]
vectors = []
vector_ids = []
id_query = {}
with open("embed.txt", "r") as f:
"fields": [
{"name": "duration", "type": DataType.INT32},
{"name": "release_year", "type": DataType.INT32},
{"name": "embedding", "type": DataType.BINARY_VECTOR, "params": {"dim": 128}},
],
"segment_row_limit": 100000,
"auto_id": False
}
# ------
# Basic create collection:
# After create collection `demo_bin_films`, we create a partition tagged "American", it means the films we
# will be inserted are from American.
# ------
client.create_collection(collection_name, collection_param)
client.create_partition(collection_name, "American")
# ------
# Basic insert entities:
# PyMilvus receive `bytes` object as a vector, so we need to generate bytes from list.
#
# Here we provide two function random binary vectors, which parameters dim refer to dimension of vectors.
# ------
def random_bin_vector(dim):
"""
This function uses numpy to specify data structure, then convert to bytes
"""
# uint8 values range is [0, 256), so we specify the high range is 256.
xb = np.random.randint(256, size=[1, (dim // 8)], dtype="uint8")
return bytes(xb[0])
def main():
# Connect to Milvus server
# You may need to change _HOST and _PORT accordingly
param = {'host': _HOST, 'port': _PORT}
# You can create a instance specified server addr and
# invoke rpc method directly
client = Milvus(**param)
# Create collection demo_collection if it dosen't exist.
collection_name = 'demo_partition_collection'
partition_tag = "random"
status, ok = client.has_collection(collection_name)
# if collection exists, then drop it
if status.OK() and ok:
client.drop_collection(collection_name)
param = {
'collection_name': collection_name,
'dimension': _DIM,
'index_file_size': _INDEX_FILE_SIZE, # optional
'metric_type': MetricType.L2 # optional
}
client.create_collection(param)
# Show collections in Milvus server
_, collections = client.show_collections()
# Describe collection
_, collection = client.describe_collection(collection_name)
print(collection)
# create partition
client.create_partition(collection_name, partition_tag=partition_tag)
# display partitions
_, partitions = client.show_partitions(collection_name)
# 10000 vectors with 16 dimension
# element per dimension is float32 type
# vectors should be a 2-D array
vectors = [[random.random() for _ in range(_DIM)] for _ in range(10000)]
# You can also use numpy to generate random vectors:
# `vectors = np.random.rand(10000, 16).astype(np.float32).tolist()`
# Insert vectors into partition of collection, return status and vectors id list
status, ids = client.insert(collection_name=collection_name, records=vectors, partition_tag=partition_tag)
# Wait for 6 seconds, until Milvus server persist vector data.
time.sleep(6)
# Get demo_collection row count
status, num = client.count_collection(collection_name)
# create index of vectors, search more rapidly
index_param = {
'nlist': 2048
}
# Create ivflat index in demo_collection
# You can search vectors without creating index. however, Creating index help to
# search faster
status = client.create_index(collection_name, IndexType.IVF_FLAT, index_param)
# describe index, get information of index
status, index = client.describe_index(collection_name)
print(index)
# Use the top 10 vectors for similarity search
query_vectors = vectors[0:10]
# execute vector similarity search, search range in partition `partition1`
search_param = {
"nprobe": 10
}
param = {
'collection_name': collection_name,
'query_records': query_vectors,
'top_k': 1,
'partition_tags': ["random"],
'params': search_param
}
status, results = client.search(**param)
if status.OK():
# indicate search result
# also use by:
# `results.distance_array[0][0] == 0.0 or results.id_array[0][0] == ids[0]`
if results[0][0].distance == 0.0 or results[0][0].id == ids[0]:
print('Query result is correct')
else:
print('Query result isn\'t correct')
# print results
print(results)
# Delete partition. You can also invoke `drop_collection()`, so that all of partitions belongs to
# designated collections will be deleted.
# status = client.drop_partition(collection_name, partition_tag)
# Delete collection. All of partitions of this collection will be dropped.
status = client.drop_collection(collection_name)
class Indexer:
'''
索引器。
'''
def __init__(self, name, host='127.0.0.1', port='19531'):
'''
初始化。
'''
self.client = Milvus(host=host, port=port)
self.collection = name
def init(self, lenient=False):
'''
创建集合。
'''
if lenient:
status, result = self.client.has_collection(
collection_name=self.collection)
if status.code != 0:
raise ExertMilvusException(status)
if result:
return
status = self.client.create_collection({
'collection_name': self.collection,
'dimension': 512,
'index_file_size': 1024,
'metric_type': MetricType.L2
})
if status.code != 0 and not (lenient and status.code == 9):
raise ExertMilvusException(status)
# 创建索引。
status = self.client.create_index(collection_name=self.collection,
index_type=IndexType.IVF_FLAT,
params={'nlist': 16384})
if status.code != 0:
raise ExertMilvusException(status)
return status
def drop(self):
'''
删除集合。
'''
status = self.client.drop_collection(collection_name=self.collection)
if status.code != 0:
raise ExertMilvusException(status)
def flush(self):
'''
写入到硬盘。
'''
status = self.client.flush([self.collection])
if status.code != 0:
raise ExertMilvusException(status)
def compact(self):
'''
压缩集合。
'''
status = self.client.compact(collection_name=self.collection)
if status.code != 0:
raise ExertMilvusException(status)
def close(self):
'''
关闭链接。
'''
self.client.close()
def new_tag(self, tag):
'''
建分块标签。
'''
status = self.client.create_partition(collection_name=self.collection,
partition_tag=tag)
if status.code != 0:
raise ExertMilvusException(status)
def list_tag(self):
'''
列举分块标签。
'''
status, result = self.client.list_partitions(
collection_name=self.collection)
if status.code != 0:
raise ExertMilvusException(status)
return result
def drop_tag(self, tag):
'''
删除分块标签。
'''
status = self.client.drop_partition(collection_name=self.collection,
partition_tag=tag)
if status.code != 0:
raise ExertMilvusException(status)
def index(self, vectors, tag=None, ids=None):
'''
添加索引
'''
params = {}
if tag != None:
params['tag'] = tag
if ids != None:
params['ids'] = ids
status, result = self.client.insert(collection_name=self.collection,
records=vectors,
**params)
if status.code != 0:
raise ExertMilvusException(status)
return result
def listing(self, ids):
'''
列举信息。
'''
status, result = self.client.get_entity_by_id(
collection_name=self.collection, ids=ids)
if status.code != 0:
raise ExertMilvusException(status)
return result
def counting(self):
'''
计算索引数。
'''
status, result = self.client.count_entities(
collection_name=self.collection)
if status.code != 0:
raise ExertMilvusException(status)
return result
def unindex(self, ids):
'''
去掉索引。
'''
status = self.client.delete_entity_by_id(
collection_name=self.collection, id_array=ids)
if status.code != 0:
raise ExertMilvusException(status)
def search(self, vectors, top_count=100, tags=None):
'''
搜索。
'''
params = {'params': {'nprobe': 16}}
if tags != None:
params['partition_tags'] = tags
status, results = self.client.search(collection_name=self.collection,
query_records=vectors,
top_k=top_count,
**params)
if status.code != 0:
raise ExertMilvusException(status)
return results
class MilvusClient(object):
def __init__(self, collection_name=None, host=None, port=None, timeout=60):
"""
Milvus client wrapper for python-sdk.
Default timeout set 60s
"""
self._collection_name = collection_name
try:
start_time = time.time()
if not host:
host = SERVER_HOST_DEFAULT
if not port:
port = SERVER_PORT_DEFAULT
logger.debug(host)
logger.debug(port)
# retry connect for remote server
i = 0
while time.time() < start_time + timeout:
try:
self._milvus = Milvus(host=host, port=port, try_connect=False, pre_ping=False)
if self._milvus.server_status():
logger.debug("Try connect times: %d, %s" % (i, round(time.time() - start_time, 2)))
break
except Exception as e:
logger.debug("Milvus connect failed: %d times" % i)
i = i + 1
if time.time() > start_time + timeout:
raise Exception("Server connect timeout")
except Exception as e:
raise e
self._metric_type = None
if self._collection_name and self.exists_collection():
self._metric_type = metric_type_to_str(self.describe()[1].metric_type)
self._dimension = self.describe()[1].dimension
def __str__(self):
return 'Milvus collection %s' % self._collection_name
def set_collection(self, name):
self._collection_name = name
def check_status(self, status):
if not status.OK():
logger.error(self._collection_name)
logger.error(status.message)
logger.error(self._milvus.server_status())
logger.error(self.count())
raise Exception("Status not ok")
def check_result_ids(self, result):
for index, item in enumerate(result):
if item[0].distance >= epsilon:
logger.error(index)
logger.error(item[0].distance)
raise Exception("Distance wrong")
def create_collection(self, collection_name, dimension, index_file_size, metric_type):
if not self._collection_name:
self._collection_name = collection_name
if metric_type not in METRIC_MAP.keys():
raise Exception("Not supported metric_type: %s" % metric_type)
metric_type = METRIC_MAP[metric_type]
create_param = {'collection_name': collection_name,
'dimension': dimension,
'index_file_size': index_file_size,
"metric_type": metric_type}
status = self._milvus.create_collection(create_param)
self.check_status(status)
def create_partition(self, tag_name):
status = self._milvus.create_partition(self._collection_name, tag_name)
self.check_status(status)
def drop_partition(self, tag_name):
status = self._milvus.drop_partition(self._collection_name, tag_name)
self.check_status(status)
def list_partitions(self):
status, tags = self._milvus.list_partitions(self._collection_name)
self.check_status(status)
return tags
@time_wrapper
def insert(self, X, ids=None, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
status, result = self._milvus.insert(collection_name, X, ids)
self.check_status(status)
return status, result
def insert_rand(self):
insert_xb = random.randint(1, 100)
X = [[random.random() for _ in range(self._dimension)] for _ in range(insert_xb)]
X = utils.normalize(self._metric_type, X)
count_before = self.count()
status, _ = self.insert(X)
self.check_status(status)
self.flush()
if count_before + insert_xb != self.count():
raise Exception("Assert failed after inserting")
def get_rand_ids(self, length):
while True:
status, stats = self._milvus.get_collection_stats(self._collection_name)
self.check_status(status)
segments = stats["partitions"][0]["segments"]
# random choice one segment
segment = random.choice(segments)
status, segment_ids = self._milvus.list_id_in_segment(self._collection_name, segment["name"])
if not status.OK():
logger.error(status.message)
continue
if len(segment_ids):
break
if length >= len(segment_ids):
logger.debug("Reset length: %d" % len(segment_ids))
return segment_ids
return random.sample(segment_ids, length)
def get_rand_ids_each_segment(self, length):
res = []
status, stats = self._milvus.get_collection_stats(self._collection_name)
self.check_status(status)
segments = stats["partitions"][0]["segments"]
segments_num = len(segments)
# random choice from each segment
for segment in segments:
status, segment_ids = self._milvus.list_id_in_segment(self._collection_name, segment["name"])
self.check_status(status)
res.extend(segment_ids[:length])
return segments_num, res
def get_rand_entities(self, length):
ids = self.get_rand_ids(length)
status, get_res = self._milvus.get_entity_by_id(self._collection_name, ids)
self.check_status(status)
return ids, get_res
@time_wrapper
def get_entities(self, get_ids):
status, get_res = self._milvus.get_entity_by_id(self._collection_name, get_ids)
self.check_status(status)
return get_res
@time_wrapper
def delete(self, ids, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
status = self._milvus.delete_entity_by_id(collection_name, ids)
self.check_status(status)
def delete_rand(self):
delete_id_length = random.randint(1, 100)
count_before = self.count()
logger.info("%s: length to delete: %d" % (self._collection_name, delete_id_length))
delete_ids = self.get_rand_ids(delete_id_length)
self.delete(delete_ids)
self.flush()
logger.info("%s: count after delete: %d" % (self._collection_name, self.count()))
status, get_res = self._milvus.get_entity_by_id(self._collection_name, delete_ids)
self.check_status(status)
for item in get_res:
if item:
raise Exception("Assert failed after delete")
if count_before - len(delete_ids) != self.count():
raise Exception("Assert failed after delete")
@time_wrapper
def flush(self, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
status = self._milvus.flush([collection_name])
self.check_status(status)
@time_wrapper
def compact(self, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
status = self._milvus.compact(collection_name)
self.check_status(status)
@time_wrapper
def create_index(self, index_type, index_param=None):
index_type = INDEX_MAP[index_type]
logger.info("Building index start, collection_name: %s, index_type: %s" % (self._collection_name, index_type))
if index_param:
logger.info(index_param)
status = self._milvus.create_index(self._collection_name, index_type, index_param)
self.check_status(status)
def describe_index(self):
status, result = self._milvus.get_index_info(self._collection_name)
self.check_status(status)
index_type = None
for k, v in INDEX_MAP.items():
if result._index_type == v:
index_type = k
break
return {"index_type": index_type, "index_param": result._params}
def drop_index(self):
logger.info("Drop index: %s" % self._collection_name)
return self._milvus.drop_index(self._collection_name)
def query(self, X, top_k, search_param=None, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
status, result = self._milvus.search(collection_name, top_k, query_records=X, params=search_param)
self.check_status(status)
return result
def query_rand(self):
top_k = random.randint(1, 100)
nq = random.randint(1, 100)
nprobe = random.randint(1, 100)
search_param = {"nprobe": nprobe}
_, X = self.get_rand_entities(nq)
logger.info("%s, Search nq: %d, top_k: %d, nprobe: %d" % (self._collection_name, nq, top_k, nprobe))
status, _ = self._milvus.search(self._collection_name, top_k, query_records=X, params=search_param)
self.check_status(status)
# for i, item in enumerate(search_res):
# if item[0].id != ids[i]:
# logger.warning("The index of search result: %d" % i)
# raise Exception("Query failed")
# @time_wrapper
# def query_ids(self, top_k, ids, search_param=None):
# status, result = self._milvus.search_by_id(self._collection_name, ids, top_k, params=search_param)
# self.check_result_ids(result)
# return result
def count(self, name=None):
if name is None:
name = self._collection_name
logger.debug(self._milvus.count_entities(name))
row_count = self._milvus.count_entities(name)[1]
if not row_count:
row_count = 0
logger.debug("Row count: %d in collection: <%s>" % (row_count, name))
return row_count
def drop(self, timeout=120, name=None):
timeout = int(timeout)
if name is None:
name = self._collection_name
logger.info("Start delete collection: %s" % name)
status = self._milvus.drop_collection(name)
self.check_status(status)
i = 0
while i < timeout:
if self.count(name=name):
time.sleep(1)
i = i + 1
continue
else:
break
if i >= timeout:
logger.error("Delete collection timeout")
def describe(self):
# logger.info(self._milvus.get_collection_info(self._collection_name))
return self._milvus.get_collection_info(self._collection_name)
def show_collections(self):
return self._milvus.list_collections()
def exists_collection(self, collection_name=None):
if collection_name is None:
collection_name = self._collection_name
_, res = self._milvus.has_collection(collection_name)
# self.check_status(status)
return res
def clean_db(self):
collection_names = self.show_collections()[1]
for name in collection_names:
logger.debug(name)
self.drop(name=name)
@time_wrapper
def preload_collection(self):
status = self._milvus.load_collection(self._collection_name, timeout=3000)
self.check_status(status)
return status
def get_server_version(self):
_, res = self._milvus.server_version()
return res
def get_server_mode(self):
return self.cmd("mode")
def get_server_commit(self):
return self.cmd("build_commit_id")
def get_server_config(self):
return json.loads(self.cmd("get_config *"))
def get_mem_info(self):
result = json.loads(self.cmd("get_system_info"))
result_human = {
# unit: Gb
"memory_used": round(int(result["memory_used"]) / (1024*1024*1024), 2)
}
return result_human
def cmd(self, command):
status, res = self._milvus._cmd(command)
logger.info("Server command: %s, result: %s" % (command, res))
self.check_status(status)
return res
def main():
milvus = Milvus(uri=uri)
param = {
'collection_name': collection_name,
'dimension': _DIM,
'index_file_size': 32,
#'metric_type': MetricType.IP
'metric_type': MetricType.L2
}
# show collections in Milvus server
_, collections = milvus.list_collections()
# 创建 collection
milvus.create_collection(param)
# 创建 collection partion
milvus.create_partition(collection_name, partition_tag)
print(f'collections in Milvus: {collections}')
# Describe demo_collection
_, collection = milvus.get_collection_info(collection_name)
print(f'descript demo_collection: {collection}')
# build fake vectors
vectors = [[random.random() for _ in range(_DIM)] for _ in range(10)]
vectors1 = [[random.random() for _ in range(_DIM)] for _ in range(10)]
status, id = milvus.insert(collection_name=collection_name,
records=vectors,
ids=list(range(10)),
partition_tag=partition_tag)
print(f'status: {status} | id: {id}')
if not status.OK():
print(f"insert failded: {status}")
status1, id1 = milvus.insert(collection_name=collection_name,
records=vectors1,
ids=list(range(10, 20)),
partition_tag=partition_tag)
print(f'status1: {status1} | id1: {id1}')
ids_deleted = list(range(10))
status_delete = milvus.delete_entity_by_id(collection_name=collection_name,
id_array=ids_deleted)
if status_delete.OK():
print(f'delete successful')
# Flush collection insered data to disk
milvus.flush([collection_name])
# Get demo_collection row count
status, result = milvus.count_entities(collection_name)
print(f"demo_collection row count: {result}")
# Obtain raw vectors by providing vector ids
status, result_vectors = milvus.get_entity_by_id(collection_name,
list(range(10, 20)))
# create index of vectors, search more repidly
index_param = {'nlist': 2}
# create ivflat index in demo_collection
status = milvus.create_index(collection_name, IndexType.IVF_FLAT,
index_param)
if status.OK():
print(f"create index ivf_flat succeeed")
# use the top 10 vectors for similarity search
query_vectors = vectors1[0:2]
# execute vector similariy search
search_param = {"nprobe": 16}
param = {
'collection_name': collection_name,
'query_records': query_vectors,
'top_k': 1,
'params': search_param
}
status, results = milvus.search(**param)
if status.OK():
if results[0][0].distance == 0.0:
print('query result is correct')
else:
print('not correct')
print(results)
else:
print(f'search failed: {status}')
# 清除已经存在的collection
milvus.drop_collection(collection_name=collection_name)
milvus.close()
{
"name": "embedding",
"type": DataType.FLOAT_VECTOR,
"params": {
"dim": 32
}
},
],
"segment_row_limit":
16384,
"auto_id":
False
}
client.create_collection(collection_name, collection_param)
client.create_partition(collection_name, "Movie")
print("--------get collection info--------")
collection = client.get_collection_info(collection_name)
pprint(collection)
partitions = client.list_partitions(collection_name)
print("\n----------list partitions----------")
pprint(partitions)
import codecs
with codecs.open("movie_vectors.txt", "r", encoding='utf-8',
errors='ignore') as f:
lines = f.readlines()
ids = [int(line.split(":")[0]) for line in lines]
embeddings = []
class ANN(object):
def __init__(self, host='10.119.33.90', port='19530', show_info=False):
self.client = Milvus(host, port)
if show_info:
logger.info({
"ClientVersion": self.client.client_version(),
"ServerVersion": self.client.server_version()
})
def create_collection(self,
collection_name,
collection_param,
partition_tag=None,
overwrite=True):
"""
:param collection_name:
:param collection_param:
collection_param = {
"fields": [
# Milvus doesn't support string type now, but we are considering supporting it soon.
# {"name": "title", "type": DataType.STRING},
{"name": "category_", "type": DataType.INT32},
{"name": "vector", "type": DataType.FLOAT_VECTOR, "params": {"dim": 768}},
],
"segment_row_limit": 4096,
"auto_id": False
}
:param overwrite:
:return:
"""
if self.client.has_collection(collection_name) and overwrite:
self.client.drop_collection(collection_name)
self.client.flush()
time.sleep(5)
self.client.create_collection(collection_name, collection_param)
elif self.client.has_collection(collection_name):
print(f"{collection_name} already exist !!!")
else:
self.client.create_collection(collection_name, collection_param)
if partition_tag is not None:
self.client.create_partition(collection_name,
partition_tag=partition_tag)
def create_index(self,
collection_name,
field_name,
index_type='IVF_FLAT',
metric_type='IP',
index_params=None):
"""
MetricType:
INVALID = 0
L2 = 1
IP = 2
# Only supported for byte vectors
HAMMING = 3
JACCARD = 4
TANIMOTO = 5
#
SUBSTRUCTURE = 6
SUPERSTRUCTURE = 7
IndexType:
INVALID = 0
FLAT = 1
IVFLAT = 2
IVF_SQ8 = 3
RNSG = 4
IVF_SQ8H = 5
IVF_PQ = 6
HNSW = 11
ANNOY = 12
# alternative name
IVF_FLAT = IVFLAT
IVF_SQ8_H = IVF_SQ8H
class DataType(IntEnum):
NULL = 0
INT8 = 1
INT16 = 2
INT32 = 3
INT64 = 4
STRING = 20
BOOL = 30
FLOAT = 40
DOUBLE = 41
VECTOR = 100
UNKNOWN = 9999
class RangeType(IntEnum):
LT = 0 # less than
LTE = 1 # less than or equal
EQ = 2 # equal
GT = 3 # greater than
GTE = 4 # greater than or equal
NE = 5 # not equal
:return:
"""
if index_params is None:
index_params = {'nlist': 1024}
params = {
'index_type': index_type,
# 'index_file_size': 1024,
'params': index_params,
'metric_type': metric_type,
}
self.client.create_index(collection_name, field_name,
params) # field_name='embedding'
def batch_insert(self, collection_name, entities, batch_size=100000):
# 分区
n = len(entities[0]['values'])
num_part = n // batch_size + 1
ids = []
values_list = [_['values'] for _ in entities]
for i in range(num_part):
for e, values in zip(entities, values_list):
e['values'] = values[i * batch_size:(i + 1) * batch_size]
ids += self.client.insert(collection_name, entities)
self.client.flush()
return ids
def search(self): # todo: 获取相同的信息
pass
def drop_collection(self, collection_name):
if self.client.has_collection(collection_name):
self.client.drop_collection(collection_name)
def drop_partition(self, collection_name, partition_tag):
if self.client.has_partition(collection_name, partition_tag):
self.client.drop_partition(collection_name,
partition_tag,
timeout=30)
