def validate_insert(_collection_name):
milvus = Milvus(**server_config)
milvus.flush([_collection_name])
status, count = milvus.count_entities(_collection_name)
assert count == 10 * 10000, "Insert validate fail. Vectors num is not matched."
# drop collcetion
print("Drop collection ...")
milvus.drop_collection(_collection_name)
milvus.close()
def milvus_test(usr_features, IS_INFER, mov_features=None, ids=None):
_HOST = '127.0.0.1'
_PORT = '19530' # default value
table_name = 'recommender_demo'
milvus = Milvus(_HOST, _PORT)
if IS_INFER:
status = milvus.drop_collection(table_name)
time.sleep(3)
status, ok = milvus.has_collection(table_name)
if not ok:
if mov_features is None:
print("Insert vectors is none!")
sys.exit(1)
param = {
'collection_name': table_name,
'dimension': 200,
'index_file_size': 1024, # optional
'metric_type': MetricType.IP # optional
}
print(milvus.create_collection(param))
insert_vectors = normaliz_data(mov_features)
status, ids = milvus.insert(collection_name=table_name,
records=insert_vectors,
ids=ids)
time.sleep(1)
status, result = milvus.count_entities(table_name)
print("rows in table recommender_demo:", result)
search_vectors = normaliz_data(usr_features)
param = {
'collection_name': table_name,
'query_records': search_vectors,
'top_k': 5,
'params': {
'nprobe': 16
}
}
time1 = time.time()
status, results = milvus.search(**param)
time2 = time.time()
print("Top\t", "Ids\t", "Title\t", "Score")
for i, re in enumerate(results[0]):
title = paddle.dataset.movielens.movie_info()[int(re.id)].title
print(i, "\t", re.id, "\t", title, "\t", float(re.distance) * 5)
def milvus_test(usr_features, mov_features, ids):
_HOST = '127.0.0.1'
_PORT = '19530' # default value
milvus = Milvus(_HOST, _PORT)
table_name = 'paddle_demo1'
status, ok = milvus.has_collection(table_name)
if not ok:
param = {
'collection_name': table_name,
'dimension': 200,
'index_file_size': 1024, # optional
'metric_type': MetricType.IP # optional
}
milvus.create_collection(param)
insert_vectors = normaliz_data([usr_features.tolist()])
status, ids = milvus.insert(collection_name=table_name,
records=insert_vectors,
ids=ids)
time.sleep(1)
status, result = milvus.count_entities(table_name)
print("rows in table paddle_demo1:", result)
# status, table = milvus.count_entities(table_name)
search_vectors = normaliz_data([mov_features.tolist()])
param = {
'collection_name': table_name,
'query_records': search_vectors,
'top_k': 1,
'params': {
'nprobe': 16
}
}
status, results = milvus.search(**param)
print("Searched ids:", results[0][0].id)
print("Score:", float(results[0][0].distance) * 5)
status = milvus.drop_collection(table_name)
{
"name": "embedding",
"values": embeddings,
"type": DataType.FLOAT_VECTOR
},
]
# ------
# Basic insert:
# After preparing the data, we are going to insert them into our collection.
# The number of films inserted should be 8657.
# ------
ids = client.insert(collection_name, hybrid_entities, ids)
client.flush([collection_name])
after_flush_counts = client.count_entities(collection_name)
print(" > There are {} films in collection `{}` after flush".format(
after_flush_counts, collection_name))
# ------
# Basic create index:
# Now that we have inserted all the films into Milvus, we are going to build index with these data.
#
# While building index, we have to indicate which `field` to build index for, the `index_type`,
# `metric_type` and params for the specific index type. In our case, we want to build a `IVF_FLAT`
# index, so the specific params are "nlist". See pymilvus documentation
# (https://milvus-io.github.io/milvus-sdk-python/pythondoc/v0.3.0/index.html) for `index_type` we
# support and the params accordingly.
#
# If there are already index for a collection and you call `create_index` with different params, the
# older index will be replaced by new one.
class MilvusHelper:
def __init__(self):
try:
self.client = Milvus(host=MILVUS_HOST, port=MILVUS_PORT)
LOGGER.debug(
"Successfully connect to Milvus with IP:{} and PORT:{}".format(
MILVUS_HOST, MILVUS_PORT))
except Exception as e:
LOGGER.error("Failed to connect Milvus: {}".format(e))
sys.exit(1)
# Return if Milvus has the collection
def has_collection(self, collection_name):
try:
status = self.client.has_collection(collection_name)[1]
return status
except Exception as e:
LOGGER.error("Failed to load data to Milvus: {}".format(e))
sys.exit(1)
# Create milvus collection if not exists
def create_colllection(self, collection_name):
try:
if not self.has_collection(collection_name):
collection_param = {
'collection_name': collection_name,
'dimension': VECTOR_DIMENSION,
'index_file_size': INDEX_FILE_SIZE,
'metric_type': METRIC_TYPE
}
status = self.client.create_collection(collection_param)
if status.code != 0:
raise Exception(status.message)
LOGGER.debug(
"Create Milvus collection: {}".format(collection_name))
except Exception as e:
LOGGER.error("Failed to load data to Milvus: {}".format(e))
sys.exit(1)
# Batch insert vectors to milvus collection
def insert(self, collection_name, vectors):
try:
self.create_colllection(collection_name)
status, ids = self.client.insert(collection_name=collection_name,
records=vectors)
if not status.code:
LOGGER.debug(
"Insert vectors to Milvus in collection: {} with {} rows".
format(collection_name, len(vectors)))
return ids
else:
raise Exception(status.message)
except Exception as e:
LOGGER.error("Failed to load data to Milvus: {}".format(e))
sys.exit(1)
# Create IVF_FLAT index on milvus collection
def create_index(self, collection_name):
try:
index_param = {'nlist': 16384}
status = self.client.create_index(collection_name,
IndexType.IVF_FLAT, index_param)
if not status.code:
LOGGER.debug(
"Successfully create index in collection:{} with param:{}".
format(collection_name, index_param))
return status
else:
raise Exception(status.message)
except Exception as e:
LOGGER.error("Failed to create index: {}".format(e))
sys.exit(1)
# Delete Milvus collection
def delete_collection(self, collection_name):
try:
status = self.client.drop_collection(
collection_name=collection_name)
if not status.code:
LOGGER.debug(
"Successfully drop collection: {}".format(collection_name))
return status
else:
raise Exception(status.message)
except Exception as e:
LOGGER.error("Failed to drop collection: {}".format(e))
sys.exit(1)
# Search vector in milvus collection
def search_vectors(self, collection_name, vectors, top_k):
try:
search_param = {'nprobe': 16}
status, result = self.client.search(
collection_name=collection_name,
query_records=vectors,
top_k=top_k,
params=search_param)
if not status.code:
LOGGER.debug("Successfully search in collection: {}".format(
collection_name))
return result
else:
raise Exception(status.message)
except Exception as e:
LOGGER.error("Failed to search vectors in Milvus: {}".format(e))
sys.exit(1)
# Get the number of milvus collection
def count(self, collection_name):
try:
status, num = self.client.count_entities(
collection_name=collection_name)
if not status.code:
LOGGER.debug(
"Successfully get the num:{} of the collection:{}".format(
num, collection_name))
return num
else:
raise Exception(status.message)
except Exception as e:
LOGGER.error("Failed to count vectors in Milvus: {}".format(e))
sys.exit(1)
def main():
# Specify server addr when create milvus client instance
milvus = Milvus(_HOST, _PORT)
# Create collection demo_collection if it dosen't exist.
collection_name = 'example_async_collection_'
status, ok = milvus.has_collection(collection_name)
if not ok:
param = {
'collection_name': collection_name,
'dimension': _DIM,
'index_file_size': 128, # optional
'metric_type': MetricType.L2 # optional
}
status = milvus.create_collection(param)
if not status.OK():
print("Create collection failed: {}".format(status.message),
file=sys.stderr)
print("exiting ...", file=sys.stderr)
sys.exit(1)
# Show collections in Milvus server
_, collections = milvus.list_collections()
# Describe demo_collection
_, collection = milvus.get_collection_info(collection_name)
print(collection)
# 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(100000)]
# You can also use numpy to generate random vectors:
# `vectors = np.random.rand(10000, 16).astype(np.float32)`
def _insert_callback(status, ids):
if status.OK():
print("Insert successfully")
else:
print("Insert failed.", status.message)
# Insert vectors into demo_collection, adding callback function
insert_future = milvus.insert(collection_name=collection_name,
records=vectors,
_async=True,
_callback=_insert_callback)
# Or invoke result() to get results:
# insert_future = milvus.insert(collection_name=collection_name, records=vectors, _async=True)
# status, ids = insert_future.result()
insert_future.done()
# Flush collection inserted data to disk.
def _flush_callback(status):
if status.OK():
print("Flush successfully")
else:
print("Flush failed.", status.message)
flush_future = milvus.flush([collection_name],
_async=True,
_callback=_flush_callback)
# Or invoke result() to get results:
# flush_future = milvus.flush([collection_name], _async=True)
# status = flush_future.result()
flush_future.done()
def _compact_callback(status):
if status.OK():
print("Compact successfully")
else:
print("Compact failed.", status.message)
compact_furure = milvus.compact(collection_name,
_async=True,
_cakkback=_compact_callback)
# Or invoke result() to get results:
# compact_future = milvus.compact(collection_name, _async=True)
# status = compact_future.result()
compact_furure.done()
# Get demo_collection row count
status, result = milvus.count_entities(collection_name)
# present collection info
_, info = milvus.get_collection_stats(collection_name)
print(info)
# create index of vectors, search more rapidly
index_param = {'nlist': 2048}
def _index_callback(status):
if status.OK():
print("Create index successfully")
else:
print("Create index failed.", status.message)
# Create ivflat index in demo_collection
# You can search vectors without creating index. however, Creating index help to
# search faster
print("Creating index: {}".format(index_param))
index_future = milvus.create_index(collection_name,
IndexType.IVF_FLAT,
index_param,
_async=True,
_callback=_index_callback)
# Or invoke result() to get results:
# index_future = milvus.create_index(collection_name, IndexType.IVF_FLAT, index_param, _async=True)
# status = index_future.result()
index_future.done()
# describe index, get information of index
status, index = milvus.get_index_info(collection_name)
print(index)
# Use the top 10 vectors for similarity search
query_vectors = vectors[0:10]
# execute vector similarity search
search_param = {"nprobe": 16}
print("Searching ... ")
def _search_callback(status, results):
# if status.OK():
# print("Search successfully")
# else:
# print("Search failed.", status.message)
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)
else:
print("Search failed. ", status)
param = {
'collection_name': collection_name,
'query_records': query_vectors,
'top_k': 1,
'params': search_param,
"_async": True,
"_callback": _search_callback
}
search_future = milvus.search(**param)
# Or invoke result() to get results:
#
# param = {
# 'collection_name': collection_name,
# 'query_records': query_vectors,
# 'top_k': 1,
# 'params': search_param,
# "_async": True,
# }
# search_future = milvus.search(param)
# status, results = index_future.result()
search_future.done()
# Delete demo_collection
status = milvus.drop_collection(collection_name)
def main():
# Specify server addr when create milvus client instance
# milvus client instance maintain a connection pool, param
# `pool_size` specify the max connection num.
# 获取服务端的连接
milvus = Milvus(_HOST, _PORT)
# Create collection demo_collection if it dosen't exist.
# 创建collection
collection_name = 'example_collection_'
# 看是否有这个collection
status, ok = milvus.has_collection(collection_name)
# 如果没有则创建
if not ok:
param = {
'collection_name': collection_name,
'dimension': _DIM,
'index_file_size': _INDEX_FILE_SIZE, # optional
'metric_type': MetricType.L2 # optional
}
# 创建collection
milvus.create_collection(param)
# Show collections in Milvus server
# 查看所有的collection
_, collections = milvus.list_collections()
print(collections)
# Describe demo_collection
# 得到当前的collection
_, collection = milvus.get_collection_info(collection_name)
print(collection)
# 10000 vectors with 128 dimension
# element per dimension is float32 type
# vectors should be a 2-D array
# 创建10个长度为8的向量
vectors = [[random.random() for _ in range(_DIM)] for _ in range(10)]
print(vectors)
# You can also use numpy to generate random vectors:
# vectors = np.random.rand(10000, _DIM).astype(np.float32)
# Insert vectors into demo_collection, return status and vectors id list
# 把这10个向量都插入milvus
status, ids = milvus.insert(collection_name=collection_name,
records=vectors)
if not status.OK():
print("Insert failed: {}".format(status))
print(ids)
# Flush collection inserted data to disk.
# 数据落盘
milvus.flush([collection_name])
# Get demo_collection row count
# 得到当前row的数量
status, result = milvus.count_entities(collection_name)
print(status)
print(result)
# present collection statistics info
# 查看collection的统计数据
_, info = milvus.get_collection_stats(collection_name)
print(info)
# Obtain raw vectors by providing vector ids
# 得到前十个数据
status, result_vectors = milvus.get_entity_by_id(collection_name, ids[:10])
print(result_vectors)
# create index of vectors, search more rapidly
# 创建索引
index_param = {'nlist': 2048}
# Create ivflat index in demo_collection
# You can search vectors without creating index. however, Creating index help to
# search faster
# 创建ivf_flat
print("Creating index: {}".format(index_param))
status = milvus.create_index(collection_name, IndexType.IVF_FLAT,
index_param)
# describe index, get information of index
# 得到索引的信息
status, index = milvus.get_index_info(collection_name)
print(index)
# Use the top 10 vectors for similarity search
# 对前10个数据进行query
query_vectors = vectors[0:10]
# execute vector similarity search
# 索引的搜索的中心点数量
search_param = {"nprobe": 16}
print("Searching ... ")
param = {
'collection_name': collection_name,
'query_records': query_vectors,
'top_k': 1,
'params': search_param,
}
# 进行搜索
status, results = milvus.search(**param)
if status.OK():
print(results)
# 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)
else:
print("Search failed. ", status)
# Delete demo_collection
# 删除掉collection
status = milvus.drop_collection(collection_name)
def main():
milvus = Milvus(_HOST, _PORT)
# num = random.randint(1, 100000)
num = 100000
# Create collection demo_collection if it dosen't exist.
collection_name = 'example_hybrid_collections_{}'.format(num)
if milvus.has_collection(collection_name):
milvus.drop_collection(collection_name)
collection_param = {
"fields": [{
"field": "A",
"type": DataType.INT32
}, {
"field": "B",
"type": DataType.INT32
}, {
"field": "C",
"type": DataType.INT64
}, {
"field": "Vec",
"type": DataType.FLOAT_VECTOR,
"params": {
"dim": 128,
"metric_type": "L2"
}
}],
"segment_size":
100
}
milvus.create_collection(collection_name, collection_param)
milvus.compact(collection_name)
# milvus.create_partition(collection_name, "p_01", timeout=1800)
# pars = milvus.list_partitions(collection_name)
# ok = milvus.has_partition(collection_name, "p_01", timeout=1800)
# assert ok
# ok = milvus.has_partition(collection_name, "p_02")
# assert not ok
# for p in pars:
# if p == "_default":
# continue
# milvus.drop_partition(collection_name, p)
# milvus.drop_collection(collection_name)
# sys.exit(0)
A_list = [random.randint(0, 255) for _ in range(num)]
vec = [[random.random() for _ in range(128)] for _ in range(num)]
hybrid_entities = [{
"field": "A",
"values": A_list,
"type": DataType.INT32
}, {
"field": "B",
"values": A_list,
"type": DataType.INT32
}, {
"field": "C",
"values": A_list,
"type": DataType.INT64
}, {
"field": "Vec",
"values": vec,
"type": DataType.FLOAT_VECTOR,
"params": {
"dim": 128
}
}]
for slice_e in utils.entities_slice(hybrid_entities):
ids = milvus.insert(collection_name, slice_e)
milvus.flush([collection_name])
print("Flush ... ")
# time.sleep(3)
count = milvus.count_entities(collection_name)
milvus.delete_entity_by_id(collection_name, ids[:1])
milvus.flush([collection_name])
print("Get entity be id start ...... ")
entities = milvus.get_entity_by_id(collection_name, ids[:1])
et = entities.dict()
milvus.delete_entity_by_id(collection_name, ids[1:2])
milvus.flush([collection_name])
print("Create index ......")
milvus.create_index(collection_name, "Vec", {
"index_type": "IVF_FLAT",
"metric_type": "L2",
"params": {
"nlist": 100
}
})
print("Create index done.")
info = milvus.get_collection_info(collection_name)
print(info)
stats = milvus.get_collection_stats(collection_name)
print("\nstats\n")
print(stats)
query_hybrid = \
{
"bool": {
"must": [
{
"term": {
"A": [1, 2, 5]
}
},
{
"range": {
"B": {"GT": 1, "LT": 100}
}
},
{
"vector": {
"Vec": {
"topk": 10, "query": vec[: 10000], "params": {"nprobe": 10}
}
}
}
],
},
}
# print("Start searach ..", flush=True)
# results = milvus.search(collection_name, query_hybrid)
# print(results)
#
# for r in list(results):
# print("ids", r.ids)
# print("distances", r.distances)
t0 = time.time()
count = 0
results = milvus.search(collection_name, query_hybrid, fields=["B"])
for r in list(results):
# print("ids", r.ids)
# print("distances", r.distances)
for rr in r:
count += 1
# print(rr.entity.get("B"))
print("Search cost {} s".format(time.time() - t0))
# for result in results:
# for r in result:
# print(f"{r}")
# itertor entity id
# for result in results:
# for r in result:
# # get distance
# dis = r.distance
# id_ = r.id
# # obtain all field name
# fields = r.entity.fields
# for f in fields:
# # get field value by field name
# # fv = r.entity.
# fv = r.entity.value_of_field(f)
# print(fv)
milvus.drop_collection(collection_name)
print("- release_year: {}".format(current_entity.release_year))
print("- duration: {}".format(current_entity.duration))
print("- embedding: {}".format(current_entity.embedding))
# ------
# Basic delete:
# Now let's see how to delete things in Milvus.
# You can simply delete entities by their ids.
# ------
client.delete_entity_by_id(collection_name, ids=[1, 4])
client.flush() # flush is important
result = client.get_entity_by_id(collection_name, ids=[1, 4])
counts_delete = sum([1 for entity in result if entity is not None])
counts_in_collection = client.count_entities(collection_name)
print("\n----------delete id = 1, id = 4----------")
print("Get {} entities by id 1, 4".format(counts_delete))
print("There are {} entities after delete films with 1, 4".format(counts_in_collection))
# ------
# Basic delete:
# You can drop partitions we create, and drop the collection we create.
# ------
client.drop_partition(collection_name, partition_tag='American')
if collection_name in client.list_collections():
client.drop_collection(collection_name)
# ------
# Summary:
# Now we've went through all basic communications pymilvus can do with Milvus server, hope it's helpful!
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
def main():
# Specify server addr when create milvus client instance
# milvus client instance maintain a connection pool, param
# `pool_size` specify the max connection num.
milvus = Milvus(_HOST, _PORT)
# Create collection demo_collection if it dosen't exist.
collection_name = 'example_collection_'
status, ok = milvus.has_collection(collection_name)
if not ok:
param = {
'collection_name': collection_name,
'dimension': _DIM,
'index_file_size': _INDEX_FILE_SIZE, # optional
'metric_type': MetricType.L2 # optional
}
milvus.create_collection(param)
# Show collections in Milvus server
_, collections = milvus.list_collections()
# Describe demo_collection
_, collection = milvus.get_collection_info(collection_name)
print(collection)
# element per dimension is float32 type
# vectors should be a 2-D array
vectors = text2vec(index_sentences)
print(vectors)
# Insert vectors into demo_collection, return status and vectors id list
status, ids = milvus.insert(collection_name=collection_name,
records=vectors)
if not status.OK():
print("Insert failed: {}".format(status))
else:
print(ids)
#create a quick lookup table to easily access the indexed text/sentences given the ids
look_up = {}
for ID, sentences in zip(ids, index_sentences):
look_up[ID] = sentences
for k in look_up:
print(k, look_up[k])
# Flush collection inserted data to disk.
milvus.flush([collection_name])
# Get demo_collection row count
status, result = milvus.count_entities(collection_name)
# present collection statistics info
_, info = milvus.get_collection_stats(collection_name)
print(info)
# Obtain raw vectors by providing vector ids
status, result_vectors = milvus.get_entity_by_id(collection_name, ids)
# 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
print("Creating index: {}".format(index_param))
status = milvus.create_index(collection_name, IndexType.IVF_FLAT,
index_param)
# describe index, get information of index
status, index = milvus.get_index_info(collection_name)
print(index)
# Use the query sentences for similarity search
query_vectors = text2vec(query_sentences)
# execute vector similarity search
search_param = {"nprobe": 16}
print("Searching ... ")
param = {
'collection_name': collection_name,
'query_records': query_vectors,
'top_k': 1,
'params': search_param,
}
status, results = milvus.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
for res in results:
for ele in res:
print('id:{}, text:{}, distance: {}'.format(
ele.id, look_up[ele.id], ele.distance))
else:
print("Search failed. ", status)
# Delete demo_collection
status = milvus.drop_collection(collection_name)
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
"duration": 165,
"release_year": 2012,
"embedding": [random.random() for _ in range(8)]
}
]
ids = client.insert(collection_name, Batmans)
print("\n----------insert batmans----------")
print("Films are inserted and the ids are: {}".format(ids))
# ------
# Basic insert entities:
# After insert entities into collection, we need to flush collection to make sure its on disk,
# so that we are able to retrieve it.
# ------
before_flush_counts = client.count_entities(collection_name)
client.flush([collection_name])
after_flush_counts = client.count_entities(collection_name)
print("\n----------flush----------")
print("There are {} films in collection `{}` before flush".format(before_flush_counts, collection_name))
print("There are {} films in collection `{}` after flush".format(after_flush_counts, collection_name))
# ------
# Basic insert entities:
# We can get the detail of collection statistics info by `get_collection_stats`
# ------
info = client.get_collection_stats(collection_name)
print("\n----------get collection stats----------")
pprint(info)
# ------
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()
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"
# create collection
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.list_collections()
# Describe collection
_, collection = client.get_collection_info(collection_name)
print(collection)
# create partition
client.create_partition(collection_name, partition_tag=partition_tag)
# display partitions
_, partitions = client.list_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_entities(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.get_index_info(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)
# Drop 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 MyMilvus():
def __init__(self, name, host, port, collection_param):
self.host = host
self.port = port
self.client = Milvus(host, port)
self.collection_name = name
self.collection_param = collection_param
# self.collection_param = {
# "fields": [
# {"name": "release_year", "type": DataType.INT32},
# {"name": "embedding", "type": DataType.FLOAT_VECTOR, "params": {"dim": 8}},
# ],
# "segment_row_limit": 4096,
# "auto_id": False
# }
def create_collection(self):
if self.collection_name not in self.client.list_collections():
# self.client.drop_collection(self.collection_name)
self.client.create_collection(self.collection_name, self.collection_param)
# ------
# Basic create index:
# Now that we have a collection in Milvus with `segment_row_limit` 4096, we can create index or
# insert entities.
#
# We can call `create_index` BEFORE we insert any entities or AFTER. However Milvus won't actually
# start build index task if the segment row count is smaller than `segment_row_limit`. So if we want
# to make Milvus build index, we need to insert number of entities larger than `segment_row_limit`.
#
# We are going to use data in `films.csv` so you can checkout the structure. And we need to group
# data with same fields together, so here is a example of how we obtain the data in files and transfer
# them into what we need.
# ------
ids = [] # ids
titles = [] # titles
release_years = [] # release year
embeddings = [] # embeddings
films = []
with open('films.csv', 'r') as csvfile:
reader = csv.reader(csvfile)
films = [film for film in reader]
for film in films:
ids.append(int(film[0]))
titles.append(film[1])
release_years.append(int(film[2]))
embeddings.append(list(map(float, film[3][1:][:-1].split(','))))
hybrid_entities = [
{"name": "release_year", "values": release_years, "type": DataType.INT32},
{"name": "embedding", "values": embeddings, "type": DataType.FLOAT_VECTOR},
]
# ------
# Basic insert:
# After preparing the data, we are going to insert them into our collection.
# The number of films inserted should be 8657.
# ------
ids = self.client.insert(self.collection_name, hybrid_entities, ids)
self.client.flush([self.collection_name])
after_flush_counts = self.client.count_entities(self.collection_name)
print(" > There are {} films in collection `{}` after flush".format(after_flush_counts, self.collection_name))
# ------
# Basic create index:
# Now that we have inserted all the films into Milvus, we are going to build index with these data.
#
# While building index, we have to indicate which `field` to build index for, the `index_type`,
# `metric_type` and params for the specific index type. In our case, we want to build a `IVF_FLAT`
# index, so the specific params are "nlist". See pymilvus documentation
# (https://milvus-io.github.io/milvus-sdk-python/pythondoc/v0.3.0/index.html) for `index_type` we
# support and the params accordingly.
#
# If there are already index for a collection and you call `create_index` with different params, the
# older index will be replaced by new one.
# ------
self.client.create_index(self.collection_name, "embedding",
{"index_type": "IVF_FLAT", "metric_type": "L2", "params": {"nlist": 100}})
# ------
# Basic create index:
# We can get the detail of the index by `get_collection_info`.
# ------
info = self.client.get_collection_info(self.collection_name)
pprint(info)
# ------
# Basic hybrid search entities:
# If we want to use index, the specific index params need to be provided, in our case, the "params"
# should be "nprobe", if no "params" given, Milvus will complain about it and raise a exception.
# ------
# query_embedding = [random.random() for _ in range(8)]
# query_hybrid = {
# "bool": {
# "must": [
# {
# "term": {"release_year": [2002, 1995]}
# },
# {
# "vector": {
# "embedding": {"topk": 3,
# "query": [query_embedding],
# "metric_type": "L2",
# "params": {"nprobe": 8}}
# }
# }
# ]
# }
# }
# ------
# Basic hybrid search entities
# ------
# results = client.search(collection_name, query_hybrid, fields=["release_year", "embedding"])
# for entities in results:
# for topk_film in entities:
# current_entity = topk_film.entity
# print("==")
# print("- id: {}".format(topk_film.id))
# print("- title: {}".format(titles[topk_film.id]))
# print("- distance: {}".format(topk_film.distance))
#
# print("- release_year: {}".format(current_entity.release_year))
# print("- embedding: {}".format(current_entity.embedding))
# ------
# Basic delete index:
# You can drop index for a field.
# ------
self.client.drop_index(self.collection_name, "embedding")
if self.collection_name in self.client.list_collections():
self.client.drop_collection(self.collection_name)