milvus/shards/mishards/service_handler.py

737 lines
30 KiB
Python

import logging
import time
import json
import ujson
import multiprocessing
from milvus.grpc_gen import milvus_pb2, milvus_pb2_grpc, status_pb2
from milvus.client import types as Types
from milvus import MetricType
from mishards import (db, exceptions)
from mishards.grpc_utils import mark_grpc_method
from mishards.grpc_utils.grpc_args_parser import GrpcArgsParser as Parser
logger = logging.getLogger(__name__)
class ServiceHandler(milvus_pb2_grpc.MilvusServiceServicer):
MAX_NPROBE = 2048
MAX_TOPK = 2048
def __init__(self, tracer, router, max_workers=multiprocessing.cpu_count(), **kwargs):
self.collection_meta = {}
self.error_handlers = {}
self.tracer = tracer
self.router = router
self.max_workers = max_workers
def _reduce(self, source_ids, ids, source_diss, diss, k, reverse):
if source_diss[k - 1] <= diss[0]:
return source_ids, source_diss
if diss[k - 1] <= source_diss[0]:
return ids, diss
source_diss.extend(diss)
diss_t = enumerate(source_diss)
diss_m_rst = sorted(diss_t, key=lambda x: x[1])[:k]
diss_m_out = [id_ for _, id_ in diss_m_rst]
source_ids.extend(ids)
id_m_out = [source_ids[i] for i, _ in diss_m_rst]
return id_m_out, diss_m_out
def _do_merge(self, files_n_topk_results, topk, reverse=False, **kwargs):
status = status_pb2.Status(error_code=status_pb2.SUCCESS,
reason="Success")
if not files_n_topk_results:
return status, [], []
merge_id_results = []
merge_dis_results = []
calc_time = time.time()
for files_collection in files_n_topk_results:
if isinstance(files_collection, tuple):
status, _ = files_collection
return status, [], []
if files_collection.status.error_code != 0:
return files_collection.status, [], []
row_num = files_collection.row_num
# row_num is equal to 0, result is empty
if not row_num:
continue
ids = files_collection.ids
diss = files_collection.distances # distance collections
# TODO: batch_len is equal to topk, may need to compare with topk
batch_len = len(ids) // row_num
for row_index in range(row_num):
id_batch = ids[row_index * batch_len: (row_index + 1) * batch_len]
dis_batch = diss[row_index * batch_len: (row_index + 1) * batch_len]
if len(merge_id_results) < row_index:
raise ValueError("merge error")
elif len(merge_id_results) == row_index:
# TODO: may bug here
merge_id_results.append(id_batch)
merge_dis_results.append(dis_batch)
else:
merge_id_results[row_index], merge_dis_results[row_index] = \
self._reduce(merge_id_results[row_index], id_batch,
merge_dis_results[row_index], dis_batch,
batch_len,
reverse)
calc_time = time.time() - calc_time
logger.info('Merge takes {}'.format(calc_time))
id_mrege_list = []
dis_mrege_list = []
for id_results, dis_results in zip(merge_id_results, merge_dis_results):
id_mrege_list.extend(id_results)
dis_mrege_list.extend(dis_results)
return status, id_mrege_list, dis_mrege_list
def _do_query(self,
context,
collection_id,
collection_meta,
vectors,
topk,
search_params,
partition_tags=None,
**kwargs):
metadata = kwargs.get('metadata', None)
routing = {}
p_span = None if self.tracer.empty else context.get_active_span(
).context
with self.tracer.start_span('get_routing', child_of=p_span):
routing = self.router.routing(collection_id,
partition_tags=partition_tags,
metadata=metadata)
logger.info('Routing: {}'.format(routing))
metadata = kwargs.get('metadata', None)
all_topk_results = []
with self.tracer.start_span('do_search', child_of=p_span) as span:
if len(routing) == 0:
ft = self.router.connection().search(collection_id, topk, vectors, list(partition_tags), search_params, _async=True)
ret = ft.result(raw=True)
all_topk_results.append(ret)
else:
futures = []
for addr, files_tuple in routing.items():
search_file_ids, ud_file_ids = files_tuple
logger.info(f"<{addr}> needed update segment ids {ud_file_ids}")
conn = self.router.query_conn(addr, metadata=metadata)
start = time.time()
ud_file_ids and conn.reload_segments(collection_id, ud_file_ids)
span = kwargs.get('span', None)
span = span if span else (None if self.tracer.empty else
context.get_active_span().context)
with self.tracer.start_span('search_{}'.format(addr),
child_of=span):
future = conn.search_in_segment(collection_name=collection_id,
file_ids=search_file_ids,
query_records=vectors,
top_k=topk,
params=search_params, _async=True)
futures.append(future)
for f in futures:
ret = f.result(raw=True)
all_topk_results.append(ret)
reverse = collection_meta.metric_type == Types.MetricType.IP
with self.tracer.start_span('do_merge', child_of=p_span):
return self._do_merge(all_topk_results,
topk,
reverse=reverse,
metadata=metadata)
def _create_collection(self, collection_schema):
return self.router.connection().create_collection(collection_schema)
@mark_grpc_method
def CreateCollection(self, request, context):
_status, unpacks = Parser.parse_proto_CollectionSchema(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
_status, _collection_schema = unpacks
# if _status.error_code != 0:
# logging.warning('[CreateCollection] collection schema error occurred: {}'.format(_status))
# return _status
logger.info('CreateCollection {}'.format(_collection_schema['collection_name']))
_status = self._create_collection(_collection_schema)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
def _has_collection(self, collection_name, metadata=None):
return self.router.connection(metadata=metadata).has_collection(collection_name)
@mark_grpc_method
def HasCollection(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return milvus_pb2.BoolReply(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
bool_reply=False)
logger.info('HasCollection {}'.format(_collection_name))
_status, _bool = self._has_collection(_collection_name,
metadata={'resp_class': milvus_pb2.BoolReply})
return milvus_pb2.BoolReply(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
bool_reply=_bool)
@mark_grpc_method
def CreatePartition(self, request, context):
_collection_name, _tag = Parser.parse_proto_PartitionParam(request)
_status = self.router.connection().create_partition(_collection_name, _tag)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
@mark_grpc_method
def DropPartition(self, request, context):
_collection_name, _tag = Parser.parse_proto_PartitionParam(request)
_status = self.router.connection().drop_partition(_collection_name, _tag)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
@mark_grpc_method
def HasPartition(self, request, context):
_collection_name, _tag = Parser.parse_proto_PartitionParam(request)
_status, _ok = self.router.connection().has_partition(_collection_name, _tag)
return milvus_pb2.BoolReply(status=status_pb2.Status(error_code=_status.code,
reason=_status.message), bool_reply=_ok)
@mark_grpc_method
def ShowPartitions(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return milvus_pb2.PartitionList(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
partition_array=[])
logger.info('ShowPartitions {}'.format(_collection_name))
_status, partition_array = self.router.connection().list_partitions(_collection_name)
return milvus_pb2.PartitionList(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
partition_tag_array=[param.tag for param in partition_array])
def _drop_collection(self, collection_name):
return self.router.connection().drop_collection(collection_name)
@mark_grpc_method
def DropCollection(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
logger.info('DropCollection {}'.format(_collection_name))
_status = self._drop_collection(_collection_name)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
def _create_index(self, collection_name, index_type, param):
return self.router.connection().create_index(collection_name, index_type, param)
@mark_grpc_method
def CreateIndex(self, request, context):
_status, unpacks = Parser.parse_proto_IndexParam(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
_collection_name, _index_type, _index_param = unpacks
logger.info('CreateIndex {}'.format(_collection_name))
# TODO: interface create_collection incompleted
_status = self._create_index(_collection_name, _index_type, _index_param)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
def _add_vectors(self, param, metadata=None):
return self.router.connection(metadata=metadata).insert(
None, None, insert_param=param)
@mark_grpc_method
def Insert(self, request, context):
logger.info('Insert')
# TODO: Ths SDK interface add_vectors() could update, add a key 'row_id_array'
_status, _ids = self._add_vectors(
metadata={'resp_class': milvus_pb2.VectorIds}, param=request)
return milvus_pb2.VectorIds(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
vector_id_array=_ids)
@mark_grpc_method
def Search(self, request, context):
metadata = {'resp_class': milvus_pb2.TopKQueryResult}
collection_name = request.collection_name
topk = request.topk
if len(request.extra_params) == 0:
raise exceptions.SearchParamError(message="Search parma loss", metadata=metadata)
params = ujson.loads(str(request.extra_params[0].value))
logger.info('Search {}: topk={} params={}'.format(
collection_name, topk, params))
# if nprobe > self.MAX_NPROBE or nprobe <= 0:
# raise exceptions.InvalidArgumentError(
# message='Invalid nprobe: {}'.format(nprobe), metadata=metadata)
if topk > self.MAX_TOPK or topk <= 0:
raise exceptions.InvalidTopKError(
message='Invalid topk: {}'.format(topk), metadata=metadata)
collection_meta = self.collection_meta.get(collection_name, None)
if not collection_meta:
status, info = self.router.connection(
metadata=metadata).get_collection_info(collection_name)
if not status.OK():
raise exceptions.CollectionNotFoundError(collection_name,
metadata=metadata)
self.collection_meta[collection_name] = info
collection_meta = info
start = time.time()
query_record_array = []
if int(collection_meta.metric_type) >= MetricType.HAMMING.value:
for query_record in request.query_record_array:
query_record_array.append(bytes(query_record.binary_data))
else:
for query_record in request.query_record_array:
query_record_array.append(list(query_record.float_data))
status, id_results, dis_results = self._do_query(context,
collection_name,
collection_meta,
query_record_array,
topk,
params,
partition_tags=getattr(request, "partition_tag_array", []),
metadata=metadata)
now = time.time()
logger.info('SearchVector takes: {}'.format(now - start))
topk_result_list = milvus_pb2.TopKQueryResult(
status=status_pb2.Status(error_code=status.error_code,
reason=status.reason),
row_num=len(request.query_record_array) if len(id_results) else 0,
ids=id_results,
distances=dis_results)
return topk_result_list
@mark_grpc_method
def SearchInFiles(self, request, context):
raise NotImplemented()
# @mark_grpc_method
# def SearchByID(self, request, context):
# metadata = {'resp_class': milvus_pb2.TopKQueryResult}
#
# collection_name = request.collection_name
#
# topk = request.topk
#
# if len(request.extra_params) == 0:
# raise exceptions.SearchParamError(message="Search param loss", metadata=metadata)
# params = ujson.loads(str(request.extra_params[0].value))
#
# logger.info('Search {}: topk={} params={}'.format(
# collection_name, topk, params))
#
# if topk > self.MAX_TOPK or topk <= 0:
# raise exceptions.InvalidTopKError(
# message='Invalid topk: {}'.format(topk), metadata=metadata)
#
# collection_meta = self.collection_meta.get(collection_name, None)
#
# if not collection_meta:
# status, info = self.router.connection(
# metadata=metadata).describe_collection(collection_name)
# if not status.OK():
# raise exceptions.CollectionNotFoundError(collection_name,
# metadata=metadata)
#
# self.collection_meta[collection_name] = info
# collection_meta = info
#
# start = time.time()
#
# query_record_array = []
# if int(collection_meta.metric_type) >= MetricType.HAMMING.value:
# for query_record in request.query_record_array:
# query_record_array.append(bytes(query_record.binary_data))
# else:
# for query_record in request.query_record_array:
# query_record_array.append(list(query_record.float_data))
#
# partition_tags = getattr(request, "partition_tag_array", [])
# ids = getattr(request, "id_array", [])
# search_result = self.router.connection(metadata=metadata).search_by_ids(collection_name, ids, topk, partition_tags, params)
# # status, id_results, dis_results = self._do_query(context,
# # collection_name,
# # collection_meta,
# # query_record_array,
# # topk,
# # params,
# # partition_tags=getattr(request, "partition_tag_array", []),
# # metadata=metadata)
#
# now = time.time()
# logger.info('SearchVector takes: {}'.format(now - start))
# return search_result
# #
# # topk_result_list = milvus_pb2.TopKQueryResult(
# # status=status_pb2.Status(error_code=status.error_code,
# # reason=status.reason),
# # row_num=len(request.query_record_array) if len(id_results) else 0,
# # ids=id_results,
# # distances=dis_results)
# # return topk_result_list
# # raise NotImplemented()
def _describe_collection(self, collection_name, metadata=None):
return self.router.connection(metadata=metadata).get_collection_info(collection_name)
@mark_grpc_method
def DescribeCollection(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return milvus_pb2.CollectionSchema(status=status_pb2.Status(
error_code=_status.code, reason=_status.message), )
metadata = {'resp_class': milvus_pb2.CollectionSchema}
logger.info('DescribeCollection {}'.format(_collection_name))
_status, _collection = self._describe_collection(metadata=metadata,
collection_name=_collection_name)
if _status.OK():
return milvus_pb2.CollectionSchema(
collection_name=_collection_name,
index_file_size=_collection.index_file_size,
dimension=_collection.dimension,
metric_type=_collection.metric_type,
status=status_pb2.Status(error_code=_status.code,
reason=_status.message),
)
return milvus_pb2.CollectionSchema(
collection_name=_collection_name,
status=status_pb2.Status(error_code=_status.code,
reason=_status.message),
)
def _collection_info(self, collection_name, metadata=None):
return self.router.connection(metadata=metadata).get_collection_stats(collection_name)
@mark_grpc_method
def ShowCollectionInfo(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return milvus_pb2.CollectionInfo(status=status_pb2.Status(
error_code=_status.code, reason=_status.message), )
metadata = {'resp_class': milvus_pb2.CollectionInfo}
_status, _info = self._collection_info(metadata=metadata, collection_name=_collection_name)
_info_str = ujson.dumps(_info)
if _status.OK():
return milvus_pb2.CollectionInfo(
status=status_pb2.Status(error_code=_status.code,
reason=_status.message),
json_info=_info_str
)
return milvus_pb2.CollectionInfo(
status=status_pb2.Status(error_code=_status.code,
reason=_status.message),
)
def _count_collection(self, collection_name, metadata=None):
return self.router.connection(
metadata=metadata).count_entities(collection_name)
@mark_grpc_method
def CountCollection(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
status = status_pb2.Status(error_code=_status.code,
reason=_status.message)
return milvus_pb2.CollectionRowCount(status=status)
logger.info('CountCollection {}'.format(_collection_name))
metadata = {'resp_class': milvus_pb2.CollectionRowCount}
_status, _count = self._count_collection(_collection_name, metadata=metadata)
return milvus_pb2.CollectionRowCount(
status=status_pb2.Status(error_code=_status.code,
reason=_status.message),
collection_row_count=_count if isinstance(_count, int) else -1)
def _get_server_version(self, metadata=None):
return self.router.connection(metadata=metadata).server_version()
def _cmd(self, cmd, metadata=None):
return self.router.connection(metadata=metadata)._cmd(cmd)
@mark_grpc_method
def Cmd(self, request, context):
_status, _cmd = Parser.parse_proto_Command(request)
logger.info('Cmd: {}'.format(_cmd))
if not _status.OK():
return milvus_pb2.StringReply(status=status_pb2.Status(
error_code=_status.code, reason=_status.message))
metadata = {'resp_class': milvus_pb2.StringReply}
if _cmd == 'conn_stats':
stats = self.router.readonly_topo.stats()
return milvus_pb2.StringReply(status=status_pb2.Status(
error_code=status_pb2.SUCCESS),
string_reply=json.dumps(stats, indent=2))
# if _cmd == 'version':
# _status, _reply = self._get_server_version(metadata=metadata)
# else:
# _status, _reply = self.router.connection(
# metadata=metadata).server_status()
_status, _reply = self._cmd(_cmd, metadata=metadata)
return milvus_pb2.StringReply(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
string_reply=_reply)
def _show_collections(self, metadata=None):
return self.router.connection(metadata=metadata).list_collections()
@mark_grpc_method
def ShowCollections(self, request, context):
logger.info('ShowCollections')
metadata = {'resp_class': milvus_pb2.CollectionName}
_status, _results = self._show_collections(metadata=metadata)
return milvus_pb2.CollectionNameList(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
collection_names=_results)
def _preload_collection(self, collection_name):
return self.router.connection().load_collection(collection_name)
@mark_grpc_method
def PreloadCollection(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
logger.info('PreloadCollection {}'.format(_collection_name))
_status = self._preload_collection(_collection_name)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
def ReloadSegments(self, request, context):
raise NotImplementedError("Not implemented in mishards")
def _describe_index(self, collection_name, metadata=None):
return self.router.connection(metadata=metadata).get_index_info(collection_name)
@mark_grpc_method
def DescribeIndex(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return milvus_pb2.IndexParam(status=status_pb2.Status(
error_code=_status.code, reason=_status.message))
metadata = {'resp_class': milvus_pb2.IndexParam}
logger.info('DescribeIndex {}'.format(_collection_name))
_status, _index_param = self._describe_index(collection_name=_collection_name,
metadata=metadata)
if not _index_param:
return milvus_pb2.IndexParam(status=status_pb2.Status(
error_code=_status.code, reason=_status.message))
_index_type = _index_param._index_type
grpc_index = milvus_pb2.IndexParam(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
collection_name=_collection_name, index_type=_index_type)
grpc_index.extra_params.add(key='params', value=ujson.dumps(_index_param._params))
return grpc_index
def _get_vectors_by_id(self, collection_name, ids, metadata):
return self.router.connection(metadata=metadata).get_entity_by_id(collection_name, ids)
@mark_grpc_method
def GetVectorsByID(self, request, context):
_status, unpacks = Parser.parse_proto_VectorIdentity(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
metadata = {'resp_class': milvus_pb2.VectorsData}
_collection_name, _ids = unpacks
logger.info('GetVectorByID {}'.format(_collection_name))
_status, vectors = self._get_vectors_by_id(_collection_name, _ids, metadata)
_rpc_status = status_pb2.Status(error_code=_status.code, reason=_status.message)
if not vectors:
return milvus_pb2.VectorsData(status=_rpc_status, )
if len(vectors) == 0:
return milvus_pb2.VectorsData(status=_rpc_status, vectors_data=[])
if isinstance(vectors[0], bytes):
records = [milvus_pb2.RowRecord(binary_data=v) for v in vectors]
else:
records = [milvus_pb2.RowRecord(float_data=v) for v in vectors]
response = milvus_pb2.VectorsData(status=_rpc_status)
response.vectors_data.extend(records)
return response
def _get_vector_ids(self, collection_name, segment_name, metadata):
return self.router.connection(metadata=metadata).list_id_in_segment(collection_name, segment_name)
@mark_grpc_method
def GetVectorIDs(self, request, context):
_status, unpacks = Parser.parse_proto_GetVectorIDsParam(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
metadata = {'resp_class': milvus_pb2.VectorIds}
_collection_name, _segment_name = unpacks
logger.info('GetVectorIDs {}'.format(_collection_name))
_status, ids = self._get_vector_ids(_collection_name, _segment_name, metadata)
if not ids:
return milvus_pb2.VectorIds(status=status_pb2.Status(
error_code=_status.code, reason=_status.message), )
return milvus_pb2.VectorIds(status=status_pb2.Status(
error_code=_status.code, reason=_status.message),
vector_id_array=ids
)
def _delete_by_id(self, collection_name, id_array):
return self.router.connection().delete_entity_by_id(collection_name, id_array)
@mark_grpc_method
def DeleteByID(self, request, context):
_status, unpacks = Parser.parse_proto_DeleteByIDParam(request)
if not _status.OK():
logging.error('DeleteByID {}'.format(_status.message))
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
_collection_name, _ids = unpacks
logger.info('DeleteByID {}'.format(_collection_name))
_status = self._delete_by_id(_collection_name, _ids)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
def _drop_index(self, collection_name):
return self.router.connection().drop_index(collection_name)
@mark_grpc_method
def DropIndex(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
logger.info('DropIndex {}'.format(_collection_name))
_status = self._drop_index(_collection_name)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
def _flush(self, collection_names):
return self.router.connection().flush(collection_names)
@mark_grpc_method
def Flush(self, request, context):
_status, _collection_names = Parser.parse_proto_FlushParam(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
logger.info('Flush {}'.format(_collection_names))
_status = self._flush(_collection_names)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
def _compact(self, collection_name):
return self.router.connection().compact(collection_name)
@mark_grpc_method
def Compact(self, request, context):
_status, _collection_name = Parser.parse_proto_CollectionName(request)
if not _status.OK():
return status_pb2.Status(error_code=_status.code,
reason=_status.message)
logger.info('Compact {}'.format(_collection_name))
_status = self._compact(_collection_name)
return status_pb2.Status(error_code=_status.code,
reason=_status.message)