# Milvus Developer Guides by Rentong Guo, Sep 15, 2020 ## 1. System Overview In this section, we sketch the system design of Milvus , including data model, data organization, architecture, and state synchronization. #### 1.1 Data Model Milvus exposes the following set of data features to applications: * a data model based on schematized relational tables, in that rows must have primary-keys, * a query language specifies data definition, data manipulation, and data query, where data definition includes create, drop, and data manipulation includes insert, upsert, delete, and data query falls into three types, primary key search, approximate nearest neighbor search (ANNS), ANNS with predicates. The requests' execution order is strictly in accordance with their issue-time order. We take proxy's issue time as a requst's issue time. For a batch request, all its sub-requests share a same issue time. In cases there are multiple proxies, issue time from different proxies are regarded as coming from a central clock. Transaction is currently not supported by Milvus. Only batch requests such as batch insert/delete/query are supported. A batch insert/delete is guaranteed to become visible atomically. #### 1.2 Data Organization

In Milvus, 'collection' refers to the concept of table. A collection can be optionally divided into several 'partitions'. Both collection and partition are the basic execution scopes of queries. When use parition, users should clearly know how a collection should be partitioned. In most cases, parition leads to more flexible data management and more efficient quering. For a partitioned collection, queries can be executed both on the collection or a set of specified partitions. Each collection or parition contains a set of 'segment groups'. Segment group is the basic unit of data-to-node mapping. It's also the basic unit of replica. For instance, if a query node failed, its segment groups will be redistributed accross other nodes. If a query node is overloaded, part of its segment groups will be migrated to underloaded ones. If a hot collection/partition is detected, its segment groups will be replicated to smooth the system load skewness. 'Segment' is the finest unit of data organization. It is where the data and indexes are actually kept. Each segment contains a set of rows. In order to reduce the memory footprint during a query execution and to fully utilize SIMD, the physical data layout within segments is organized in a column-based manner. #### 1.3 Architecture Overview

The main components, proxy, WAL, query node and write node can scale to multiple instances. These components scale seperately for better tradeoff between availability and cost. The WAL forms a hash ring. Requests (i.e. inserts and deletes) from clients will be repacked by proxy. Operations shared identical hash value (the hash value of primary key) will be routed to the same hash bucket. In addtion, some preprocessing work will be done by proxy, such as static validity checking, primary key assignment (if not given by user), timestamp assignment. The query/write nodes are linked to the hash ring, with each node covers some portion of the buckets. Once the hash function and bucket coverage are settled, the chain 'proxy -> WAL -> query/write node' will act as a producer-consumer pipeline. Logs in each bucket is a determined operation stream. Via performing the operation stream in order, the query nodes keep themselves up to date. The query nodes hold all the indexes in memory. Since building index is time-consuming, the query nodes will dump their index to disk (store engine) for fast failure recovery and cross node index copy. The write nodes are stateless. They simply transforms the newly arrived WALs to binlog format, then append the binlog to store enginey. Note that not all the components are necessarily replicated. The system provides failure tolerance by maintaining multiple copies of WAL and binlog. When there is no in-memory index replica and there occurs a query node failure, other query nodes will take over its indexes by loading the dumped index files, or rebuilding them from binlog and WALs. The links from query nodes to the hash ring will also be adjusted such that the failure node's input WAL stream can be properly handled by its neighbors. #### 1.4 State Synchronization

Data in Milvus have three different forms, namely WAL, binlog, and index. As mentioned in the previous section, WAL can be viewed as a determined operation stream. Other two data forms keep themselves up to date by performing the operation stream in time order. Each of the WAL is attached with a timestamp, which is the time when the log is sent to the hash bucket. Binlog records, table rows, index cells will also keep that timestamp. In this way, different data forms can offer consistent snapshot for a given time T. For example, requests such as "fetch binlogs before T for point-in-time recovery", "get the row with primary key K at time T", "launch a similarity search at time T for vector V" perform on binlog, index respectively. Though different data forms these three requests are performed, they observe identical snapshot, namely all the state changes before T. For better throughput, Milvus allows asynchronous state synchronization between WAL and index/binlog/table. Whenever the data is not fresh enough to satisfiy a query, the query will be suspended until the data is up-to-date, or timeout will be returned. ## 2. Schema #### 2.1 Collection Schema ``` go type CollectionSchema struct { Name string Description string AutoId bool Fields []FieldSchema } ``` #### 2.2 Field Schema ``` go type FieldSchema struct { Name string Description string DataType DataType TypeParams map[string]string IndexParams map[string]string } ``` ###### 2.2.1 Data Types ###### 2.2.2 Type Params ###### 2.2.3 Index Params ## 3. Request In this section, we introduce the RPCs of milvus service. A brief description of the RPCs is listed as follows. | RPC | description | | :----------------- | ------------------------------------------------------------ | | CreateCollection | create a collection base on schema statement | | DropCollection | drop a collection | | HasCollection | whether or not a collection exists | | DescribeCollection | show a collection's schema and its descriptive statistics | | ShowCollections | list all collections | | CreatePartition | create a partition | | DropPartition | drop a partition | | HasPartition | whether or not a partition exists | | DescribePartition | show a partition's name and its descriptive statistics | | ShowPartitions | list a collection's all partitions | | Insert | insert a batch of rows into a collection or a partition | | Search | query the columns of a collection or a partition with ANNS statements and boolean expressions | #### 3.1 Definition Requests ###### 3.2.1 Collection * CreateCollection * DropCollection * HasCollection * DescribeCollection * ShowCollections ###### 3.2.2 Partition * CreatePartition * DropPartition * HasPartition * DescribePartition * ShowPartitions #### 3.2 Manipulation Requsts ###### 3.2.1 Insert * Insert ###### 3.2.2 Delete * DeleteByID #### 3.3 Query ## 4. Time #### 4.1 Timestamp Before we discuss timestamp, let's take a brief review of Hybrid Logical Clock (HLC). HLC uses 64bits timestamps which are composed of a 46-bits physical component (thought of as and always close to local wall time) and a 18-bits logical component (used to distinguish between events with the same physical component).

HLC's logical part is advanced on each request. The phsical part can be increased in two cases: A. when the local wall time is greater than HLC's physical part, B. or the logical part overflows. In either cases, the physical part will be updated, and the logical part will be set to 0. Keep the physical part close to local wall time may face non-monotonic problems such as updates to POSIX time that could turn time backward. HLC avoids such problems, since if 'local wall time < HLC's physical part' holds, only case B is satisfied, thus montonicity is guaranteed. Milvus does not support transaction, but it should gurantee the deterministic execution of the multi-way WAL. The timestamp attached to each request should - have its physical part close to wall time (has an acceptable bounded error, a.k.a. uncertainty interval in transaction senarios), - and be globally unique. HLC leverages on physical clocks at nodes that are synchronized using the NTP. NTP usually maintain time to within tens of milliseconds over local networks in datacenter. Asymmetric routes and network congestion occasionally cause errors of hundreds of milliseconds. Both the normal time error and the spike are acceptable for Milvus use cases. The interface of Timestamp is as follows. ``` type timestamp struct { physical uint64 // 18-63 bits logical uint64 // 0-17 bits } type Timestamp uint64 ``` #### 4.2 Timestamp Oracle ```go type timestampOracle struct { client *etcd.Client // client of a reliable meta service, i.e. etcd client rootPath string // this timestampOracle's working root path on the reliable kv service saveInterval uint64 lastSavedTime uint64 tso Timestamp // monotonically increasing timestamp } func (tso *timestampOracle) GetTimestamp(count uint32) ([]Timestamp, error) func (tso *timestampOracle) saveTimestamp() error func (tso *timestampOracle) loadTimestamp() error ``` #### 4.2 Timestamp Allocator ```go type TimestampAllocator struct { Alloc(count uint32) ([]Timestamp, error) } func (allocator *TimestampAllocator) Start() error func (allocator *TimestampAllocator) Close() error func NewTimestampAllocator() *TimestampAllocator ``` ###### 4.2.1 Batch Allocation of Timestamps ###### 4.2.2 Expiration of Timestamps #### 4.5 T_safe ## 5. Basic Components #### 5.1 Watchdog ``` go type ActiveComponent interface { Id() string Status() Status Clean() Status Restart() Status } type ComponentHeartbeat interface { Id() string Status() Status Serialize() string } type Watchdog struct { targets [] *ActiveComponent heartbeats ComponentHeartbeat chan } // register ActiveComponent func (dog *Watchdog) Register(target *ActiveComponent) // called by ActiveComponents func (dog *Watchdog) PutHeartbeat(heartbeat *ComponentHeartbeat) // dump heatbeats as log stream func (dog *Watchdog) dumpHeartbeat(heartbeat *ComponentHeartbeat) ``` #### 5.2 Global Parameter Table ``` go type GlobalParamsTable struct { params memoryKV } func (gparams *GlobalParamsTable) Save(key, value string) error func (gparams *GlobalParamsTable) Load(key string) (string, error) func (gparams *GlobalParamsTable) LoadRange(key, endKey string, limit int) ([]string, []string, error) func (gparams *GlobalParamsTable) Remove(key string) error ``` #### 5.3 Message Stream ``` go type MsgType uint32 const { USER_REQUEST MsgType = 1 TIME_TICK = 2 } type TsMsg interface { SetTs(ts Timestamp) Ts() Timestamp Type() MsgType } type MsgPack struct { BeginTs Timestamp EndTs Timestamp Msgs []*TsMsg } type TsMsgMarshaler interface { Marshal(input *TsMsg) ([]byte, error) Unmarshal(input []byte) (*TsMsg, error) } type MsgStream interface { SetMsgMarshaler(marshal *TsMsgMarshaler, unmarshal *TsMsgMarshaler) Produce(*MsgPack) error Consume() *MsgPack // message can be consumed exactly once } type HashFunc func(*MsgPack) map[int32]*MsgPack type PulsarMsgStream struct { client *pulsar.Client msgHashFunc HashFunc // return a map from produceChannel idx to *MsgPack producers []*pulsar.Producer consumers []*pulsar.Consumer msgMarshaler *TsMsgMarshaler msgUnmarshaler *TsMsgMarshaler } func (ms *PulsarMsgStream) SetProducerChannels(channels []string) func (ms *PulsarMsgStream) SetConsumerChannels(channels []string) func (ms *PulsarMsgStream) SetMsgMarshaler(marshal *TsMsgMarshaler, unmarshal *TsMsgMarshaler) func (ms *PulsarMsgStream) SetMsgHashFunc(hashFunc *HashFunc) func (ms *PulsarMsgStream) Produce(msgs *MsgPack) error func (ms *PulsarMsgStream) Consume() (*MsgPack, error) //return messages in one time tick type PulsarTtMsgStream struct { client *pulsar.Client msgHashFunc (*MsgPack) map[int32]*MsgPack // return a map from produceChannel idx to *MsgPack producers []*pulsar.Producer consumers []*pulsar.Consumer msgMarshaler *TsMsgMarshaler msgUnmarshaler *TsMsgMarshaler inputBuf []*TsMsg unsolvedBuf []*TsMsg msgPacks []*MsgPack } func (ms *PulsarMsgStream) SetProducerChannels(channels []string) func (ms *PulsarMsgStream) SetConsumerChannels(channels []string) func (ms *PulsarMsgStream) SetMsgMarshaler(marshal *TsMsgMarshaler, unmarshal *TsMsgMarshaler) func (ms *PulsarMsgStream) SetMsgHashFunc(hashFunc *HashFunc) func (ms *PulsarMsgStream) Produce(msgs *MsgPack) error func (ms *PulsarMsgStream) Consume() *MsgPack //return messages in one time tick ``` #### 5.4 ID Allocator ```go type IdAllocator struct { Alloc(count uint32) ([]int64, error) } func (allocator *IdAllocator) Start() error func (allocator *IdAllocator) Close() error func NewIdAllocator() *IdAllocator ``` ## 6. Proxy #### 6.1 Proxy Instance ```go type Proxy struct { servicepb.UnimplementedMilvusServiceServer masterClient mpb.MasterClient timeTick *timeTick ttStream *MessageStream scheduler *taskScheduler tsAllocator *TimestampAllocator ReqIdAllocator *IdAllocator RowIdAllocator *IdAllocator SegIdAssigner *segIdAssigner } func (proxy *Proxy) Start() error func NewProxy(ctx context.Context) *Proxy ``` #### Global Parameter Table ```go type GlobalParamsTable struct { } func (*paramTable GlobalParamsTable) ProxyId() int64 func (*paramTable GlobalParamsTable) ProxyAddress() string func (*paramTable GlobalParamsTable) MasterAddress() string func (*paramTable GlobalParamsTable) PulsarAddress() string func (*paramTable GlobalParamsTable) TimeTickTopic() string func (*paramTable GlobalParamsTable) InsertTopics() []string func (*paramTable GlobalParamsTable) QueryTopic() string func (*paramTable GlobalParamsTable) QueryResultTopics() []string func (*paramTable GlobalParamsTable) Init() error var ProxyParamTable GlobalParamsTable ``` #### 6.2 Task ``` go type task interface { Id() int64 // return ReqId PreExecute() error Execute() error PostExecute() error WaitToFinish() error Notify() error } ``` * Base Task ```go type baseTask struct { Type ReqType ReqId int64 Ts Timestamp ProxyId int64 } func (task *baseTask) PreExecute() error func (task *baseTask) Execute() error func (task *baseTask) PostExecute() error func (task *baseTask) WaitToFinish() error func (task *baseTask) Notify() error ``` * Insert Task Take insertTask as an example: ```go type insertTask struct { baseTask SegIdAssigner *segIdAssigner RowIdAllocator *IdAllocator rowBatch *RowBatch } func (task *InsertTask) Execute() error func (task *InsertTask) WaitToFinish() error func (task *InsertTask) Notify() error ``` #### 6.2 Task Scheduler * Base Task Queue ```go type baseTaskQueue struct { unissuedTasks *List activeTasks map[int64]*task utLock sync.Mutex // lock for UnissuedTasks atLock sync.Mutex // lock for ActiveTasks } func (queue *baseTaskQueue) AddUnissuedTask(task *task) func (queue *baseTaskQueue) FrontUnissuedTask() *task func (queue *baseTaskQueue) PopUnissuedTask(id int64) *task func (queue *baseTaskQueue) AddActiveTask(task *task) func (queue *baseTaskQueue) PopActiveTask(id int64) *task func (queue *baseTaskQueue) TaskDoneTest(ts Timestamp) bool ``` *AddUnissuedTask(task \*task)* will put a new task into *unissuedTasks*, while maintaining the list by timestamp order. *TaskDoneTest(ts Timestamp)* will check both *unissuedTasks* and *unissuedTasks*. If no task found before *ts*, then the function returns *true*, indicates that all the tasks before *ts* are completed. * Data Definition Task Queue ```go type ddTaskQueue struct { baseTaskQueue lock sync.Mutex } func (queue *ddTaskQueue) Enqueue(task *task) error func newDdTaskQueue() *ddTaskQueue ``` Data definition tasks (i.e. *CreateCollectionTask*) will be put into *DdTaskQueue*. If a task is enqueued, *Enqueue(task \*task)* will set *Ts*, *ReqId*, *ProxyId*, then push it into *queue*. The timestamps of the enqueued tasks should be strictly monotonically increasing. As *Enqueue(task \*task)* will be called in parallel, setting timestamp and queue insertion need to be done atomically. * Data Manipulation Task Queue ```go type dmTaskQueue struct { baseTaskQueue } func (queue *dmTaskQueue) Enqueue(task *task) error func newDmTaskQueue() *dmTaskQueue ``` Insert tasks and delete tasks will be put into *DmTaskQueue*. If a *insertTask* is enqueued, *Enqueue(task \*task)* will set *Ts*, *ReqId*, *ProxyId*, *SegIdAssigner*, *RowIdAllocator*, then push it into *queue*. The *SegIdAssigner* and *RowIdAllocator* will later be used in the task's execution phase. * Data Query Task Queue ```go type dqTaskQueue struct { baseTaskQueue } func (queue *dqTaskQueue) Enqueue(task *task) error func newDqTaskQueue() *dqTaskQueue ``` Queries will be put into *DqTaskQueue*. * Task Scheduler ``` go type taskScheduler struct { DdQueue *ddTaskQueue DmQueue *dmTaskQueue DqQueue *dqTaskQueue tsAllocator *TimestampAllocator ReqIdAllocator *IdAllocator } func (sched *taskScheduler) scheduleDdTask() *task func (sched *taskScheduler) scheduleDmTask() *task func (sched *taskScheduler) scheduleDqTask() *task func (sched *taskScheduler) Start() error func (sched *taskScheduler) TaskDoneTest(ts Timestamp) bool func newTaskScheduler(ctx context.Context, tsAllocator *TimestampAllocator, ReqIdAllocator *IdAllocator) *taskScheduler ``` *scheduleDdTask()* selects tasks in a FIFO manner, thus time order is garanteed. The policy of *scheduleDmTask()* should target on throughput, not tasks' time order. Note that the time order of the tasks' execution will later be garanteed by the timestamp & time tick mechanism. The policy of *scheduleDqTask()* should target on throughput. It should also take visibility into consideration. For example, if an insert task and a query arrive in a same time tick and the query comes after insert, the query should be scheduled in the next tick thus the query can see the insert. *TaskDoneTest(ts Timestamp)* will check all the three task queues. If no task found before *ts*, then the function returns *true*, indicates that all the tasks before *ts* are completed. * Statistics ```go // ActiveComponent interfaces func (sched *taskScheduler) Id() String func (sched *taskScheduler) Status() Status func (sched *taskScheduler) Clean() Status func (sched *taskScheduler) Restart() Status func (sched *taskScheduler) heartbeat() // protobuf message taskSchedulerHeartbeat { string id uint64 dd_queue_length uint64 dm_queue_length uint64 dq_queue_length uint64 num_dd_done uint64 num_dm_done uint64 num_dq_done } ``` #### 6.3 Time Tick * Time Tick ``` go type timeTick struct { lastTick Timestamp currentTick Timestamp wallTick Timestamp tickStep Timestamp syncInterval Timestamp tsAllocator *TimestampAllocator scheduler *taskScheduler ttStream *MessageStream ctx context.Context } func (tt *timeTick) Start() error func (tt *timeTick) synchronize() error func newTimeTick(ctx context.Context, tickStep Timestamp, syncInterval Timestamp, tsAllocator *TimestampAllocator, scheduler *taskScheduler, ttStream *MessageStream) *timeTick ``` *Start()* will enter a loop. On each *tickStep*, it tries to send a *TIME_TICK* typed *TsMsg* into *ttStream*. After each *syncInterval*, it sychronizes its *wallTick* with *tsAllocator* by calling *synchronize()*. When *currentTick + tickStep < wallTick* holds, it will update *currentTick* with *wallTick* on next tick. Otherwise, it will update *currentTick* with *currentTick + tickStep*. * Statistics ```go // ActiveComponent interfaces func (tt *timeTick) ID() String func (tt *timeTick) Status() Status func (tt *timeTick) Clean() Status func (tt *timeTick) Restart() Status func (tt *timeTick) heartbeat() // protobuf message TimeTickHeartbeat { string id uint64 last_tick } ``` ## 8. Query Node #### 8.1 Collection and Segment Meta ###### 8.1.1 Collection ``` go type Collection struct { Name string Id uint64 Fields map[string]FieldMeta SegmentsId []uint64 cCollectionSchema C.CCollectionSchema } ``` ###### 8.1.2 Field Meta ```go type FieldMeta struct { Name string Id uint64 IsPrimaryKey bool TypeParams map[string]string IndexParams map[string]string } ``` ###### 8.1.3 Segment ``` go type Segment struct { Id uint64 ParitionName string CollectionId uint64 OpenTime Timestamp CloseTime Timestamp NumRows uint64 cSegment C.CSegmentBase } ``` #### 8.2 Message Streams ```go type ManipulationReqUnmarshaler struct {} // implementations of MsgUnmarshaler interfaces func (unmarshaler *InsertMsgUnmarshaler) Unmarshal(input *pulsar.Message) (*TsMsg, error) type QueryReqUnmarshaler struct {} // implementations of MsgUnmarshaler interfaces func (unmarshaler *QueryReqUnmarshaler) Unmarshal(input *pulsar.Message) (*TsMsg, error) ``` #### 8.3 Query Node ## 4. Storage Engine #### 4.X Interfaces ## 5. Master #### 5.1 Interfaces (RPC) | RPC | description | | :----------------- | ------------------------------------------------------------ | | CreateCollection | create a collection base on schema statement | | DropCollection | drop a collection | | HasCollection | whether or not a collection exists | | DescribeCollection | show a collection's schema and its descriptive statistics | | ShowCollections | list all collections | | CreatePartition | create a partition | | DropPartition | drop a partition | | HasPartition | whether or not a partition exists | | DescribePartition | show a partition's name and its descriptive statistics | | ShowPartitions | list a collection's all partitions | | AllocTimestamp | allocate a batch of consecutive timestamps | | AllocId | allocate a batch of consecutive IDs | | AssignSegmentId | assign segment id to insert rows (master determines which segment these rows belong to) | | | | | | | #### 5.2 Master Instance ```go type Master interface { tso timestampOracle // timestamp oracle ddScheduler ddRequestScheduler // data definition request scheduler metaTable metaTable // in-memory system meta collManager collectionManager // collection & partition manager segManager segmentManager // segment manager } ``` * Timestamp allocation Master serves as a centrol clock of the whole system. Other components (i.e. Proxy) allocates timestamps from master via RPC *AllocTimestamp*. All the timestamp allocation requests will be handled by the timestampOracle singleton. See section 4.2 for the details about timestampOracle. * Request Scheduling * System Meta * Collection Management * Segment Management #### 5.3 Data definition Request Scheduler ###### 5.2.1 Task Master receives data definition requests via grpc. Each request (described by a proto) will be wrapped as a task for further scheduling. The task interface is ```go type task interface { Type() ReqType Ts() Timestamp Execute() error WaitToFinish() error Notify() error } ``` A task example is as follows. In this example, we wrap a CreateCollectionRequest (a proto) as a createCollectionTask. The wrapper need to contain task interfaces. ``` go type createCollectionTask struct { req *CreateCollectionRequest cv int chan } // Task interfaces func (task *createCollectionTask) Type() ReqType func (task *createCollectionTask) Ts() Timestamp func (task *createCollectionTask) Execute() error func (task *createCollectionTask) Notify() error func (task *createCollectionTask) WaitToFinish() error ``` ###### 5.2.2 Scheduler ```go type ddRequestScheduler struct { reqQueue *task chan } func (rs *ddRequestScheduler) Enqueue(task *task) error func (rs *ddRequestScheduler) schedule() *task // implement scheduling policy ``` #### 5.4 Meta Table ```go type metaTable struct { client *etcd.Client // client of a reliable kv service, i.e. etcd client rootPath string // this metaTable's working root path on the reliable kv service tenantMeta map[int64]TenantMeta // tenant id to tenant meta proxyMeta map[int64]ProxyMeta // proxy id to proxy meta collMeta map[int64]CollectionMeta // collection id to collection meta segMeta map[int64]SegmentMeta // segment id to segment meta } ```