milvus/shards

Mishards - An Experimental Sharding Middleware

中文版

Milvus aims to achieve efficient similarity search and analytics for massive-scale vectors. A standalone Milvus instance can easily handle vector search among billion-scale vectors. However, for 10 billion, 100 billion or even larger datasets, a Milvus cluster is needed.

Ideally, this cluster can be accessed and used just as the standalone instance, meanwhile it satisfies the business requirements such as low latency and high concurrency.

This page meant to demonstrates how to use Mishards, an experimental sharding middleware for Milvus, to establish an orchestrated cluster.

What is Mishards

Mishards is a middleware that is developed using Python. It provides unlimited extension of memory and computation capacity through request forwarding, read/write splitting, horizontal scalability and dynamic extension. It works as the proxy of the Milvus system.

Using Mishards in Milvus cluster deployment is an experimental feature available for user test and feedback.

How Mishards works

Mishards splits the upstream requests to sub-requests and forwards them to Milvus servers. When the search computation is completed, all results are collected by Mishards and sent back to the client.

Below graph is a demonstration of the process:

Mishards example codes

Below examples codes demonstrate how to build from source code a Milvus server with Mishards on a standalone machine, as well as how to use Kubernetes to establish Milvus cluster with Mishards.

Before executing these examples, make sure you meet the prerequisites of Milvus installation.

Build from source code

Prequisites

Make sure Python 3.6 or higher is installed.

Start Milvus and Mishards from source code

Follow below steps to start a standalone Milvus instance with Mishards from source code:

1. Clone milvus repository.

   git clone <milvus repo http/ssh url>
   

2. Install Mishards dependencies.

   $ cd milvus/shards
   $ pip install -r requirements.txt
   

3. Start Milvus server.

   $ sudo nvidia-docker run --rm -d -p 19530:19530 -v /tmp/milvus/db:/opt/milvus/db milvusdb/milvus
   

4. Update path permissions.

   $ sudo chown -R $USER:$USER /tmp/milvus
   

5. Configure Mishards environmental variables.

   $ cp mishards/.env.example mishards/.env
   

6. Start Mishards server.

   $ python mishards/main.py
   

Docker example

The all_in_one example shows how to use Docker container to start 2 Milvus instances, 1 Mishards instance and 1 Jaeger instance.

1. Install Docker Compose.

2. Build docker images for these instances.

   $ make build
   

3. Start all instances.

   $ make deploy
   

4. Confirm instance status.

   $ make probe_deploy
        Pass ==> Pass: Connected
        Fail ==> Error: Fail connecting to server on 127.0.0.1:19530. Timeout
   

To check the service tracing, open the Jaeger page on your browser.

To stop all instances, use the following command:

$ make clean_deploy

Kubernetes example

Using Kubernetes to deploy Milvus cluster requires that the developers have a basic understanding of general concepts of Kubernetes.

This example mainly demonstrates how to use Kubernetes to establish a Milvus cluster containing 2 Milvus instances（1 read instance and 1 write instance), 1 MySQL instance and 1 Mishards instance.

This example does not include tasks such as setting up Kubernetes cluster, installing shared storage and using command tools such as kubectl.

Below is the architecture of Milvus cluster built upon Kubernetes:

Prerequisites

• A Kubernetes cluster is already established.
• nvidia-docker 2.0 is already installed.
• Shared storage is already installed.
• kubectl is installed and can access the Kubernetes cluster.

Use Kubernetes to build a Milvus cluster

1. Start Milvus cluster

   $ make cluster
   

2. Confirm that Mishards is connected to Milvus.

   $ make probe_cluster
     Pass ==> Pass: Connected
   

To check cluster status:

$ make cluster_status

To delete the cluster:

$ make clean_cluster

To add a read instance:

$ cd kubernetes_demo
$ ./start.sh scale-ro-server 2 

To add a proxy instance:

$ cd kubernetes_demo
$ ./start.sh scale-proxy 2 

To check cluster logs:

$ kubectl logs -f --tail=1000 -n milvus milvus-ro-servers-0 

Mishards Unit test

Unit test

$ cd milvus/shards
$ make test

Code coverage test

$ cd milvus/shards
$ make coverage

Code format check

$ cd milvus/shards
$ make style

Mishards configuration

Overall configuration

Name	Required	Type	Default	Description
Debug	No	boolean	True	Choose if to enable Debug work mode.
TIMEZONE	No	string	UTC	Timezone
MAX_RETRY	No	integer	3	The maximum retry times allowed to connect to Milvus.
SERVER_PORT	No	integer	19530	Define the server port of Mishards.
WOSERVER	Yes	string		Define the address of Milvus write instance. Currently, only static settings are supported. Format for reference: tcp://127.0.0.1:19530.

Metadata

Name	Required	Type	Default	Description
SQLALCHEMY_DATABASE_URI	Yes	string		Define the database address for metadata storage. Format standard: RFC-738-style. For example: mysql+pymysql://root:root@127.0.0.1:3306/milvus?charset=utf8mb4.
SQL_ECHO	No	boolean	False	Choose if to print SQL statements.
SQLALCHEMY_DATABASE_TEST_URI	No	string		Define the database address of metadata storage in test environment.
SQL_TEST_ECHO	No	boolean	False	Choose if to print SQL statements in test environment.

Service discovery

Name	Required	Type	Default	Description
DISCOVERY_PLUGIN_PATH	No	string		Define the search path to locate the plug-in. The default path is used if the value is not set.
DISCOVERY_CLASS_NAME	No	string	static	Under the plug-in search path, search the class based on the class name, and instantiate it. Currently, the system provides 2 classes: static and kubernetes.
DISCOVERY_STATIC_HOSTS	No	list	[]	When DISCOVERY_CLASS_NAME is static , define a comma-separated service address list, for example192.168.1.188,192.168.1.190.
DISCOVERY_STATIC_PORT	No	integer	19530	When DISCOVERY_CLASS_NAME is static, define the server port.
DISCOVERY_KUBERNETES_NAMESPACE	No	string		When DISCOVERY_CLASS_NAME is kubernetes, define the namespace of Milvus cluster.
DISCOVERY_KUBERNETES_IN_CLUSTER	No	boolean	False	When DISCOVERY_CLASS_NAME is kubernetes , choose if to run the server in Kubernetes.
DISCOVERY_KUBERNETES_POLL_INTERVAL	No	integer	5 (Seconds)	When DISCOVERY_CLASS_NAME is kubernetes , define the listening cycle of the server.
DISCOVERY_KUBERNETES_POD_PATT	No	string		When DISCOVERY_CLASS_NAME is kubernetes , map the regular expression of Milvus Pod.
DISCOVERY_KUBERNETES_LABEL_SELECTOR	No	string		When SD_PROVIDER is kubernetes, map the label of Milvus Pod. For example: tier=ro-servers.

Tracing

Name	Required	Type	Default	Description
TRACER_PLUGIN_PATH	No	string		Define the search path to locate the tracing plug-in. The default path is used if the value is not set.
TRACER_CLASS_NAME	No	string		Under the plug-in search path, search the class based on the class name, and instantiate it. Currently, only Jaeger is supported.
TRACING_SERVICE_NAME	No	string	mishards	When TRACING_CLASS_NAME is Jaeger, the name of the tracing service.
TRACING_SAMPLER_TYPE	No	string	const	When TRACING_CLASS_NAME is Jaeger, the sampling type of the tracing service.
TRACING_SAMPLER_PARAM	No	integer	1	When TRACING_CLASS_NAME is Jaeger, the sampling frequency of the tracing service.
TRACING_LOG_PAYLOAD	No	boolean	False	When TRACING_CLASS_NAME is Jaeger, choose if to sample Payload.

Logging

Name	Required	Type	Default	Description
LOG_LEVEL	No	string	DEBUG	Log recording levels. Currently supports DEBUG ,INFO ,WARNING and ERROR.
LOG_PATH	No	string	/tmp/mishards	Log recording path.
LOG_NAME	No	string	logfile	Log recording name.

Routing

Name	Required	Type	Default	Description
ROUTER_PLUGIN_PATH	No	string		Define the search path to locate the routing plug-in. The default path is used if the value is not set.
ROUTER_CLASS_NAME	No	string	FileBasedHashRingRouter	Under the plug-in search path, search the class based on the class name, and instantiate it. Currently, only FileBasedHashRingRouter is supported.
ROUTER_CLASS_TEST_NAME	No	string	FileBasedHashRingRouter	Under the plug-in search path, search the class based on the class name, and instantiate it. Currently, FileBasedHashRingRouter is supported for test environment only.