基于Milevus的ANN服务踩坑

基于Milvus的ANN服务踩坑

坑1：Exception in thread “main” java.lang.NoClassDefFoundError: org/slf4j/LoggerFactory

SLF4J jar 需要被添加到classpath

添加依赖。

 
     org.slf4j
     slf4j-api
     1.7.30
 

坑2：NoClassDefFoundError: org/apache/logging/log4j/util/StackLocatorUtil

项目中已经有log4j了，与milevus-sdk.pom中的log4j产生冲突，将其exclusion即可。

    io.milvus
    milvus-sdk-java

    0.6.0
    
        
            log4j-slf4j-impl
            org.apache.logging.log4j
        
    

坑3：NoClassDefFoundError: com/google/protobuf/GeneratedMessageV3

milevus-sdk中的protobuf依赖没生效，需要在当前项目中显示添加protobuf依赖。（注意版本要一致）

    com.google.protobuf
    protobuf-java-util
    3.11.0

坑4：NoSuchMethodError:com.google.common.base.Preconditions.checkArgument(ZLjava/lang/String;J)V

原因是当前guava版本过低。

升级：

    com.google.guava
    guava

    23.6-jre

坑5：查询较慢
1、可能是milevus服务端参数cache配置过小，（cache_config.cpu_cache_capacity），此配置默认是4G，如果存的向量容量大于4G，就会很慢。
2、可能是没有显示手动创建索引

使用case

先连接milevus，再创建collection，并在collection里创建index，之后根据需要在collection里创建分区，最后往collection里对应的分区里插数据。
连接milevus

public static MilvusClient getMilevusClient(String host, int port) {

        logger.info("prepared to getMilevusClient...");
        if (milvusClient == null || !milvusClient.isConnected()) {

            milvusClient = new MilvusGrpcClient();
            ConnectParam connectParam = new ConnectParam.Builder()
                    .withHost(host)
                    .withPort(port)
                    .withConnectTimeout(10, TimeUnit.SECONDS)
                    .build();
            logger.info("getMilevusClient milvusClient: {}, connected: {}", milvusClient, milvusClient.isConnected());
            try {
                Response connectResponse = milvusClient.connect(connectParam);
            } catch (ConnectFailedException e) {
                System.out.println("Failed to connect to Milvus server: " + e.toString());
                logger.error("Failed to connect to Milvus server: {}", e);
            }

            logger.info("getMilevusClient milevus isConnected: {}", milvusClient.isConnected());
            return milvusClient.isConnected() ? milvusClient : null;
        }

        return milvusClient;
    }

检查collection并创建
现在milevus不支持字符串索引，所以需要将业务端id与milevusid的id映射关系存下来，这里选型MongoDB。

try {
    // check partiton exist
    HasPartitionResponse isPartitionExist = milvusClient.hasPartition(collectionName, partitionName);
    if (isPartitionExist.ok() && !isPartitionExist.hasPartition()) {
        // create partition
        milvusClient.createPartition(collectionName, partitionName);
    }

    // batch insert id mapping to mongo
    MongoUtils.batchInsertDocList(mongoClient, ANN_DATABASE_NAME, ANN_COLLECTION_NAME, documentList);
    logger.info("success to insert to mongo, documentList size is: {}", documentList.size());
    System.out.println("batch inserting to mongo...");

    // batch insert to milvus
    InsertParam insertParam = new InsertParam
            .Builder(collectionName)
            .withPartitionTag(partitionName)
            .withFloatVectors(batchVectorList)
            .withVectorIds(batchMilevusIdList)
            .build();
    logger.info("batchVectorList size: {}", batchVectorList.size());
    logger.info("batchMilevusIdList size: {}", batchMilevusIdList.size());
    ListenableFuture<InsertResponse> insertResponse = milvusClient.insertAsync(insertParam);
    
} catch (Exception e) {
    logger.error("batch insert occur error, exp: {}", e);
    System.out.println(Arrays.toString(e.getStackTrace()));
}

注意
如果选择内积作为衡量向量相似度的指标（即IP），需要在向量导入前做归一化。

 static List<Float> normalizeVector(List<Float> vector) {
     float squareSum = vector.stream().map(x -> x * x).reduce((float) 0, Float::sum);
     final float norm = (float) Math.sqrt(squareSum);
     vector = vector.stream().map(x -> x / norm).collect(Collectors.toList());
     return vector;
 }