Design patterns for implementing Hive Metastore for Amazon EMR on EKS

In fashionable information architectures, the necessity to handle and question huge datasets effectively, constantly, and precisely is paramount. For organizations that take care of massive information processing, managing metadata turns into a vital concern. That is the place Hive Metastore (HMS) can function a central metadata retailer, taking part in an important function in these fashionable information architectures.

HMS is a central repository of metadata for Apache Hive tables and different information lake desk codecs (for instance, Apache Iceberg), offering shoppers (comparable to Apache Hive, Apache Spark, and Trino) entry to this data utilizing the Metastore Service API. Over time, HMS has grow to be a foundational element for information lakes, integrating with a various ecosystem of open supply and proprietary instruments.

In non-containerized environments, there was sometimes just one method to implementing HMS—operating it as a service in an Apache Hadoop cluster. With the appearance of containerization in information lakes by means of applied sciences comparable to Docker and Kubernetes, a number of choices for implementing HMS have emerged. These choices provide higher flexibility, permitting organizations to tailor HMS deployment to their particular wants and infrastructure.

On this publish, we are going to discover the structure patterns and exhibit their implementation utilizing Amazon EMR on EKS with Spark Operator job submission sort, guiding you thru the complexities that can assist you select the perfect method in your use case.

Answer overview

Previous to Hive 3.0, HMS was tightly built-in with Hive and different Hadoop ecosystem parts. Hive 3.0 launched a Standalone Hive Metastore. This new model of HMS features as an impartial service, decoupled from different Hive and Hadoop parts comparable to HiveServer2. This separation allows varied functions, comparable to Apache Spark, to work together straight with HMS with out requiring a full Hive and Hadoop atmosphere set up. You may study extra about different parts of Apache Hive on the Design web page.

On this publish, we are going to use a Standalone Hive Metastore for example the structure and implementation particulars of varied design patterns. Any reference to HMS refers to a Standalone Hive Metastore.

The HMS broadly consists of two essential parts:

• Backend database: The database is a persistent information retailer that holds all of the metadata, comparable to desk schemas, partitions, and information places.
• Metastore service API: The Metastore service API is a stateless service that manages the core performance of the HMS. It handles learn and write operations to the backend database.

Containerization and Kubernetes gives varied structure and implementation choices for HMS, together with, operating:

On this publish, we’ll use Apache Spark as the info processing framework to exhibit these three architectural patterns. Nevertheless, these patterns aren’t restricted to Spark and might be utilized to any information processing framework, comparable to Hive or Trino, that depends on HMS for managing metadata and accessing catalog data.

Be aware that in a Spark utility, the driving force is accountable for querying the metastore to fetch desk schemas and places, then distributes this data to the executors. Executors course of the info utilizing the places offered by the driving force, by no means needing to question the metastore straight. Therefore, within the three patterns described within the following sections, solely the driving force communicates with the HMS, not the executors.

HMS as sidecar container

On this sample, HMS runs as a sidecar container throughout the identical pod as the info processing framework, comparable to Apache Spark. This method makes use of Kubernetes multi-container pod performance, permitting each HMS and the info processing framework to function collectively in the identical pod. The next determine illustrates this structure, the place the HMS container is a part of Spark driver pod.

This sample is suited to small-scale deployments the place simplicity is the precedence. As a result of HMS is co-located with the Spark driver, it reduces community overhead and offers a simple setup. Nevertheless, it’s vital to notice that on this method HMS operates completely throughout the scope of the father or mother utility and isn’t accessible by different functions. Moreover, row conflicts may come up when a number of jobs try to insert information into the identical desk concurrently. To handle this, you need to ensure that no two jobs are writing to the identical desk concurrently.

Contemplate this method if you happen to desire a primary structure. It’s preferrred for organizations the place a single workforce manages each the info processing framework (for instance, Apache Spark) and HMS, and there’s no want for different functions to make use of HMS.

Cluster devoted HMS

On this sample, HMS runs in a number of pods managed by means of a Kubernetes deployment, sometimes inside a devoted namespace in the identical information processing EKS cluster. The next determine illustrates this setup, with HMS decoupled from Spark driver pods and different workloads.

This sample works properly for medium-scale deployments the place reasonable isolation is sufficient, and compute and information wants might be dealt with inside a couple of clusters. It offers a steadiness between useful resource effectivity and isolation, making it preferrred to be used instances the place scaling metadata providers independently is vital, however full decoupling isn’t needed. Moreover, this sample works properly when a single workforce manages each the info processing frameworks and HMS, making certain streamlined operations and alignment with organizational obligations.

By decoupling HMS from Spark driver pods, it could actually serve a number of shoppers, comparable to Apache Spark and Trino, whereas sharing cluster assets. Nevertheless, this method may result in useful resource competition in periods of excessive demand, which might be mitigated by imposing tenant isolation on HMS pods.

Exterior HMS

On this structure sample, HMS is deployed in its personal EKS cluster deployed utilizing Kubernetes deployment and uncovered as a Kubernetes Service utilizing AWS Load Balancer Controller, separate from the info processing clusters. The next determine illustrates this setup, the place HMS is configured as an exterior service, separate from the info processing clusters.

This sample fits situations the place you need a centralized metastore service shared throughout a number of information processing clusters. HMS permits completely different information groups to handle their very own information processing clusters whereas counting on the shared metastore for metadata administration. By deploying HMS in a devoted EKS cluster, this sample offers most isolation, impartial scaling, and the pliability to function and managed as its personal impartial service.

Whereas this method gives clear separation of issues and the flexibility to scale independently, it additionally introduces larger operational complexity and doubtlessly elevated prices due to the necessity to handle an extra cluster. Contemplate this sample you probably have strict compliance necessities, want to make sure full isolation for metadata providers, or need to present a unified metadata catalog service for a number of information groups. It really works properly in organizations the place completely different groups handle their very own information processing frameworks and depend on a shared metadata retailer for information processing wants. Moreover, the separation allows specialised groups to give attention to their respective areas.

Deploy the answer

Within the the rest of this publish, you’ll discover the implementation particulars for every of the three structure patterns, utilizing EMR on EKS with Spark Operator job submission sort for example to exhibit their implementation. Be aware that this implementation hasn’t been examined with different EMR on EKS Spark job submission sorts. You’ll start by deploying the frequent parts that function the inspiration for all of the structure patterns. Subsequent, you’ll deploy the parts particular to every sample. Lastly, you’ll execute Spark jobs to hook up with the HMS implementation distinctive to every sample and confirm the profitable execution and retrieval of information and metadata.

To streamline the setup course of, we’ve automated the deployment of frequent infrastructure parts so you possibly can give attention to the important elements of every HMS structure. We’ll present detailed data that can assist you perceive every step, simplifying the setup whereas preserving the training expertise.

State of affairs

To showcase the patterns, you’ll create three clusters:

• Two EMR on EKS clusters: analytics-cluster and datascience-cluster
• An EKS cluster: hivemetastore-cluster

Each analytics-cluster and datascience-cluster function information processing clusters that run Spark workloads, whereas the hivemetastore-cluster hosts the HMS.

You’ll use analytics-cluster for example the HMS as sidecar and cluster devoted sample. You’ll use all three clusters to exhibit the exterior HMS sample.

Supply code

You could find the codebase within the AWS Samples GitHub repository.

Stipulations

Earlier than you deploy this answer, ensure that the next conditions are in place:

Arrange frequent infrastructure

Start by organising the infrastructure parts which are frequent to all three architectures.

1. Clone the repository to your native machine and set the 2 atmosphere variables. Exchange <AWS_REGION> with the AWS Area the place you need to deploy these assets.

git clone https://github.com/aws-samples/sample-emr-eks-hive-metastore-patterns.git
cd sample-emr-eks-hive-metastore-patterns
export REPO_DIR=$(pwd)
export AWS_REGION=<AWS_REGION>

2. Execute the next script to create the shared infrastructure.

cd ${REPO_DIR}/setup
./setup.sh

3. To confirm profitable infrastructure deployment, navigate to the AWS Administration Console for AWS CloudFormation, choose your stack, and verify the Occasions, Assets, and Outputs tabs for completion standing, particulars, and record of assets created.

You could have accomplished the setup of the frequent parts that function the inspiration for all architectures. You’ll now deploy the parts particular to every structure and execute Apache Spark jobs to validate the profitable implementation.

HMS in a sidecar container

To implement HMS utilizing the sidecar container sample, the Spark utility requires setting each sidecar and catalog properties within the job configuration file.

1. Execute the next script to configure the analytics-cluster for sidecar sample. For this publish, we saved the HMS database credentials right into a Kubernetes Secret object. We advocate utilizing Kubernetes Exterior Secrets and techniques Operator to fetch HMS database credentials from AWS Secrets and techniques Supervisor.

cd ${REPO_DIR}/hms-sidecar
./configure-hms-sidecar.sh analytics-cluster

2. Assessment the Spark job manifest file spark-hms-sidecar-job.yaml. This file was created by substituting variables within the spark-hms-sidecar-job.tpl template within the earlier step. The next samples spotlight key sections of the manifest file.

spec:
  driver:
  ...
    sidecars:
      # Hive Metastore Sidecar container
      - identify: hive-metastore
        picture: ${AWS_ACCOUNT_ID}.dkr.ecr.${AWS_REGION}.amazonaws.com/hive-metastore:newest
        env:
          # These settings configure metastore to make use of Amazon Postgres RDS as backend database, connecting to it through jdbc URL.
          - identify: HIVE_METASTORE_DB_TYPE
            worth: postgres
          - identify: HIVE_METASTORE_DB_DRIVER
            worth: org.postgresql.Driver
          - identify: HIVE_METASTORE_DB_URL
            worth: jdbc:postgresql://${HMS_RDS_PROXY_ENDPOINT}:5432/hivemetastore
          # The warehouse location is specified as an S3 bucket
          - identify: HIVE_METASTORE_WAREHOUSE_LOC
            worth: s3a://${S3_BUCKET_NAME}/warehouse
          - identify: AWS_REGION
            worth: ${AWS_REGION}
          # The database username and password are handed through atmosphere variables. The password is retrieved from a Kubernetes secret
          - identify: HIVE_METASTORE_DB_USER
            worth: hive_metastore_user
          - identify: HIVE_METASTORE_DB_PASSWORD
            valueFrom:
              secretKeyRef:
                identify: hms-rds-password
                key: HIVE_METASTORE_DB_PASSWORD

Spark job configuration

spec:  
  sparkConf:
    # Hive Catalog properties
    # Units spark to make use of Hive because the catalog for SQL operations
    spark.sql.catalogImplementation: "hive"
    # HMS is uncovered on localhost:8080 because the sidecar runs in the identical pod. Spark connects to the sidecar through this URI
    spark.hadoop.hive.metastore.uris: "thrift://localhost:9083"
    # The info location is ready to S3 bucket
    spark.sql.warehouse.dir: "s3a://${S3_BUCKET_NAME}/warehouse"
    # Retrieve non permanent AWS credentials by assuming a job utilizing an internet identification token
    spark.hadoop.fs.s3a.aws.credentials.supplier: "com.amazonaws.auth.WebIdentityTokenCredentialsProvider"
    # Configure Spark to make use of S3A filesystem
    spark.hadoop.fs.s3a.impl: "org.apache.hadoop.fs.s3a.S3AFileSystem"

Submit the Spark job and confirm the HMS as sidecar container setup

On this sample, you’ll submit Spark jobs in analytics-cluster. The Spark jobs will hook up with the HMS service operating as a sidecar container within the driver pod.

1. Run the Spark job to confirm that the setup was profitable.

kubectl apply -f spark-hms-sidecar-job.yaml

2. Describe the sparkapplication object.

kubectl get sparkapplication -n emr
kubectl describe sparkapplication spark-hms-sidecar-job --namespace emr

3. Listing the pods and observe the variety of containers connected to the driving force pod. Wait till the Standing modifications from ContainerCreating to Working (ought to take just some seconds).

4. View the driving force logs to validate the output.

kubectl logs spark-hms-sidecar-job-driver --namespace emr

5. Should you encounter the next error, await a couple of minutes and rerun the earlier command.

Error from server (BadRequest): container "spark-kubernetes-driver" in pod "spark-hms-sidecar-driver" is ready to start out: ContainerCreating

6. After profitable completion of the job, you see the next message within the logs. The tabular output efficiently validates the setup of HMS as a sidecar container.

...
24/09/17 21:44:00 INFO metastore: Making an attempt to hook up with metastore with URI thrift://localhost:9083
24/09/17 21:44:00 INFO metastore: Opened a connection to metastore, present connections: 1
24/09/17 21:44:00 INFO metastore: Related to metastore.
...
...
+-------+---+-------+---+
|sample|id |identify   |age|
+-------+---+-------+---+
|Sidecar|1 |Alice   |30 |
|Sidecar|2 |Bob     |25 |
|Sidecar|3 |Charlie |35 |
+-------+---+-------+---+

Cluster devoted HMS

To implement HMS utilizing a cluster devoted HMS sample, the Spark utility requires organising HMS URI and catalog properties within the job configuration file.

1. Execute the next script to configure the analytics-cluster for cluster devoted sample.

cd ${REPO_DIR}/hms-cluster-dedicated
./configure-hms-cluster-dedicated.sh analytics-cluster

2. Confirm the HMS deployment by itemizing the pods and viewing the logs. No Java exceptions within the logs confirms that the Hive Metastore service is operating efficiently.

kubectl get pods --namespace hive-metastore
kubectl logs <HMS-PODNAME> --namespace hive-metastore

3. Assessment the Spark job manifest file, spark-hms-cluster-dedicated-job.yaml. This file is created by substituting variables within the spark-hms-cluster-dedicated-job.tpl template within the earlier step. The next pattern highlights key sections of the manifest file.

spec: 
  sparkConf:
    # Hive Catalog properties
    # Units spark to make use of Hive because the catalog for SQL operations
    spark.sql.catalogImplementation: "hive"
    # HMS is operating in a pod and we are able to hook up with it in the identical EKS cluster through this URI
    spark.hadoop.hive.metastore.uris: "thrift://hive-metastore-svc.hive-metastore.svc.cluster.native:9083"
    # The info location is ready to S3 bucket
    spark.sql.warehouse.dir: "s3a://${S3_BUCKET_NAME}/warehouse"
    # Retrieve non permanent AWS credentials by assuming a job utilizing an internet identification token
    spark.hadoop.fs.s3a.aws.credentials.supplier: "com.amazonaws.auth.WebIdentityTokenCredentialsProvider"
    # Configure Spark to make use of S3A filesystem
    spark.hadoop.fs.s3a.impl: "org.apache.hadoop.fs.s3a.S3AFileSystem"

Submit the Spark job and confirm the cluster devoted HMS setup

On this sample, you’ll submit Spark jobs in analytics-cluster. The Spark jobs will hook up with the HMS service in the identical information processing EKS cluster.

1. Submit the job.

kubectl apply -f spark-hms-cluster-dedicated-job.yaml -n emr

2. Confirm the standing.

kubectl get sparkapplication -n emr
kubectl describe sparkapplication spark-hms-cluster-dedicated-job --namespace emr

3. Describe driver pod and observe the variety of containers connected to the driving force pod. Wait till the standing modifications from ContainerCreating to Working (ought to take just some seconds).

4. View the driving force logs to validate the output.

kubectl logs spark-hms-cluster-dedicated-job-driver --namespace emr

5. After profitable completion of the job, you need to see the next message within the logs. The tabular output efficiently validates the setup of cluster devoted HMS.

...
24/09/17 21:44:00 INFO metastore: Making an attempt to hook up with metastore with URI thrift://hive-metastore-svc.hive-metastore.svc.cluster.native:9083
24/09/17 21:44:00 INFO metastore: Opened a connection to metastore, present connections: 1
24/09/17 21:44:00 INFO metastore: Related to metastore.
...
...
+-----------------+---+-------+---+
|sample          |id |identify   |age|
+-----------------+---+-------+---+
|Cluster Devoted|1  |Alice  |30 |
|Cluster Devoted|2  |Bob    |25 |
|Cluster Devoted|3  |Charlie|35 |
+-----------------+---+-------+---+

Exterior HMS

To implement an exterior HMS sample, the Spark utility requires organising an HMS URI for the service endpoint uncovered by hivemetastore-cluster.

1. Execute the next script to configure hivemetastore-cluster for Exterior HMS sample.

cd ${REPO_DIR}/hms-external
./configure-hms-external.sh

2. Assessment the Spark job manifest file spark-hms-external-job.yaml. This file is created by substituting variables within the spark-hms-external-job.tpl template throughout the setup course of. The next pattern highlights key sections of the manifest file.

spec:
  sparkConf:
    # Hive Catalog properties
    # Units spark to make use of Hive because the catalog for SQL operations
    spark.sql.catalogImplementation: "hive"
    # HMS is operating in a cluster and we are able to hook up with it within the EKS cluster through this URI
    spark.hadoop.hive.metastore.uris: "thrift://${HMS_URI_ENDPOINT}:9083"
    # The info location is ready to S3 bucket
    spark.sql.warehouse.dir: "s3a://${S3_BUCKET_NAME}/warehouse"
    # Retrieve non permanent AWS credentials by assuming a job utilizing an internet identification token
    spark.hadoop.fs.s3a.aws.credentials.supplier: "com.amazonaws.auth.WebIdentityTokenCredentialsProvider"
    # Configure Spark to make use of S3A filesystem
    spark.hadoop.fs.s3a.impl: "org.apache.hadoop.fs.s3a.S3AFileSystem"

Submit the Spark job and confirm the HMS in a separate EKS cluster setup

To confirm the setup, submit Spark jobs in analytics-cluster and datascience-cluster. The Spark jobs will hook up with the HMS service within the hivemetastore-cluster.

Use the next steps for analytics-cluster after which for datascience-cluster to confirm that each clusters can hook up with the HMS on hivemetastore-cluster.

1. Run the spark job to check the profitable setup. Exchange <CONTEXT_NAME> with Kubernetes context for analytics-cluster after which for datascience-cluster.

kubectl config use-context <CONTEXT_NAME>
kubectl apply -f spark-hms-external-job.yaml -n emr

2. Describe the sparkapplication object.

kubectl get sparkapplication spark-hms-external-job -n emr
kubectl describe sparkapplication spark-hms-external-job --namespace emr

3. Listing the pods and observe the variety of containers connected to the driving force pod. Wait till the standing modifications from ContainerCreating to Working (ought to take just some seconds).

5. View the driving force logs to validate the output on the info processing cluster.

kubectl logs spark-hms-external-job-driver --namespace emr

6. The output ought to seem like the next. The tabular output efficiently validates the setup of HMS in a separate EKS cluster.

After profitable completion of the job, you need to be capable of see the beneath message within the logs.
...
24/09/17 21:44:00 INFO metastore: Making an attempt to hook up with metastore with URI  thrift://k8s-hivemeta-hmsexter-xxxxxx.elb.us-east-1.amazonaws.com:9083
24/09/17 21:44:00 INFO metastore: Opened a connection to metastore, present connections: 1
24/09/17 21:44:00 INFO metastore: Related to metastore.
...
...
+--------+---+-------+---+
|sample |id |identify   |age|
+--------+---+-------+---+
|Exterior|1  |Alice  |30 |
|Exterior|2  |Bob    |25 |
|Exterior|3  |Charlie|35 |
+--------+---+-------+---+

Clear up

To keep away from incurring future fees from the assets created on this tutorial, clear up your atmosphere after you’ve accomplished the steps. You are able to do this by operating the cleanup.sh script, which can safely take away all of the assets provisioned throughout the setup.

cd ${REPO_DIR}/setup
./cleanup.sh

Conclusion

On this publish, we’ve explored the design patterns for implementing the Hive Metastore (HMS) with EMR on EKS with Spark Operator, every providing distinct benefits relying in your necessities. Whether or not you select to deploy HMS as a sidecar container throughout the Apache Spark Driver pod, or as a Kubernetes deployment within the information processing EKS cluster, or as an exterior HMS service in a separate EKS cluster, the important thing concerns revolve round communication effectivity, scalability, useful resource isolation, excessive availability, and safety.

We encourage you to experiment with these patterns in your individual setups, adapting them to suit your distinctive workloads and operational wants. By understanding and making use of these design patterns, you possibly can optimize your Hive Metastore deployments for efficiency, scalability, and safety in your EMR on EKS environments. Discover additional by deploying the answer in your AWS account and share your experiences and insights with the neighborhood.

In regards to the Authors

Avinash Desireddy is a Cloud Infrastructure Architect at AWS, obsessed with constructing safe functions and information platforms. He has intensive expertise in Kubernetes, DevOps, and enterprise structure, serving to clients containerize functions, streamline deployments, and optimize cloud-native environments.

Suvojit Dasgupta is a Principal Knowledge Architect at AWS. He leads a workforce of expert engineers in designing and constructing scalable information options for AWS clients. He makes a speciality of growing and implementing modern information architectures to deal with advanced enterprise challenges.