208354c327
# 🔍 Description ## Issue References 🔗 While enabling batch implementation V2 with the following configurations ``` kyuubi.batch.impl.version=2 kyuubi.batch.submitter.enabled=true kyuubi.batch.submitter.threads=48 spark.master=yarn spark.submit.deployMode=cluster spark.yarn.submit.waitAppCompletion=false ``` I found that the batch jobs will be blocked in the DB queue once a YARN queue has no resources, this brings an issue, the subsequential batch jobs that are going to be submitted to another YARN queue also be queued in DB, rather than YARN queue. ``` mysql> select state, engine_state, count(1) from metadata where state in ('INITIALIZED', 'PENDING', 'RUNNING') group by state, engine_state; +-------------+--------------+----------+ | state | engine_state | count(1) | +-------------+--------------+----------+ | INITIALIZED | NULL | 166 | | PENDING | NULL | 1 | | RUNNING | PENDING | 148 | | RUNNING | RUNNING | 415 | +-------------+--------------+----------+ ``` ## Describe Your Solution 🔧 The submitter queue whose size is controlled by `kyuubi.batch.submitter.threads` is designed to address the `spark-submit` process concurrency issue, too many `spark-submit` processes may run out of the Kyuubi server's node CPU/memory resources and eventually crash the service. For Spark YARN cluster mode, if set `spark.yarn.submit.waitAppCompletion=false`, the local `spark-submit` process exits immediately once the Application goes ACCEPTED status, even no resource could be allocated for the AM container, we should not block such case in submitter queue. ## Types of changes 🔖 - [ ] Bugfix (non-breaking change which fixes an issue) - [ ] New feature (non-breaking change which adds functionality) - [ ] Breaking change (fix or feature that would cause existing functionality to change) ## Test Plan 🧪 Pass GA, and roll out into internal cluster. --- # Checklist 📝 - [x] This patch was not authored or co-authored using [Generative Tooling](https://www.apache.org/legal/generative-tooling.html) **Be nice. Be informative.** Closes #6028 from pan3793/batch-submit. Closes #6028 05fcc758f [Cheng Pan] Exited spark-submit process should not block batch submit queue Authored-by: Cheng Pan <chengpan@apache.org> Signed-off-by: Cheng Pan <chengpan@apache.org> |
||
|---|---|---|
| .github | ||
| .idea | ||
| bin | ||
| build | ||
| charts/kyuubi | ||
| conf | ||
| dev | ||
| docker | ||
| docs | ||
| extensions | ||
| externals | ||
| integration-tests | ||
| kyuubi-assembly | ||
| kyuubi-common | ||
| kyuubi-ctl | ||
| kyuubi-events | ||
| kyuubi-ha | ||
| kyuubi-hive-beeline | ||
| kyuubi-hive-jdbc | ||
| kyuubi-hive-jdbc-shaded | ||
| kyuubi-metrics | ||
| kyuubi-rest-client | ||
| kyuubi-server | ||
| kyuubi-util | ||
| kyuubi-util-scala | ||
| kyuubi-zookeeper | ||
| licenses | ||
| licenses-binary | ||
| .asf.yaml | ||
| .dockerignore | ||
| .gitattributes | ||
| .gitignore | ||
| .rat-excludes | ||
| .readthedocs.yaml | ||
| .scalafmt.conf | ||
| codecov.yml | ||
| CONTRIBUTING.md | ||
| LICENSE | ||
| LICENSE-binary | ||
| NOTICE | ||
| NOTICE-binary | ||
| pom.xml | ||
| README.md | ||
| scalastyle-config.xml | ||
Project - Documentation - Who's using
Apache Kyuubi
Apache Kyuubi™ is a distributed and multi-tenant gateway to provide serverless SQL on data warehouses and lakehouses.
What is Kyuubi?
Kyuubi provides a pure SQL gateway through Thrift JDBC/ODBC interface for end-users to manipulate large-scale data with pre-programmed and extensible Spark SQL engines. This "out-of-the-box" model minimizes the barriers and costs for end-users to use Spark at the client side. At the server-side, Kyuubi server and engines' multi-tenant architecture provides the administrators a way to achieve computing resource isolation, data security, high availability, high client concurrency, etc.
- A HiveServer2-like API
- Multi-tenant Spark Support
- Running Spark in a serverless way
Target Users
Kyuubi's goal is to make it easy and efficient for anyone to use Spark(maybe other engines soon) and facilitate users to handle big data like ordinary data. Here, anyone means that users do not need to have a Spark technical background but a human language, SQL only. Sometimes, SQL skills are unnecessary when integrating Kyuubi with Apache Superset, which supports rich visualizations and dashboards.
In typical big data production environments with Kyuubi, there should be system administrators and end-users.
- System administrators: A small group consists of Spark experts responsible for Kyuubi deployment, configuration, and tuning.
- End-users: Focus on business data of their own, not where it stores, how it computes.
Additionally, the Kyuubi community will continuously optimize the whole system with various features, such as History-Based Optimizer, Auto-tuning, Materialized View, SQL Dialects, Functions, etc.
Usage scenarios
Port workloads from HiveServer2 to Spark SQL
In typical big data production environments, especially secured ones, all bundled services manage access control lists to restricting access to authorized users. For example, Hadoop YARN divides compute resources into queues. With Queue ACLs, it can identify and control which users/groups can take actions on particular queues. Similarly, HDFS ACLs control access of HDFS files by providing a way to set different permissions for specific users/groups.
Apache Spark is a unified analytics engine for large-scale data processing. It provides a Distributed SQL Engine, a.k.a, the Spark Thrift Server(STS), designed to be seamlessly compatible with HiveServer2 and get even better performance.
HiveServer2 can identify and authenticate a caller, and then if the caller also has permissions for the YARN queue and HDFS files, it succeeds. Otherwise, it fails. However, on the one hand, STS is a single Spark application. The user and queue to which STS belongs are uniquely determined at startup. Consequently, STS cannot leverage cluster managers such as YARN and Kubernetes for resource isolation and sharing or control the access for callers by the single user inside the whole system. On the other hand, the Thrift Server is coupled in the Spark driver's JVM process. This coupled architecture puts a high risk on server stability and makes it unable to handle high client concurrency or apply high availability such as load balancing as it is stateful.
Kyuubi extends the use of STS in a multi-tenant model based on a unified interface and relies on the concept of multi-tenancy to interact with cluster managers to finally gain the ability of resources sharing/isolation and data security. The loosely coupled architecture of the Kyuubi server and engine dramatically improves the client concurrency and service stability of the service itself.
DataLake/Lakehouse Support
The vision of Kyuubi is to unify the portal and become an easy-to-use data lake management platform. Different kinds of workloads, such as ETL processing and BI analytics, can be supported by one platform, using one copy of data, with one SQL interface.
- Logical View support via Kyuubi DataLake Metadata APIs
- Multiple Catalogs support
- SQL Standard Authorization support for DataLake(coming)
Cloud Native Support
Kyuubi can deploy its engines on different kinds of Cluster Managers, such as, Hadoop YARN, Kubernetes, etc.
The Kyuubi Ecosystem(present and future)
The figure below shows our vision for the Kyuubi Ecosystem. Some of them have been realized, some in development, and others would not be possible without your help.
Online Documentation 
Quick Start
Ready? Getting Started with Kyuubi.
Contributing
Project & Community Status
Aside
The project took its name from a character of a popular Japanese manga - Naruto.
The character is named Kyuubi Kitsune/Kurama, which is a nine-tailed fox in mythology.
Kyuubi spread the power and spirit of fire, which is used here to represent the powerful Apache Spark.
Its nine tails stand for end-to-end multi-tenancy support of this project.