[Subtask][DISCUSS] Docker image tag options

ulysses-you commented 1 year ago

Code of Conduct

[X] I agree to follow this project's Code of Conduct

Search before asking

[X] I have searched in the issues and found no similar issues.

Describe the subtask

Kyuubi supports multi-engines, e.g. Spark, Flink, Hive... So there is a chance to provide some custom image tags since we are using a separate repo.

An idea of tag options: version	tag	derived tags
1.6.0 Kyuubi only	1.6.0-scala_2.12-java8-ubuntu	1.6.0, latest
1.6.0 with Spark default version binary tar	1.6.0-spark3.3-scala_2.12-java8-ubuntu	1.6.0-spark3.3, 1.6.0-spark
1.6.0 with Flink default version binary tar	1.6.0-flink1.5-scala_2.12-java8-ubuntu	1.6.0-flink3.3, 1.6.0-flink
1.6.0 with all default version binary tar	1.6.0-all-scala_2.12-java8-ubuntu	1.6.0-all

Parent issue

https://github.com/apache/kyuubi/issues/4158

Are you willing to submit PR?

[ ] Yes. I would be willing to submit a PR with guidance from the Kyuubi community to improve.
[ ] No. I cannot submit a PR at this time.

ulysses-you commented 1 year ago

This issue is used to disscuss what image tag should be pushed during each release. cc @yaooqinn @pan3793 @zwangsheng @hddong @yanghua @turboFei @cfmcgrady @cxzl25 and all. Any thought is welcome !

zwangsheng commented 1 year ago

+1 for provided basic, spark, flink and all images.

For tag, maybe we should omit scala_2.12-java8-ubuntu as not provided another options.

We can indicate those versions in the description, so that users can perceive.

ulysses-you commented 1 year ago

yea, scala_2.12-java8-ubuntu is more like a style of DOI. For apache/kyuubi repo, we can use the derived tag

bowenliang123 commented 1 year ago

As for the choice of operation system, I would prefer options in either alpine or centos. Considering the maturity in industry production and package management,

ubuntu is less used as base docker image in production
centos is more common in production in practise , but it's no longer maintained from official
alpine is popular and practically used as small-size Linux base images in Docker, generally used from app to service deployment. plus, the apk is easy to use and includes most common package

bowenliang123 commented 1 year ago

And as Kyuubi contains extensions for engine, consider release an all-in-one tag version shipping the possible available extensions and configs for best practice or proficiency as well. e.g. including,

Auxiliary Optimization with enabling configs (https://kyuubi.readthedocs.io/en/master/extensions/engines/spark/rules.html)
Authz Plugin for Spark without enabling (https://kyuubi.readthedocs.io/en/master/security/authorization/spark/index.html)
SQL Lineage Support without enabling (https://kyuubi.readthedocs.io/en/master/extensions/engines/spark/lineage.html)
Hive Dialect Support without enabling (https://kyuubi.readthedocs.io/en/master/extensions/engines/spark/jdbc-dialect.html)

bowenliang123 commented 1 year ago

As for the choice of operation system type of base images, I would prefer alpine or centos. From a prospective of production maturity or package management

ubuntu is less used in server or containers deployment in production, for Ops means different tool set and packages
centos is more popular in the pass , but good enough for production . lacking of further official maintainance.
alpine is one of most popular base images with minimized image size while providing comprehensive features. OpenJDK also ships official version with it in new versions although alpine uses glibc. apk in alpine is easy to use and covering most common packages.

bowenliang123 commented 1 year ago

Also, to maximize both performance, accessibility, and potentiality in best practices and potentiality out of box, consider release an fine-tuned all-in-one image for each engine with engine/server plugins and suggested configs of by default. eg.

[engine connector] Spark Hive Connector (https://kyuubi.readthedocs.io/en/master/connector/spark/hive.html?highlight=hive%20connector#hive-connector-integration)
[engine plugin] Auxiliary Optimization, enabled by default (https://kyuubi.readthedocs.io/en/master/extensions/engines/spark/rules.html)
[engine plugin] Authz Plugin for Ranger, disabled by default
[engine plugin] SQL Lineage Support, disabled by default
[engine plugin] Hive Dialect Support, disabled by default (https://kyuubi.readthedocs.io/en/master/extensions/engines/spark/jdbc-dialect.html) ...

turboFei commented 1 year ago

As for the choice of operation system, I would prefer options in either alpine or centos.

Ubuntu looks fine for me, all the hadoop images in ebay are based on ubuntu or migrating to ubuntu.

-1 for centos.

turboFei commented 1 year ago

spark official docker image is also using ubuntu, https://github.com/apache/spark-docker/tree/master/3.3.1

bowenliang123 commented 1 year ago

centos is just an option for discussion as many companies already have local proxy for packages , I am good with both ubuntu and alpine. But considering in 2 facts,

ubuntu is good for software builds with development tools, but does not contribute too in runtime. For container deployment , size matters. The compressed size of distribution will be up to 1G (and 1.5G decompressed), as 400MB OS lib with ubuntu + 250MB JDK + 250 MB Kyuubi with extensions. As with alpine, the whole size can be shrunk down to 600-700MB in total and 1G in decompressed.
never heard of using ubuntu in production either for server deployment or container images in the infrastructure industry (e.g. finance, power grid and etc.).
from base image adoption rate, even debian has a greater share than ubuntu, and ubuntu is also much less than alpine

I am quite sure the community will agree to ship a version based on ubuntu, but it is never the first choice for production for me and other users with similar background or consideration. And that's why I raise the discussion for a second option of OS base lib here.

turboFei commented 1 year ago

never heard of using ubuntu in production either for server deployment or container images in the infrastructure industry (e.g. finance, power grid and etc.).

our old hadoop nodes images were based on centos before, but now, all the hadoop nodes are based on ubuntu or in the migration.

I think If I did not say that, you will not know that we are using ubuntu.

There might be some limitations for my view, because I just know the current status in ebay.

bowenliang123 commented 1 year ago

Yes, thanks for sharing the status in your company. With all the discussions, ubuntu is a balanced choice for the base image. Even for the image size, it is not as extremely small as alpine but already smaller than centos. ubuntu is still the first option for this topic, since it is adopted in Spark's official image (as you just pointed out), but also in Flink's (https://github.com/apache/flink-docker). Meanwhile, it's less attractive for the situation in my case if this's the only option for base os, and it's pushing me away from practice and having more thoughts in reassembling the docker images in private distribution.

ulysses-you commented 1 year ago

@bowenliang123 can you show some projects who ships centOS to image ? That may help us to see the difference.

bowenliang123 commented 1 year ago

centos is no longer supported and we could skip this option for kyuubi. And my concern is for enterprise-level production use,

support yum like package management and rpm installers
make use of the existing mirror proxy or private repo of the company, which is the key to align for Ops in deployment or management for both raw machines or containers.

apache / kyuubi