Marquez Monthly Community Meeting

The Marquez Community Meeting occurs on the fourth Thursday of each month. Meetings are held on Zoom.

August 25, 2022

July 28, 2022

Attendees:

• TSC:
  • Willy Lulciuc, Co-creator of Marquez
  • Michael Collado, Staff Software Engineer, Astronomer
• And:
  • Michael Robinson, Software Engineer, Dev. Rel., Astronomer
  • Minkyu Park, Senior Engineer, Astronomer
  • John Thomas, Software Engineer, Dev. Rel., Astronomer
  • Ross Turk, Senior Director of Community, Astronomer
  • Ryan Hatter, Customer Reliability Engineer, Astronomer
  • Howard Yoo, Staff Product Manager, Astronomer

Agenda:

1. Announcements
2. Introducing the Marquez blog
3. Architecture review: the lineage graph
4. Discussion
   1. Marquez issue #2048

Meeting:

Notes:

1. Announcements [Willy]
2. Introducing the Marquez Blog [Michael R. and Ross]
   1. new blog can be found at marquezproject.ai/blog
   2. designed and built by Ross
   3. to contribute a blog post on GitHub:
      1. write post in Markdown, place it in new directory in OpenLineage/website/contents/blog 
      2. OR: open an issue first to suggest a topic or get feedback on your idea
      3. artwork: Ross happy to make the images; tag him
      4. Ross also happy to document the artwork creation process for others
3. Architecture review: the lineage graph [Willy]
   1. What is Marquez doing in the background to surface lineage metadata at the run level during execution?
   2. What is a current lineage graph?
      1. bigraph with nodes for jobs and datasets
      2. run-level lineage is collected from OpenLineage events
      3. representation of job is based on datasets and the inputs and outputs they produce
      4. datasets stitched together using OpenLineage `ID` (global and unique)
      5. versioning of jobs enabled by OpenLineage `JobVersion`
         1. Marquez keeps track of changes to code and datasets behind the scenes
   3. Marquez data model
      1. Marquez keeps track of:
         1. job versions
         2. runs of each version
         3. sources
      2. each node represents the latest, or current, version of the job's lineage
      3. `Job` is `ID` and arrays representing input and output datasets
   4. Demo
      1. UI defaults to latest/current graph
      2. prior versions accessible via `version history` tab
      3. selecting a version makes another job node/datasets visible
      4. makes "time travel" possible in your pipeline
      5. all of this possible thanks to the OpenLineage spec
   5. Q & A
      1. If a job has not completed, will you not see metadata? [Howard]
         • no – a job has to complete in order for versioning logic to be applied 
      2. Is a job version associated with the code that produced it? [Ryan]
         • yes – if the code is provided as a source location facet
         • Marquez will determine if the code has changed
         • changes to schema also monitored using dataset versioning; this tied to job version
4. Discussion
   1. Howard: issue 2048: 
      1. There is an edge case (using a custom extractor) where the TaskMetadata's given input or output dataset would NOT have the fields populated (`dataset.fields = []`).
      2. Having this type of metadata makes Marquez overwrite the existing version of the dataset with empty fields
      3. Proposal: Marquez should try to reuse the dataset instead of rewriting 
   2. Agreed; question remains about how to do it [Willy]
      1. behavior reflects versioning logic
      2. possible solution: use `null` value in OL spec rather than empty array
      3. challenge: we want to avoid making assumptions

June 23, 2022

Attendees:

• TSC
  • Willy Lulciuc, Co-creator of Marquez
  • Julien Le Dem, Chief Architect, Astronomer
• And
  • Martin Fiser, Head of Professional Services, Keboola
  • Michael Robinson, Software Engineer, Dev. Rel., Astronomer
  • Minkyu Park, Senior Engineer, Astronomer
  • John Thomas, Support Engineer, Astronomer
  • Naga Raghavarapu, Principal Software Engineer, Oracle
  • Ross Turk, Senior Director of Community, Astronomer

Agenda:

• Announcements
• Recent release: 0.23.0
• User story by Martin Fiser (Keboola)
• Open discussion

Meeting:

Notes:

• Announcements [Willy]

  • Mqz/OL swag is still available!
  • Willy talked Mqz at OS Summit (LinuxCon)

• Recent Release 0.23.0 [Michael R.]

  • Added
    • Update docker-compose.yml: Randomly map postgres db port (#2000, @RNHTTR)
    • Job parent hierarchy (#1935 #1980 #1992, @collado-mike)
  • Changed
    • Set default limit for listing datasets and jobs in UI from 2000 to 25 (#2018, @wslulciuc)
  • Fixed
    • Return the tag for postgresql to 12.1.0 (#2015, @rossturk)

• Keboola Use Case [Martin]

  • Topic: OL integration with the Keboola platform
    • Overview of platform
      • modern data experience: data stack as a service
      • all-in-one service
      • writers/reverse ETL through component framework
      • enables version control, governance, etc., in workspaces
      • much metadata produced and collected, permitting visibility across entire pipeline
        • pipeline jobs
        • storage events
        • data loads/unloads
        • user-generated metadata
    • Purpose of OL integration
      • data governance to support users' feeding data to external tools
      • OL a "language" for speaking to various tools
      • offer API for OL information
      • native Keboola component
        • feeds OL information to an endpoint (e.g., Marquez)
        • can be orchestrated on customizable interval 
        • supports SSH
        • exports full job information to the endpoint
    • Demo
      • users have multiple projects on the platform
      • a few hundred components are offered to users out of the box (e.g., Google Drive, SQL, Python, Google Sheets)
      • metadata manually pushable to OpenLineage endpoint
      • orchestrator could benefit from parent/job support
    • Challenges
      • need: richer metadata 
        • component config
        • info about tables
      • lighter UI
        • reflects feedback about legibility
        • icon customizability
      • namespaces
        • connectivity between projects
      • more integrations
      • rounded logo
    • Q & A
      • Are you interested in contributing? [Julien]
        • would like to; possibly in the future
      • Would you like to open issues? (custom facets, UI) [Willy]
        • not currently able to
      • Are you using any integrations? java or python [Willy]
        • component can be anything in the docker container
        • multiple languages used in development
      • Customers using it already? [Conor]
        • some testing is going on
        • not in production yet
        • no plans to offer Marquez to customers
      • Does it work for every connector? [Conor]
        • each will produce at least a job
      • Auth model [Willy]
        • problem: slippery slope [Martin]
        • recommended at ingress level [Willy]
        • not a focus at the moment
        • contributions to related issues welcome
      • Is data discovery offered? [Naga]
        • built in with API 
        • additional tools can be added if integration would be seamless

May 26, 2022

Attendees:

TSC:

• Willy Lulciuc, Co-creator of Marquez
• Peter Hicks, Senior Engineer, Astronomer

And:

• Ross Turk, Senior Director of Community, Astronomer
• Minkyu Park, Senior Engineer, Astronomer
• John Thomas, Support Engineer, Astronomer
• Michael Robinson, Developer Relations Engineer, Astronomer
• Joshua Wankowski, Associate Data Engineer, Northwestern Mutual
• Sam Holmberg, Software Engineer, Astronomer
• Dako Dakov, R&D Manager, VMware
• Agita Jaunzeme, Community Manager, VMware
• Radmila Radovanvic, Senior Data Engineer, Northwestern Mutual
• Gage Russell, Data Engineer, Q2
• Rae Green, Developer, Q2ebanking
• Dimira Petrova, Supervisor of Data Analytics, VMware
• Martin Fiser, Head of Professional Services, Keboola
• Naga Raghavarapu, Principal Software Engineer, Oracle
• Antoni Ivanov, Staff Engineer, VMware

Agenda:

• Announcements
• Use cases from Northwestern Mutual and VMware
• New feature: linking job runs and datasets

Meeting:

• Recording
• Password: WMz0&@Gm

Notes:

• Announcements [Willy]

  • Marquez stickers are now available: https://www.astronomer.io/datakin-swag
  • Michael C. is presenting today at Airflow Summit @ 7 pm PT: https://airflowsummit.org/program/
  • Willy will be talking Mqz at Open Source Summit in June: https://sched.co/11NgS

• Northwestern Mutual Use Case [Joshua]

  • Big-picture role of Mqz at NWM
    • Mqz used to track data usage as a whole
    • Mqz critical at NWM to data ops, has special future here
  • Company background
    • Massive insurance co. with investment management arm
    • 150+ history with many customer touch points
    • Massive data with lots of users
  • Rationale for adoption
    
    • OL is where I spend most of my time
    • These tools will be the industry standards for dataset usage going forward
    • We desired one data standard, not random internal standards
  • Breakdown of use case
    • We track the HOW of usage from initial consumption to end usage
    • We record data product usage over time
    • Bonus: improved security
      • can see how/which users are actually using data
      • allows comparison to security frameworks, double-checking of work
    • Visualization is key
      • helps in building reports and modeling huge data systems
      • we can check the entire platform stack from ingest to updates, normalization, end-usage
  • Personal perspective
    • Mqz is data ops for data processing
    • Will we have a data ops center in the future like we have currently with NOCs?
    • The visual language is the key strength of the tool
    • This is the future of data
  • Q & A
    • Are screenshots available? Do you use Spark? [Naga]
      • Can't share due to proprietary concerns
    • How much data? [Naga]
      • Can't be specific, but it's a lot!
    • It's exciting to see others excited about the project. Are you using any custom integrations? [Willy]
      • Yes, custom integrations support streaming and ingestions across the platform

• VMware use case [Antoni]

  • Demo of VDK
  • Our motivation
    • Verification problems
    • OLMqz was the solution
    • The common standard provided by OL is essential
  • Why Mqz?
    • It's helpful in debugging complex jobs, troubleshooting
    • It's key to understanding usage for maintenance – e.g., enabling removal of irrelevant datasets, jobs
    • The shared metadata is useful
  • Diagram of architecture
  • Code demo
  • Suggestions
    • Add visualization of parent/child relationships [note: see PR 1935]
    • Make output searchable by metadata (e.g., make it possible to find all late jobs)
  • Our stack
    • Postgres, Presto, Snowflake, Greenplum db, Trino
  • Q & A
    • How many integrations in use? [Gage]
      • 100 teams, 1000s of tables
    • Are you using the Python client? [Willy]
      • Yes
    • It's amazing to get this feedback [Willy]
    • The grouping of jobs is hard, but we're addressing this
    • Feel free to open issues and contribute
• New feature linking job runs to datasets [Peter]
  • Recently added to jobs: created_by available on dataset views
  • Dataset versions also now available on version history tab
    • Allows for historical introspection in case of an issue
      • Allows for seeing if the code changed, for example
• Open discussion
  
  • Is anyone using the Python client for OL? [Gage]
    • Based on today's discussion, the answer is yes
  • Projects, docs are coming [Willy]
    • You can also use the Airflow integration for insight into the Python client
  • Column-level lineage has been added to OL [Willy]
    • We worked with Microsoft on the spec
    • Look for this in the API in the next few months
    • Feedback on this appreciated
  • What's in the roadmap for multi-tenancy? How can this be used in Mqz? [Naga]
    • For every event, route it through Kafka –  we're working with a company to help us document this a bit more [Willy]
    • Alternate approach: use a namespace to add metadata
    • Issue with this: access control (see the project roadmap for more info) 

April 28, 2022

Attendees:

TSC:

• Willy Lulciuc, Co-creator of Marquez
• Michael Collado, Staff Software Engineer, Astronomer
• Julien Le Dem, Chief Architect, Astronomer

And:

• Ross Turk, Senior Director of Community, Astronomer
• Minkyu Park, Senior Engineer, Astronomer
• John Thomas, Support Engineer, Astronomer
• Michael Robinson, Developer Relations Engineer, Astronomer
• Gage Russell, Data Engineer, Q2
• Paweł Leszczyński, Data Engineer, GetInData
• Joshua Wankowski, Associate Data Engineer, Northwest Mutual
• Dillon Stadther

Agenda:

• 0.22.0 preview [Willy]
• lifecycleStateChange support [Pawel]
• Updates to job renaming and symlinking [Michael C.]

Meeting:

Notes:

• Announcements [Willy]:

  • Cool swag is available! https://www.astronomer.io/datakin-swag
  • Willy has two talks about Marquez upcoming:
  • Airflow Summit: https://airflowsummit.org/program/
  • Open Source Summit: https://sched.co/11NgS

• 0.22.0 Preview [Willy]:

  • lifecycleStateChange support will offer visibility into dataset lifecycle changes, including deleting of tables
  • Pawel:
    • change motivated by desire for more information about datasets
    • approach started out with the Spark integration
    • still more information about lifecycle changes is possible/desirable
    • additional feature idea: notification console friendly to backend developers
  • Additional possibility: grayed out nodes on graph for deleted datasets, logging to show lifecycle history
  • Pawel: panel on website could display changes to dataset over X days
    • Agreed. Create an issue and we can build on that idea.
  • Helm chart addition
    • allows annotations, e.g. Prometheus metrics
  • Support for renaming and redirection
    • introducing job hierarchy
    • symlink will permit visibility into name changes to datasets

• Updates to job renaming and symlinking [Michael C.]

  • stemmed from desire to tie linked jobs together, e.g., jobs called by DAGs, even in cases where identical code is part of different chains
  • challenge: linking old jobs to fully qualified version
  • motivating factor: changes to job names results in junk nodes on graph
  • there was no way to remove the old job names from the graph
  • but there is frequently a need to keep track of old job names
  • hence the idea of symlinking a job
  • currently there's no API to do this
  • updating must be done manually currently
    • add the UUID of the new job to the db
    • from that point on, the job history will redirect to the new job (with a 301)
  • future: API will make this possible programmatically
  • Willy: is documentation needed for this?
    • Yes, I will post a change to the README
    • We want to do the same thing for datasets
• Open discussion
  • Gage: is a home repo coming?
    • Willy: Minkyu has looked into this
    • Willy: we want to add the Helm chart to the new website
    • Willy: this is on our radar
  • New release coming soon!

March 31, 2022

Attendees:

TSC:

• Willy Lulciuc, Co-creator of Marquez
• Michael Collado, Staff Engineer, Astronomer
• Julien Le Dem, Chief Architect, Astronomer
• Peter Hicks, Senior Engineer, Astronomer

And:

• Ross Turk, Sr. Director of Community, Astronomer
• Minkyu Park, Senior Engineer, Astronomer
• John Thomas, Support Engineer, Astronomer
• Michael Robinson, Developer Relations Engineer, Astronomer
• Howard Yoo, Staff Product Manager, Astronomer

Agenda:

• Website update
• Backlog and roadmap discussion
• Open discussion

Meeting:

Slides

Notes:

Announcements [Michael R.]

• Marquez stickers are now available: https://www.astronomer.io/datakin-swag
• Willy and Julien gave a talk on OpenLineage, Airflow and Marquez at Data Council Austin on March 23
• The project's Github star count stands at 983. Have you starred the project yet?
• 1k stars are a requirement for graduation status from the LFAI. The project is nearing completion of all requirements, so formal application will be possible soon.

Website [Ross]

• The project now has a new website.
• Appropriately, it's an open-source project; PRs are welcome.
• Tech: Gatsby, Github Projects
• Dev: run yarn deploy to work on it
• Plans: blog page. Proposals for posts welcome – post them in Slack or open a PR if you prefer.

Backlog and roadmap [Willy]

• Issue: currently, PRs are driven by a small team (e.g., Peter's view for dataset versions, Pawel's lifecycle PR)
• How to get the broader community involved? Want people to have more input/control over the issues we take up.
• Solution: Github's Roadmap feature. Milestones and releases visible there. Choose Marquez on the Projects tab.
• Process: review issues on monthly basis, move to roadmap, then release.
• Question from Howard about how to propose new features
• Follow-up work: discussion of how to prioritize issues; documentation needed about how to label new issues (e.g., as "features")
• Comment from Michael C.: it's possible to add new columns to the roadmap, in addition to new issues.

Open discussion

• Michael C.: please note issue #1928: supporting job grouping and hierarchy.
  • Problem: the project does not track parent/child job relationships, despite this nomenclature being used in OpenLineage to describe related jobs.
  • Proposal: a parent_job_id column should be added to the jobs table and to the runs table, both being uuids. 
• Michael R.: please note that the meeting typically takes place on the 4th Thursday of each month.

February 24, 2022

Attendees:

TSC:

• Willy Lulciuc, Co-creator of Marquez
• Michael Collado, Staff Engineer, Datakin

And:

• Minkyu Park, Senior Engineer, Datakin
• Michael Robinson, Developer Relations Engineer, Datakin
• Ross Turk, VP of Marketing, Datakin

Agenda:

• Review of integrations to create runs and associate metadata with runs (replaced with OpenLineage)
• Demo: How to collect OpenLineage events with the lineage API to send metadata to Marquez
• Demo: OL Java client
• Dataset lifecycle management
• Open discussion

Meeting:

Slides

Notes:

• Announcements [Willy]

  • Release date of 0.21.0 is now 2/28
  • Confusion in the community about which Java client to use is being addressed in OpenLineage PR #480
    • We hope to have this merged for the next OL release

• Integrations and OL demo [Willy]

  • OL integration
    • Available at openlineage.io/integration/, where you can also find instructions for installing and configuring it
    • Requirements.txt needs to install airflow
    • Set OpenLineage URL to local instance of Marquez
    • Marquez is moving towards using a task listener to pull metadata in real time 
    • For now use the OL Airflow DAG
    • You can still use the OL backend; there are limitations there, however
  • Spark integration

    • When doing the Spark submit command you need to provide configuration - specify the extra listener (thanks to Michael C for his work on this)

    • Point the host to your deployment

    • See the OL website for more details (openlineage.io/integration/spark-spark)
  • Upcoming: Flink and Kafka
  • Your feedback on these integrations appreciated
  • There are many connections you can use in your platform by switching over to OL to collect metadata

• OL Java client demo [Willy]

  • The Java client employs a workflow with interface
  • Definition of run method required
  • Instance of database required
  • This ex: simpleworkflow with database via newDatabase method
  • Relies on a Job class
  • In Marquez you can see the calls
  • For the code see https://github.com/DatakinHQ/demo/tree/main/custom/java/simple

• Dataset lifecycle management [Willy]

  • Marquez can now capture changes to dataset names
  • Community voiced desire for this feature
  • Marquez now supports soft deletes of datasets
  • See PR #1847
  • Support of lifecycle now more concrete: can see the phases datasets go through

• Open discussion

  • Julien and Willy will be speaking in-person at the Data Council conference in Austin next month (March 23-24)
  • Michael C. will be presenting virtually at the Subsurface LIVE conference (March 2-3); topic: Spark 

January 27, 2022

Attendees:

TSC:

• Willy Lulciuc, Co-creator of Marquez
• Julien Le Dem, CTO of Datakin
• Michael Collado, Staff Engineer, Datakin
• Peter Hicks, Senior Engineer, Datakin
• Kevin Mellott, Assistant Director of Data Engineering, Northwestern Mutual

And:

• Ross Turk, VP of Marketing, Datakin
• Minkyu Park, Senior Engineer, Datakin
• John Thomas, Support Engineer, Datakin
• Michael Robinson, Developer Relations Engineer, Datakin

Agenda:

• Marquez recent releases overview [Willy] 
  • Marquez release 0.21.0 overview
    • Upgrade to Java17
• Migrating integrations to OpenLineage [Willy]
• Cloud-based development instance of Marquez via Gitpod [Peter]
• Open discussion

Meeting:

Slides

Notes:

• 0.21.0 overview [Willy]

  • Features:
    • Bug fixes
    • Removal of excess code
    • Upgrade to Java17
      • API image migrated
      • Eclipse Temurin integrated
      • All CI deployment updated to support Java17
  • Discussion [Kevin, Willy, Michael C.]:
    • Support for Java client possible in lower version
    • Proposed: schedule separate meeting about this

• Migrating integrations to OpenLineage [Willy]

  • Spark library in Marquez now deprecated
  • Use of OpenLineage Spark integration recommended going forward
    • review the docs about how to configure your instance
    • remember to add underscore to marquez_airflow
  • OpenLineage integration allows task listener
    • workaround: import DAG from OpenLineage
  • See the changelog: environment variables for the Airflow instance have changed

• Cloud-based development instance of Marquez [Peter]

  • Enabled by integration of Gitpod
  • Docker image in the cloud with Marquez and UI
  • Ideal for those not ready to install everything locally or who are having issues with their OS
  • Fast (30 seconds), eliminates risk
  • API also available
  • Can be made private or public
  • Big advantage: shareable within organizations via URL
  • Supports everything one could do locally in VS Code or similar IDE
  • Discussion [Willy, Peter, Kevin, Julien]:
    • common use case: potential users want to see metadata from their org and share the tool
    • potential side-effect: increase in Docker pulls
    • availability of metrics unknown
    • email address required

• Open Discussion

  • Advantages of possible move from CircleCI to Github Actions 
    • CircleCI downsides: outages, billing issues [Willy]
    • Julien proposed: moving to Github actions eventually after running both in parallel
    • Kevin asked to experiment with Github Actions and report back
  • Issue #1800: add support for table operations reported from OpenLineage
    • Formal solution needed [Willy]
    • Willy proposed: deploy in two modes and use flags (Julien agreed)
  • NodeID
    • An easy win: add a field that returns a nodeID [Willy]
    • Willy proposed: prioritize in next release

Marquez Workflow Group Calendar Overview

Effective March 22, 2019: Group calendars are managed within LF AI Foundation Groups.io subgroups (mail lists); with each sub-group (mail list) having a unique group calendar. Meeting invites from these group calendars are sent to the applicable sub-group (mail list). In order to see the various group calendars you must:

• Be logged into LF AI Foundation Groups.io

• Be subscribed to the sub-group(mail-list) you're interested in

• Thereafter, you will see all the calendars for the sub-groups you subscribe to under your LF AI Foundation Group Calendar via Groups.io OR

• You can also view a specific group calendar via the Wiki (if the group has created a Wiki group calendar) whether you are a member of the sub-group (mail list) or not

  • Example: LF AI TAC Group Calendar (tac-general@lists...) via Wiki

View Instructions on How to Subscribe to LF AI Group Calendars

For detailed information on LF AI meeting management processes view this page: LF AI Foundation - Community Meetings and Calendars

Marquez Meetings List

Marquez Group Calendar