[FEATURE]: Build and display dataset lineage to partition/schedule code migrations more effectively

[rwforest]

Is there an existing issue for this?

• I have searched the existing issues

Problem statement

Table mapping does not solve everything, chances are there's still error after migration. Since HMS lineage is there, UCX should target merging with dependencyGraph

flowchart TD

storage_path -->|reads| view
storage_path -->|reads| table
storage_path -->|reads| notebook
storage_path -->|reads| py_file
storage_path -->|reads| redash_query
storage_path -->|reads| pipeline

table --> view
view --> table

table -->|reads| notebook
notebook -->|writes| table

table -->|reads| pipeline
pipeline -->|writes| table

table -->|reads| py_file
py_file -->|writes| table

table -->|reads| redash_query
redash_query -->|writes| table
redash_query --> dashboard
dashboard --> warehouse

table -->|reads| lakeview_dashboard
lakeview_dashboard --> warehouse

notebook --> pipeline
pipeline --> job
notebook --> job
wheel --> job

py_file --> job
py_file --> git_repo
py_file --> wheel

notebook --> git_repo
git_repo -->|?| job

cluster_policy --> cluster
cluster_policy --> job
job --> task
task --> cluster
cluster --> init_script

warehouse -.-> cluster

Loading

Proposed Solution

Merge HMS lineage with dependencyGraph. While it is dependent on the version of the DBR, it should start with the highest runtime and then backfill anything that's not captured using other means. Static lineage parsing or Spark listener.

scope:

asset	has_owner_user	listing speed	AST analysis required
storage path	no	slow (via AST analysis)	yes
view	no	medium, via tables.scala	yes
table	no	medium, via tables.scala	no
pipeline	yes	fast	** yes
notebook	yes	slow, via workflow linter	yes
wheel	no	slow, via linter	yes
job	yes	fast	no
cluster	yes	fast	no
cluster_policy	yes	fast	no
git_repo	no	-	no
py_file	no	slow, via workflow linter	yes
redash query	yes	medium	yes
redash dashboard	yes	medium	no
lakeview dashboard	yes	fast	yes
warehouse	yes	fast	no

Optionally, we can create multiple copies of the same graph with starting points from a single table to show full migration scope.

Outcome:

• show dashboard with Asset (link), Asset type, Owner, Failures columns, sorted by their required order. Filter on owner.
• topologically sort all assets in order to schedule migration
• (optionally) process graph to group islands of dependencies in order to parallelise migration or find opportunities
• new table: lineage

Additional Context

No response

[nfx]

@rwforest The request is not clear, could you elaborate?

[rwforest]

@nfx When doing table replacement, source to target table is only the first pass. How do we guarantee that anything will run successfully? That's the goal of the migration. So the goal here is, for example, given 1000 notebooks, I need to be able to understand which chain of commands are the most critical. In terms of dependencyGraph we can check the incoming and outgoing edges. If we know a notebook is an orphan, then we don't attempt to fix it. Right now even if you fix all the cells, it is only the beginning of all the errors that will surface eventually.

[JCZuurmond]

@rwforest : Could you clarify the following:

1. What is the HMS dependency graph?
2. Why is this dependent on the Databricks runtime?

Generally speaking, I would like capability to plan code migrations using UCX based on a dependency graph of the (to-be-migrated) tables linted from the code

[rwforest]

@JCZuurmond there's a feature in Databricks that never made it to public I think it was to build run time dataset dependency graph, and I was told by @FastLee that it is dependent on the DBR version. I believe the linted code is static lineage but I got some notebooks that are heavily parameterized.

And I agree for the planning part. Is there a roadmap on some planning UI? I can't imagine how someone would plan code migration using csv.

[JCZuurmond]

A planning UI is not on the roadmap. We are considering to include this issue in our upcoming planning, but no decision yet