feat: Unify Feature Transformations and Feature Views #5747

franciscojavierarceo · 2025-11-28T18:20:26Z

What this PR does / why we need it:

This PR refactors the transformation system with a cleaner architecture that separates transformation logic from execution. Transformations should focus on HOW to transform data, while FeatureViews handle WHEN and WHERE to execute.

Changes:

Added: feature_transformation field to FeatureView for transformation logic
Added: transform: bool = True parameter to API methods for per-call control
Added: Schema validation for pre-transformed data (transform=False)
Added: Transformation execution logic in online/offline feature retrieval

  # Define transformation logic
  @transformation(mode="python")
  def driver_features(inputs):
      return {"conv_plus_acc": [c + a for c, a in zip(inputs["conv_rate"], inputs["acc_rate"])]}

  # Create FeatureView with transformation
  fv = FeatureView(
      name="driver_computed_features",
      source=driver_source,
      entities=[driver],
      schema=[Field(name="conv_plus_acc", dtype=Float64)],
      feature_transformation=driver_features,  # ← NEW
      online=True,
  )

  # API-level transformation control
  # Execute transformations (default)
  features = store.get_online_features(
      features=["driver_computed_features:conv_plus_acc"],
      entity_rows=entity_rows,
      transform=True  # ← NEW (default)
  )

  # Skip transformations for external batch jobs
  features = store.get_online_features(
      features=["driver_computed_features:conv_plus_acc"],
      entity_rows=spark_transformed_rows,
      transform=False  # ← NEW: skip + validate schema
  )

Which issue(s) this PR fixes:

#4584, #5716, #5689

Misc

ntkathole · 2025-12-11T11:47:34Z

sdk/python/feast/transformation/base.py

+                from feast.feature_view import FeatureView as FV
+
+                for src in sources:
+                    if not isinstance(src, (FV, type(src).__name__ == "FeatureView")):


Suggested change

if not isinstance(src, (FV, type(src).__name__ == "FeatureView")):

if not isinstance(src, FeatureView):

ntkathole · 2025-12-11T11:52:23Z

sdk/python/feast/feature_view.py

        owner: str = "",
        mode: Optional[Union["TransformationMode", str]] = None,
+        feature_transformation: Optional[Transformation] = None,
+        when: Optional[Union[TransformationTiming, str]] = None,


Need updates to proto ? when and online_enabled not serialized in from_proto.

Also, modify __eq__() and __copy__() for consider these new attributes ?

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

HaoXuAI · 2025-12-16T18:35:44Z

I think when can be a bit confusing. even for 'batch', it can be 'on_read' or 'on_write'. I think on a different angle, should transformation be in the scope of writing, and isn't transformation only responsible for transforming the data?

FeatureView as the actually data assets, I think makes more sense to having the API.

… auto-inference Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

franciscojavierarceo · 2025-12-23T19:01:57Z

I think when can be a bit confusing. even for 'batch', it can be 'on_read' or 'on_write'. I think on a different angle, should transformation be in the scope of writing, and isn't transformation only responsible for transforming the data?

FeatureView as the actually data assets, I think makes more sense to having the API.

Yeah, that's fair. @HaoXuAI I updated the PR, lmk what you think.

HaoXuAI · 2025-12-23T19:53:28Z

sdk/python/feast/feature_view.py

+        self.transform_when = transform_when
+
+        # Auto-infer online setting based on transform_when pattern
+        if transform_when in [


So why batch on read/write forces online to be true?
I think we need to think about if we really need a new tranform_when config. And if we need it, what is its relation with online/offline.
I think it might be counterintuitive that if this featureview is only used by batch or stream or on demand. what if the user defines it today as one mode and wants to change it to another one or allow both mode in the future?
My personal pint of view is removing any config for this. As long as you define the transformation, it will be applied to run before writing to destination. Use online/offline to control the destination only, if both are off, then no write. For any transformation needed on read, this should be configured at API level. Fo example, you can put get_online_feature(transform=true) to force transformation.

I think we need to think about if we really need a new tranform_when config. And if we need it, what is its relation with online/offline.

Here's how things actually work:

Batch

Sometimes data can be transformed online

On Reads

On Writes

Sometimes not, and that's basically a batch transform with no feature server transformation and pure materialization

Streaming

Handles historical data the same way it handles online serving

that's really what the transform_when indicates.

We need something other than online because the rest will be executed by the server and you can't orchestrate the transformation on write during retrieval.

I'm fine with renaming something but we need something equivalent to write_to_online_store that's more obvious.

online = true means write to online store.

transform_on_write should be removed from the server. it's a hacky way. #5283. this PR doesn't get enough reviews I think.
#5300. I think this one is also not ideal implementation as well.

When you use get_online_feature, you should assign it a transform=true to run transformation in "retrieval".

I've read through the feature server code and on demand feature view implementation. I think a lot code needs to be refactored. And if you think there is some lines of code won't work you can tag me. I'll try to explain it

Yes we agreed on this direction and I'm going to implement things that way. Just need to finish the work.

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

franciscojavierarceo requested a review from a team as a code owner November 28, 2025 18:20

franciscojavierarceo force-pushed the refactor-odfv branch 2 times, most recently from 2899413 to 649c429 Compare December 1, 2025 21:47

franciscojavierarceo added the ok-to-test label Dec 2, 2025

ntkathole reviewed Dec 11, 2025

View reviewed changes

franciscojavierarceo added 5 commits December 11, 2025 17:28

feat: Add unified transformation

c774fc7

feat: Unify transformations

c4432c0

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

feat: Unify Transformations

fc85c27

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

feat: Unify Transformations

dd364b8

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

updated docs

c77cb83

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

ntkathole force-pushed the refactor-odfv branch from 8b578bf to c77cb83 Compare December 11, 2025 11:58

franciscojavierarceo and others added 4 commits December 16, 2025 16:57

Merge branch 'master' into refactor-odfv

c5c495d

Merge branch 'master' into refactor-odfv

b4aea01

refactor: separate transformation logic from execution decisions with…

686654c

… auto-inference Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

format

2728d9d

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

HaoXuAI reviewed Dec 23, 2025

View reviewed changes

franciscojavierarceo added 6 commits December 24, 2025 13:10

incorporaitng feedback

6b9deec

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

updated

92825a5

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

fix

d485b20

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

linter

dda2f34

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

cleanup

3408523

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

cleanup

2fab232

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: Unify Feature Transformations and Feature Views #5747

feat: Unify Feature Transformations and Feature Views #5747

Uh oh!

franciscojavierarceo commented Nov 28, 2025 •

edited

Loading

Uh oh!

ntkathole Dec 11, 2025

Uh oh!

ntkathole Dec 11, 2025 •

edited

Loading

Uh oh!

HaoXuAI commented Dec 16, 2025 •

edited

Loading

Uh oh!

franciscojavierarceo commented Dec 23, 2025

Uh oh!

HaoXuAI Dec 23, 2025

Uh oh!

franciscojavierarceo Dec 23, 2025

Uh oh!

franciscojavierarceo Dec 23, 2025

Uh oh!

HaoXuAI Dec 25, 2025

Uh oh!

HaoXuAI Dec 25, 2025

Uh oh!

franciscojavierarceo Dec 25, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	if not isinstance(src, (FV, type(src).__name__ == "FeatureView")):
	if not isinstance(src, FeatureView):

feat: Unify Feature Transformations and Feature Views #5747

Are you sure you want to change the base?

feat: Unify Feature Transformations and Feature Views #5747

Uh oh!

Conversation

franciscojavierarceo commented Nov 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this PR does / why we need it:

Which issue(s) this PR fixes:

Misc

Uh oh!

ntkathole Dec 11, 2025

Choose a reason for hiding this comment

Uh oh!

ntkathole Dec 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

HaoXuAI commented Dec 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

franciscojavierarceo commented Dec 23, 2025

Uh oh!

HaoXuAI Dec 23, 2025

Choose a reason for hiding this comment

Uh oh!

franciscojavierarceo Dec 23, 2025

Choose a reason for hiding this comment

Uh oh!

franciscojavierarceo Dec 23, 2025

Choose a reason for hiding this comment

Uh oh!

HaoXuAI Dec 25, 2025

Choose a reason for hiding this comment

Uh oh!

HaoXuAI Dec 25, 2025

Choose a reason for hiding this comment

Uh oh!

franciscojavierarceo Dec 25, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

franciscojavierarceo commented Nov 28, 2025 •

edited

Loading

ntkathole Dec 11, 2025 •

edited

Loading

HaoXuAI commented Dec 16, 2025 •

edited

Loading