Skip to main content
Version: 0.8

0.7 to 0.8 Upgrade Guide

Overview​

Version 0.8 of the SDK comes with performance and cost improvements, new capabilities to define more powerful features, the ability to manage releases of the Tecton Materialization Runtime, and an enhanced development experience. It also includes minor changes to the behavior and naming of some Tecton objects, methods, and CLI commands. Read the general Upgrade Process page in addition to the following documentation for guidance on how to safely upgrade from 0.7 to 0.8.

To ensure a safe upgrade, Tecton disallows any destructive changes (i.e. Recreates) to your Feature Repository while upgrading. For example, tecton apply will prevent changes to a Feature View's Offline Store format while upgrading as this would normally cause re-materialization.

Step-by-Step Upgrade Flow​

These are the most important changes in Tecton 0.8:

  • Materialization clusters are required to pin a specific version of the tecton library (using the tecton_materialization_runtime on Batch Feature Views, Stream Feature Views, and Feature Tables).
  • The default Offline Store format is changing from Parquet to Delta.
  • Tecton is introducing a new required Repo Config .yaml file to configure defaults in your Feature Repository.
  • Stream jobs for Spark-based Stream Feature Views must run on on-demand instances. This may be a breaking change if your Feature Repository explicitly sets instance_availability to spot or spot_with_fallback in stream_compute for Stream Feature Views.

Most customers should follow this guidance while upgrading:

  1. Repo Config file: Use Tecton's CLI to generate a repo.yaml file with pre-filled defaults for tecton_materialization_runtime:

    • If your Feature Repository does not specify an Offline Store format, run tecton repo-config init --parquet to explicitly set the default format to Parquet (in order to avoid destructive changes or rematerialization).
    • Otherwise, run tecton repo-config init

  2. Spark-based Stream Feature Views: If stream_compute is set, ensure that instance_availability is either not specified, or is set to on_demand.

  3. Deprecated Methods, Parameters, or Attributes: Ensure that your Feature Repository does not rely on any Tecton methods, parameters, or attributes that are deprecated or removed in 0.8. This may be a breaking change if your Feature Repository relies on methods, parameters, or attributes removed in 0.8.

Please see the following sections for details on all changes in 0.8.

Changes in 0.8​

Versioning of the Tecton Materialization Runtime​

Tecton 0.8 introduces versioning of the Tecton Materialization Runtime that is deployed to Databricks & EMR clusters for orchestration of backfills and materialization. This further improves the reliability of Tecton releases beyond our robust testing and validation process by letting customers iteratively upgrade Feature Views and Tables.

In 0.8+, tecton_materialization_runtime is a required parameter on Batch Feature Views, Stream Feature Views, and Feature Tables. It must be set to an exact version of the tecton package (e.g. 0.8.0). Tecton guarantees +1 support -- for example, features applied with Tecton SDK 0.8 will support Materialization Runtime Versions of 0.8.* or 0.9.*.

This means that jobs for Feature Views & Tables applied using any version of Tecton will no longer automatically update (or automatically restart, in the case of Stream Feature Views) when Tecton releases changes to the materialization runtime. Customers should instead upgrade their Feature Views & Tables by setting the tecton_materialization_runtime parameter when upgrading their SDK to a new minor version or when recommended by Tecton Support. Tecton will also document release notes in the Changelog.

The new Repo Config, described in the following section, helps customers define a default tecton_materialization_runtime for all Feature Views & Tables in their Feature Repository.

Repo Config File​

Tecton 0.8 introduces the Repo Config, a required configuration file used to set defaults for Tecton objects in a Feature Repository. This helps developers avoid having to specify certain parameters for every new Tecton object and ultimately results in simpler feature definitions. For example, developers can use this file to set a default tecton_materialization_runtime for all Feature Views in a Feature Repository.

During tecton plan/apply/test, Tecton will look for a Repo Config file named repo.yaml in the root of your Feature Repository. To specify another file, use the --config flag (e.g. tecton plan --config my_config.yaml).

Tecton will automatically generate a new Repo Config file when initializing a new Feature Repository using tecton init. For existing Feature Repositories, run tecton repo-config init to generate a Repo Config file named repo.yaml with pre-filled defaults for tecton_materialization_runtime.

If you are also explicitly setting the Offline Store version to ParquetConfig (as described in the following section), use tecton repo-config init --parquet.

Changes to configuring offline_store​

offline_store is an optional parameter used to configure Tecton's Offline Store for Batch Feature Views, Stream Feature Views, and Feature Tables.

OfflineStoreConfig​

Tecton 0.8 introduces the new OfflineStoreConfig object for configuring offline_store and the new Publish Features functionality.

The OfflineStoreConfig object includes an optional parameter (staging_table_format) to configure the format of the Offline Store to DeltaConfig() or ParquetConfig(). Before 0.8, this was done by setting offline_store=ParquetConfig() or offline_store=DeltaConfig().

In 0.8, offline_store can be set to OfflineStoreConfig, ParquetConfig, or DeltaConfig. In a future version of Tecton, offline_store will only support being set to OfflineStoreConfig.

Changes to the Default Offline Store Format​

Tecton 0.8 changes the default format for the Offline Store from Parquet to Delta. If your Feature Repository does not already explicitly set offline_store, you'll need to set the offline_store format to Parquet on Batch Feature Views, Stream Feature Views, and Feature Tables to avoid recreates during the upgrade. This can be done by:

  1. Running tecton repo-config init --parquet, which adds the following to your Repo Config file:
offline_store:
  kind: OfflineStoreConfig
  staging_table_format:
    kind: ParquetConfig
  1. OR by explicitly adding this to your Feature Views & Tables:
offline_store = OfflineStoreConfig(staging_table_format=ParquetConfig())

Changes to Tecton on Snowflake​

Feature Views with mode="snowpark" are deprecated in 0.8 and will be removed in a future version of the SDK.

The output format of get_historical_features() now takes the Data Source parameter data_delay into consideration. See the SDK Reference for more details on this parameter.

Additionally, get_historical_features(start_time, end_time) now returns an additional column named _effective_timestamp. Please see this page for more details on this parameter.

Some users previously configured the ALPHA_SNOWFLAKE_COMPUTE_ENABLED flag to opt-in to the beta offline retrieval query engine used by Tecton on Snowflake. In 0.8, this new query engine is enabled by default when connected to a Snowflake cluster. To explicitly opt-in to this version (for example, when retrieving features in a Jupyter notebook), users can set tecton.conf.set('TECTON_OFFLINE_RETRIEVAL_COMPUTE_MODE', 'snowflake').

Changes to the Athena connector​

Tecton can be used with Athena to retrieve features in any Python environment.

In Tecton 0.7, we introduced major optimizations to queries generated using Tecton's Athena backend. Users previously opted-in to these optimizations by setting the SQL_DIALECT and ALPHA_ATHENA_COMPUTE_ENABLED configs.

In 0.8, these configs are no longer used as this behavior is enabled by default. To activate Tecton's Athena backend, set tecton.conf.set('TECTON_OFFLINE_RETRIEVAL_COMPUTE_MODE', 'athena').

To use the deprecated Athena connector, set tecton.conf.set('USE_DEPRECATED_ATHENA_RETRIEVAL', True). The deprecated connector will be removed in a future version of the Tecton SDK.

See the Athena connector documentation for more details.

Changes to get_historical_features(), run()​

In Tecton 0.7, None is not a valid entity value. If you use get_historical_features() with a spine that contains None entity, the rows with it will be dropped from the returned results. get_historical_features() and run() does not return None entity in its results.

In 0.8, the row will no longer be dropped. None value is treated as a valid key, and will return features the same way as other keys. get_historical_features() and run() will now return None entity in its results.

This behavior matches what Tecton returns from the Online Store.

Instance Type for Spark-based Stream Feature Views​

0.8 requires that stream jobs for Spark-based Stream Feature Views use On-Demand instances (instance_availability="on_demand"). This is to prevent issues with Spot Instances negatively impacting the freshness and reliability of Stream Feature Views. If not specified in stream_compute (within the DatabricksClusterConfig or EMRClusterConfig), Tecton will automatically set instance_availability to "on_demand".

This is a breaking change if you previously explicitly set instance_availability to spot or spot_with_fallback in stream_compute for Spark-based Stream Feature Views.

SDK Interfaces that are deprecated or removed in 0.8​

Methods that are deprecated in 0.8 and will be removed in the future​

Deprecated ParameterReplacement
FeatureView.deletion_status()Retrieve the Deletion Job ID (via FeatureView.delete_keys() or FeatureView.list_materialization_jobs()) and call get_materialization_job(job_id).status

Methods, parameters, & attributes that were previously deprecated and are officially removed in 0.8​

This is a breaking change if your Feature Repository still uses the removed methods, parameters, or attributes.

Removed ParameterReplacement
Workspace.get_all()tecton.list_workspaces()
FeatureView.max_batch_aggregation_intervalFeatureView.max_backfill_interval

CLI Changes​

Changes to tecton access-control​

The tecton access-control assign-role and tecton access-control unassign-role commands now apply only to the currently-selected workspace by default. Use the --workspace flag to apply ACL changes to a specific workspace or the --all-workspaces flag to apply changes to all workspaces.

Note that assigning or unassigning the admin role will always apply to all workspaces since it is a deployment-wide role.

Removal of --safety-checks and --no-safety-checks​

--safety-checks / --no-safety-checks were deprecated in 0.7 and are officially removed in 0.8. Users should use the --yes or -y flag to disable interactive safety checks when running apply, destroy, plan, and upgrade commands.

FAQ​

What does a successful upgrade look like?​

Here's an example of a successful apply during an upgrade. In this case, we've set the tecton_materialization_runtime and updated the offline_store parameter.

✅ Imported 1 Python module from the feature repository
⚠️  Running Tests: No tests found.
✅ Collecting local feature declarations
✅ Performing server-side feature validation: Initializing.
 ↓↓↓↓↓↓↓↓↓↓↓↓ Plan Start ↓↓↓↓↓↓↓↓↓↓

  ~ Update Batch Feature View
    name:           my_feature_view_1
    description:    Whether the user performing the transaction is over 18 years old.
    tecton_materialization_runtime:  -> 0.8.0
    offline_store:   0.7 Offline Store -> 0.8+ Offline Store

  ~ Update Batch Feature View
    name:           my_feature_view_2
    description:    Whether the user performing the transaction is over 18 years old.
    tecton_materialization_runtime:  -> 0.8.0
    offline_store:   0.7 Offline Store -> 0.8+ Offline Store

 ↑↑↑↑↑↑↑↑↑↑↑↑ Plan End ↑↑↑↑↑↑↑↑↑↑↑↑
 Generated plan ID is ...
 View your plan in the Web UI: ...

Are you sure you want to apply this plan to: "your_workspace"? [y/N]> y
🎉 Done! Applied changes to 2 objects in workspace "your_workspace".

What does a blocked upgrade look like?​

During an SDK version upgrade, Tecton automatically blocks destructive changes (i.e. Recreates) to prevent unintended changes.

Here's an example of a blocked destructive change:

✅ Imported 1 Python module from the feature repository
⚠️  Running Tests: No tests found.
✅ Collecting local feature declarations
⛔ Performing server-side feature validation: Finished generating plan.
Blocked destructive change to 'feature_view'. Destructive change are not allowed when upgrading SDK versions.
	tecton_materialization_runtime:  -> 0.8.0
	offline_store:
	{"stagingTableFormat":{"parquet":{}}}
	 ⬇⬇⬇⬇⬇
	{"stagingTableFormat":{"delta":{"timePartitionSize":"86400s"}}}

In this case, the staging_table_format change (in OfflineStoreConfig) is a destructive change (a Recreate) that was blocked. To resolve this, explicitly set the staging_table_format to ParquetConfig.

Was this page helpful?