Point-in-time support using time series feature tables

The data used to train a model often has time dependencies built into it. For example, if you are training a model to predict which machines on a factory floor need maintenance, you might have historical datasets that contain sensor measurements and usage data for many machines, along with target labels that indicate if the machine needed service or not. The dataset might contain data for machines both before and after a maintenance service was performed.

When you build the model, you must consider only feature values up until the time of the observed target value (needs service or does not need service). If you do not explicitly take into account the timestamp of each observation, you might inadvertently use feature values measured after the timestamp of the target value for training. This is called “data leakage” and can negatively affect the model’s performance.

Time series feature tables include a timestamp key column that ensures that each row in the training dataset represents the latest known feature values as of the row’s timestamp. You should use time series feature tables whenever feature values change over time, for example with time series data, event-based data, or time-aggregated data.

Note

• With Databricks Runtime 13.3 LTS and above, any Delta table in Unity Catalog with primary keys and timestamp keys can be used as a time series feature table.

• For better performance in point-in-time lookups, Databricks recommends that you apply Liquid Clustering (for databricks-feature-engineering 0.6.0 and above) or Z-Ordering (for databricks-feature-engineering 0.6.0 and below) on time series tables.

• Point-in-time lookup functionality is sometimes referred to as “time travel”. The point-in-time functionality in Databricks Feature Store is not related to Delta Lake time travel.

How time series feature tables work

Suppose you have the following feature tables. This data is taken from the example notebook.

The tables contain sensor data measuring the temperature, relative humidity, ambient light, and carbon dioxide in a room. The ground truth table indicates if a person was present in the room. Each of the tables has a primary key (‘room’) and a timestamp key (‘ts’). For simplicity, only data for a single value of the primary key (‘0’) is shown.

The following figure illustrates how the timestamp key is used to ensure point-in-time correctness in a training dataset. Feature values are matched based on the primary key (not shown in the diagram) and the timestamp key, using an AS OF join. The AS OF join ensures that the most recent value of the feature at the time of the timestamp is used in the training set.

As shown in the figure, the training dataset includes the latest feature values for each sensor prior to the timestamp on the observed ground truth.

If you created a training dataset without taking into account the timestamp key, you might have a row with these feature values and observed ground truth:

temp	rh	light	co2	ground truth
15.8	32	212	630	0

However, this is not a valid observation for training, because the co2 reading of 630 was taken at 8:52, after the observation of the ground truth at 8:50. The future data is “leaking” into the training set, which will impair the model’s performance.

Requirements

• For Feature Engineering in Unity Catalog: Feature Engineering in Unity Catalog client (any version).

• For Workspace Feature Store: Feature Store client v0.3.7 and above.

How to specify time-related keys

To use point-in-time functionality, you must specify time-related keys using the timeseries_columns argument (for Feature Engineering in Unity Catalog) or the timestamp_keys argument (for Workspace Feature Store). This indicates that feature table rows should be joined by matching the most recent value for a particular primary key that is not later than the value of the timestamps_keys column, instead of joining based on an exact time match.

If you do not use timeseries_columns or timestamp_keys, and only designate a timeseries column as a primary key column, feature store does not apply point-in-time logic to the timeseries column during joins. Instead, it matches only rows with an exact time match instead of matching all rows prior to the timestamp.

Create a time series feature table in Unity Catalog

In Unity Catalog, any table with a TIMESERIES primary key is a time series feature table. See Create a feature table in Unity Catalog for how to create one.

Create a time series feature table in local workspace

To create a time series feature table in the local Workspace Feature Store, the DataFrame or schema must contain a column that you designate as the timestamp key.

Starting with Feature Store client v0.13.4, timestamp key columns must be specified in the primary_keys argument. Timestamp keys are part of the “primary keys” that uniquely identify each row in the feature table. Like other primary key columns, timestamp key columns cannot contain NULL values.

fe = FeatureEngineeringClient()
# user_features_df DataFrame contains the following columns:
# - user_id
# - ts
# - purchases_30d
# - is_free_trial_active
fe.create_table(
  name="ml.ads_team.user_features",
  primary_keys=["user_id", "ts"],
  timeseries_columns="ts",
  features_df=user_features_df,
)

fs = FeatureStoreClient()
# user_features_df DataFrame contains the following columns:
# - user_id
# - ts
# - purchases_30d
# - is_free_trial_active
fs.create_table(
  name="ads_team.user_features",
  primary_keys=["user_id", "ts"],
  timestamp_keys="ts",
  features_df=user_features_df,
)

fs = FeatureStoreClient()
# user_features_df DataFrame contains the following columns:
# - user_id
# - ts
# - purchases_30d
# - is_free_trial_active
fs.create_table(
  name="ads_team.user_features",
  primary_keys="user_id",
  timestamp_keys="ts",
  features_df=user_features_df,
)

A time series feature table must have one timestamp key and cannot have any partition columns. The timestamp key column must be of TimestampType or DateType.

Databricks recommends that time series feature tables have no more than two primary key columns to ensure performant writes and lookups.

Update a time series feature table

When writing features to the time series feature tables, your DataFrame must supply values for all features of the feature table, unlike regular feature tables. This constraint reduces the sparsity of feature values across timestamps in the time series feature table.

fe = FeatureEngineeringClient()
# daily_users_batch_df DataFrame contains the following columns:
# - user_id
# - ts
# - purchases_30d
# - is_free_trial_active
fe.write_table(
  "ml.ads_team.user_features",
  daily_users_batch_df,
  mode="merge"
)

fs = FeatureStoreClient()
# daily_users_batch_df DataFrame contains the following columns:
# - user_id
# - ts
# - purchases_30d
# - is_free_trial_active
fs.write_table(
  "ads_team.user_features",
  daily_users_batch_df,
  mode="merge"
)

Streaming writes to time series feature tables is supported.

Create a training set with a time series feature table

To perform a point-in-time lookup for feature values from a time series feature table, you must specify a timestamp_lookup_key in the feature’s FeatureLookup, which indicates the name of the DataFrame column that contains timestamps against which to lookup time series features. Databricks Feature Store retrieves the latest feature values prior to the timestamps specified in the DataFrame’s timestamp_lookup_key column and whose primary keys (excluding timestamp keys) match the values in the DataFrame’s lookup_key columns, or null if no such feature value exists.

feature_lookups = [
  FeatureLookup(
    table_name="ml.ads_team.user_features",
    feature_names=["purchases_30d", "is_free_trial_active"],
    lookup_key="u_id",
    timestamp_lookup_key="ad_impression_ts"
  ),
  FeatureLookup(
    table_name="ml.ads_team.ad_features",
    feature_names=["sports_relevance", "food_relevance"],
    lookup_key="ad_id",
  )
]

# raw_clickstream DataFrame contains the following columns:
# - u_id
# - ad_id
# - ad_impression_ts
training_set = fe.create_training_set(
  df=raw_clickstream,
  feature_lookups=feature_lookups,
  exclude_columns=["u_id", "ad_id", "ad_impression_ts"],
  label="did_click",
)
training_df = training_set.load_df()

Tip

For faster lookup performance when Photon is enabled, pass use_spark_native_join=True to FeatureEngineeringClient.create_training_set. This requires databricks-feature-engineering version 0.6.0 or above.

feature_lookups = [
  FeatureLookup(
    table_name="ads_team.user_features",
    feature_names=["purchases_30d", "is_free_trial_active"],
    lookup_key="u_id",
    timestamp_lookup_key="ad_impression_ts"
  ),
  FeatureLookup(
    table_name="ads_team.ad_features",
    feature_names=["sports_relevance", "food_relevance"],
    lookup_key="ad_id",
  )
]

# raw_clickstream DataFrame contains the following columns:
# - u_id
# - ad_id
# - ad_impression_ts
training_set = fs.create_training_set(
  df=raw_clickstream,
  feature_lookups=feature_lookups,
  exclude_columns=["u_id", "ad_id", "ad_impression_ts"],
  label="did_click",
)
training_df = training_set.load_df()

Any FeatureLookup on a time series feature table must be a point-in-time lookup, so it must specify a timestamp_lookup_key column to use in your DataFrame. Point-in-time lookup does not skip rows with null feature values stored in the time series feature table.

Set a time limit for historical feature values

With Feature Store client v0.13.0 or above, or any version of Feature Engineering in Unity Catalog client, you can exclude feature values with older timestamps from the training set. To do so, use the parameter lookback_window in the FeatureLookup.

The data type of lookback_window must be datetime.timedelta, and the default value is None (all feature values are used, regardless of age).

For example, the following code excludes any feature values that are more than 7 days old:

from datetime import timedelta

feature_lookups = [
  FeatureLookup(
    table_name="ml.ads_team.user_features",
    feature_names=["purchases_30d", "is_free_trial_active"],
    lookup_key="u_id",
    timestamp_lookup_key="ad_impression_ts",
    lookback_window=timedelta(days=7)
  )
]

from datetime import timedelta

feature_lookups = [
  FeatureLookup(
    table_name="ads_team.user_features",
    feature_names=["purchases_30d", "is_free_trial_active"],
    lookup_key="u_id",
    timestamp_lookup_key="ad_impression_ts",
    lookback_window=timedelta(days=7)
  )
]

When you call create_training_set with the above FeatureLookup, it automatically performs the point-in-time join and excludes feature values older than 7 days.

The lookback window is applied during training and batch inference. During online inference, the latest feature value is always used, regardless of the lookback window.

Score models with time series feature tables

When you score a model trained with features from time series feature tables, Databricks Feature Store retrieves the appropriate features using point-in-time lookups with metadata packaged with the model during training. The DataFrame you provide to FeatureEngineeringClient.score_batch (for Feature Engineering in Unity Catalog) or FeatureStoreClient.score_batch (for Workspace Feature Store) must contain a timestamp column with the same name and DataType as the timestamp_lookup_key of the FeatureLookup provided to FeatureEngineeringClient.create_training_set or FeatureStoreClient.create_training_set.

Tip

For faster lookup performance when Photon is enabled, pass use_spark_native_join=True to FeatureEngineeringClient.score_batch. This requires databricks-feature-engineering version 0.6.0 or above.

Publish time series features to an online store

You can use FeatureEngineeringClient.publish_table (for Feature Engineering in Unity Catalog) or FeatureStoreClient.publish_table (for Workspace Feature Store) to publish time series feature tables to online stores. Databricks Feature Store provides the functionality to publish either a snapshot or a window of time series data to the online store, depending on the online store provider. The following table shows the supported modes for each provider.

Online store provider	Snapshot mode	Window mode
Amazon DynamoDB (v0.3.8 and above)	X	X
Amazon Aurora (MySQL-compatible)	X
Amazon RDS MySQL	X

Publish a time series snapshot

In snapshot mode, publish_table publishes the latest feature values for each primary key in the feature table. The online store supports primary key lookup but does not support point-in-time lookup.

For online stores that do not support time to live, Databricks Feature Store supports only snapshot publish mode. For online stores that do support time to live, the default publish mode is snapshot unless time to live (ttl) is specified in the OnlineStoreSpec at the time of creation.

Publish a time series window

In window mode, publish_table publishes all feature values for each primary key in the feature table to the online store and automatically removes expired records. A record is considered expired if the record’s timestamp (in UTC) is more than the specified time to live duration in the past. Refer to cloud-specific documentation for details on time-to-live.

The online store supports primary key lookup and automatically retrieves the feature value with the latest timestamp.

In window mode, you must provide a value for time to live (ttl) in the OnlineStoreSpec when you create the online store. The ttl cannot be changed once set. All subsequent publish calls inherit the ttl and are not required to explicitly define it in the OnlineStoreSpec.

Notebook example: Time series feature table

These example notebooks illustrate point-in-time lookups on time series feature tables.

Use this notebook in workspaces enabled for Unity Catalog.

Time series feature table example notebook (Unity Catalog)

Open notebook in new tab Copy link for import

The following notebook is designed for workspaces that are not enabled for Unity Catalog. It uses the Workspace Feature Store.

Time series feature table example notebook (workspaces not enabled for Unity Catalog)

Open notebook in new tab Copy link for import