Skip to content

Commit 8f591a2

Browse files
nanohannohkuepers
nanohanno
and
hkuepers
authored
feat: Use ASOF JOIN in Snowflake offline store query (feast-dev#4850)
* Use ASOF JOIN in Snowflake offline store query Signed-off-by: hkuepers <hanno.kuepers@ratepay.com> * Fix Snowflake query template for entityless feature views Signed-off-by: hkuepers <hanno.kuepers@ratepay.com> * Remove quotes on subquery in snowflake template Signed-off-by: hkuepers <hanno.kuepers@ratepay.com> * Use __subquery in Snowflake template for preparation Signed-off-by: hkuepers <hanno.kuepers@ratepay.com> * Fix deduplication in Snowflake query string Signed-off-by: hkuepers <hanno.kuepers@ratepay.com> * Use event_timestamp in ttl cte Signed-off-by: hkuepers <hanno.kuepers@ratepay.com> --------- Signed-off-by: hkuepers <hanno.kuepers@ratepay.com> Co-authored-by: hkuepers <hanno.kuepers@ratepay.com>
1 parent 3dbd58b commit 8f591a2

File tree

1 file changed

+44
-86
lines changed

1 file changed

+44
-86
lines changed

‎sdk/python/feast/infra/offline_stores/snowflake.py

+44-86
Original file line numberDiff line numberDiff line change
@@ -716,8 +716,8 @@ def _get_entity_df_event_timestamp_range(
716716

717717
MULTIPLE_FEATURE_VIEW_POINT_IN_TIME_JOIN = """
718718
/*
719-
Compute a deterministic hash for the `left_table_query_string` that will be used throughout
720-
all the logic as the field to GROUP BY the data
719+
0. Compute a deterministic hash for the `left_table_query_string` that will be used throughout
720+
all the logic as the field to GROUP BY the data.
721721
*/
722722
WITH "entity_dataframe" AS (
723723
SELECT *,
@@ -739,6 +739,10 @@ def _get_entity_df_event_timestamp_range(
739739
740740
{% for featureview in featureviews %}
741741
742+
/*
743+
1. Only select the required columns with entities of the featureview.
744+
*/
745+
742746
"{{ featureview.name }}__entity_dataframe" AS (
743747
SELECT
744748
{{ featureview.entities | map('tojson') | join(', ')}}{% if featureview.entities %},{% else %}{% endif %}
@@ -752,20 +756,7 @@ def _get_entity_df_event_timestamp_range(
752756
),
753757
754758
/*
755-
This query template performs the point-in-time correctness join for a single feature set table
756-
to the provided entity table.
757-
758-
1. We first join the current feature_view to the entity dataframe that has been passed.
759-
This JOIN has the following logic:
760-
- For each row of the entity dataframe, only keep the rows where the `timestamp_field`
761-
is less than the one provided in the entity dataframe
762-
- If there a TTL for the current feature_view, also keep the rows where the `timestamp_field`
763-
is higher the the one provided minus the TTL
764-
- For each row, Join on the entity key and retrieve the `entity_row_unique_id` that has been
765-
computed previously
766-
767-
The output of this CTE will contain all the necessary information and already filtered out most
768-
of the data that is not relevant.
759+
2. Use subquery to prepare event_timestamp, created_timestamp, entity columns and feature columns.
769760
*/
770761
771762
"{{ featureview.name }}__subquery" AS (
@@ -777,94 +768,61 @@ def _get_entity_df_event_timestamp_range(
777768
"{{ feature }}" as {% if full_feature_names %}"{{ featureview.name }}__{{featureview.field_mapping.get(feature, feature)}}"{% else %}"{{ featureview.field_mapping.get(feature, feature) }}"{% endif %}{% if loop.last %}{% else %}, {% endif %}
778769
{% endfor %}
779770
FROM {{ featureview.table_subquery }}
780-
WHERE "{{ featureview.timestamp_field }}" <= '{{ featureview.max_event_timestamp }}'
781-
{% if featureview.ttl == 0 %}{% else %}
782-
AND "{{ featureview.timestamp_field }}" >= '{{ featureview.min_event_timestamp }}'
783-
{% endif %}
784-
),
785-
786-
"{{ featureview.name }}__base" AS (
787-
SELECT
788-
"subquery".*,
789-
"entity_dataframe"."entity_timestamp",
790-
"entity_dataframe"."{{featureview.name}}__entity_row_unique_id"
791-
FROM "{{ featureview.name }}__subquery" AS "subquery"
792-
INNER JOIN "{{ featureview.name }}__entity_dataframe" AS "entity_dataframe"
793-
ON TRUE
794-
AND "subquery"."event_timestamp" <= "entity_dataframe"."entity_timestamp"
795-
796-
{% if featureview.ttl == 0 %}{% else %}
797-
AND "subquery"."event_timestamp" >= TIMESTAMPADD(second,-{{ featureview.ttl }},"entity_dataframe"."entity_timestamp")
798-
{% endif %}
799-
800-
{% for entity in featureview.entities %}
801-
AND "subquery"."{{ entity }}" = "entity_dataframe"."{{ entity }}"
802-
{% endfor %}
803771
),
804772
805773
/*
806-
2. If the `created_timestamp_column` has been set, we need to
807-
deduplicate the data first. This is done by calculating the
808-
`MAX(created_at_timestamp)` for each event_timestamp.
809-
We then join the data on the next CTE
774+
3. If the `created_timestamp_column` has been set, we need to
775+
deduplicate the data first. This is done by calculating the
776+
`MAX(created_at_timestamp)` for each event_timestamp and joining back on the subquery.
777+
Otherwise, the ASOF JOIN can have unstable side effects
778+
https://docs.snowflake.com/en/sql-reference/constructs/asof-join#expected-behavior-when-ties-exist-in-the-right-table
810779
*/
780+
811781
{% if featureview.created_timestamp_column %}
812782
"{{ featureview.name }}__dedup" AS (
813-
SELECT
814-
"{{featureview.name}}__entity_row_unique_id",
815-
"event_timestamp",
816-
MAX("created_timestamp") AS "created_timestamp"
817-
FROM "{{ featureview.name }}__base"
818-
GROUP BY "{{featureview.name}}__entity_row_unique_id", "event_timestamp"
783+
SELECT *
784+
FROM "{{ featureview.name }}__subquery"
785+
INNER JOIN (
786+
SELECT
787+
{{ featureview.entities | map('tojson') | join(', ')}}{% if featureview.entities %},{% else %}{% endif %}
788+
"event_timestamp",
789+
MAX("created_timestamp") AS "created_timestamp"
790+
FROM "{{ featureview.name }}__subquery"
791+
GROUP BY {{ featureview.entities | map('tojson') | join(', ')}}{% if featureview.entities %},{% else %}{% endif %} "event_timestamp"
792+
)
793+
USING({{ featureview.entities | map('tojson') | join(', ')}}{% if featureview.entities %},{% else %}{% endif %} "event_timestamp", "created_timestamp")
819794
),
820795
{% endif %}
821796
822797
/*
823-
3. The data has been filtered during the first CTE "*__base"
824-
Thus we only need to compute the latest timestamp of each feature.
798+
4. Make ASOF JOIN of deduplicated feature CTE on reduced entity dataframe.
825799
*/
826-
"{{ featureview.name }}__latest" AS (
800+
801+
"{{ featureview.name }}__asof_join" AS (
827802
SELECT
828-
"event_timestamp",
829-
{% if featureview.created_timestamp_column %}"created_timestamp",{% endif %}
830-
"{{featureview.name}}__entity_row_unique_id"
831-
FROM
832-
(
833-
SELECT *,
834-
ROW_NUMBER() OVER(
835-
PARTITION BY "{{featureview.name}}__entity_row_unique_id"
836-
ORDER BY "event_timestamp" DESC{% if featureview.created_timestamp_column %},"created_timestamp" DESC{% endif %}
837-
) AS "row_number"
838-
FROM "{{ featureview.name }}__base"
839-
{% if featureview.created_timestamp_column %}
840-
INNER JOIN "{{ featureview.name }}__dedup"
841-
USING ("{{featureview.name}}__entity_row_unique_id", "event_timestamp", "created_timestamp")
842-
{% endif %}
843-
)
844-
WHERE "row_number" = 1
803+
e.*,
804+
v.*
805+
FROM "{{ featureview.name }}__entity_dataframe" e
806+
ASOF JOIN {% if featureview.created_timestamp_column %}"{{ featureview.name }}__dedup"{% else %}"{{ featureview.name }}__subquery"{% endif %} v
807+
MATCH_CONDITION (e."entity_timestamp" >= v."event_timestamp")
808+
{% if featureview.entities %} USING({{ featureview.entities | map('tojson') | join(', ')}}) {% endif %}
845809
),
846810
847811
/*
848-
4. Once we know the latest value of each feature for a given timestamp,
849-
we can join again the data back to the original "base" dataset
812+
5. If TTL is configured filter the CTE to remove rows where the feature values are older than the configured ttl.
850813
*/
851-
"{{ featureview.name }}__cleaned" AS (
852-
SELECT "base".*
853-
FROM "{{ featureview.name }}__base" AS "base"
854-
INNER JOIN "{{ featureview.name }}__latest"
855-
USING(
856-
"{{featureview.name}}__entity_row_unique_id",
857-
"event_timestamp"
858-
{% if featureview.created_timestamp_column %}
859-
,"created_timestamp"
860-
{% endif %}
861-
)
862-
){% if loop.last %}{% else %}, {% endif %}
863814
815+
"{{ featureview.name }}__ttl" AS (
816+
SELECT *
817+
FROM "{{ featureview.name }}__asof_join"
818+
{% if featureview.ttl == 0 %}{% else %}
819+
WHERE "event_timestamp" >= TIMESTAMPADD(second,-{{ featureview.ttl }},"entity_timestamp")
820+
{% endif %}
821+
){% if loop.last %}{% else %}, {% endif %}
864822
865823
{% endfor %}
866824
/*
867-
Joins the outputs of multiple time travel joins to a single table.
825+
Join the outputs of multiple time travel joins to a single table.
868826
The entity_dataframe dataset being our source of truth here.
869827
*/
870828
@@ -877,7 +835,7 @@ def _get_entity_df_event_timestamp_range(
877835
{% for feature in featureview.features %}
878836
,{% if full_feature_names %}"{{ featureview.name }}__{{featureview.field_mapping.get(feature, feature)}}"{% else %}"{{ featureview.field_mapping.get(feature, feature) }}"{% endif %}
879837
{% endfor %}
880-
FROM "{{ featureview.name }}__cleaned"
881-
) "{{ featureview.name }}__cleaned" USING ("{{featureview.name}}__entity_row_unique_id")
838+
FROM "{{ featureview.name }}__ttl"
839+
) "{{ featureview.name }}__ttl" USING ("{{featureview.name}}__entity_row_unique_id")
882840
{% endfor %}
883841
"""

0 commit comments

Comments
 (0)