Skip to content

feat: Add ScyllaDB online store with vector search - #6508

Merged
franciscojavierarceo merged 14 commits into
feast-dev:masterfrom
zseta:scylladb-native-online-store
Jul 8, 2026
Merged

feat: Add ScyllaDB online store with vector search#6508
franciscojavierarceo merged 14 commits into
feast-dev:masterfrom
zseta:scylladb-native-online-store

Conversation

@zseta

@zseta zseta commented Jun 12, 2026

Copy link
Copy Markdown
Contributor

Summary

Adds ScyllaDB online store to Feast, including support for vector similarity search via ScyllaDB's built-in vector type and HNSW ANN index.

What's included

New online store

  • feast/infra/online_stores/scylladb_online_store/scylladb.py
    • ScyllaDBOnlineStore implementing OnlineStore
    • ScyllaDBOnlineStoreConfig with all connection and tuning options
    • Full online_write_batch / online_read support
    • retrieve_online_documents_v2 for ANN vector search
    • Native TTL support via CQL's USING TTL — automatically expires rows based on FeatureView.ttl
    • LOCAL_QUORUM as default consistency level for both reads and writes

Configuration

online_store:
  type: scylladb
  hosts: ["host_url1", "host_url2"]
  keyspace: feast_keyspace
  local_dc: AWS_US_EAST_1

Vector search

Tag any Field with vector_index: "true" and dimensions: "N" to enable ANN search:

Field(name="embedding", dtype=Array(Float32), tags={"vector_index": "true", "dimensions": "768"})

Then query with retrieve_online_documents_v2:

results = store.retrieve_online_documents_v2(
    feature_view=fv,
    requested_feature="embedding",
    query=[0.1, 0.2, ...],
    top_k=10,
)

Backed by a vector column with a CUSTOM INDEX USING 'vector_index' (HNSW).

TTL

When a FeatureView has a ttl set, rows are written with CQL's USING TTL clause so ScyllaDB automatically expires stale data. ScyllaDB does not support TTL on vector-indexed columns (it will be fixed in a future release). feast apply will raise an error if user wants to setttl for a vector column.

Changes

File Change
sdk/python/feast/infra/online_stores/scylladb_online_store/scylladb.py New — full implementation
sdk/python/feast/infra/online_stores/scylladb_online_store/__init__.py New — package marker
sdk/python/feast/repo_config.py Register "scylladb" short-name alias
pyproject.toml Add scylladb optional extra with scylla-driver>=3.28.0,<4
docs/reference/online-stores/scylladb.md New reference doc
docs/reference/online-stores/overview.md Add ScyllaDB to online store overview
docs/reference/alpha-vector-database.md Add ScyllaDB to vector store matrix
docs/getting-started/genai.md Mention ScyllaDB vector support
sdk/python/tests/integration/online_store/test_scylladb_online_store.py Integration tests
sdk/python/tests/universal/feature_repos/universal/online_store/scylladb.py Test store creator

Notes

  • Uses scylla-driver
  • Vector search requires ScyllaDB with the Vector Search feature enabled.
  • The keyspace is created automatically if it does not exist (NetworkTopologyStrategy, replication_factor: 3).

@zseta zseta changed the title feat: add ScyllaDB online store with vector search feat: Add ScyllaDB online store with vector search Jun 18, 2026
zseta added 8 commits June 18, 2026 13:58
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
@zseta
zseta force-pushed the scylladb-native-online-store branch from c23affa to 08573c6 Compare June 18, 2026 11:59
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
@zseta
zseta force-pushed the scylladb-native-online-store branch from 4a97639 to de1a61b Compare June 18, 2026 12:10
@zseta
zseta marked this pull request as ready for review June 18, 2026 12:22
@zseta
zseta requested a review from a team as a code owner June 18, 2026 12:22
@patelchaitany

patelchaitany commented Jun 19, 2026

Copy link
Copy Markdown
Contributor

Nice work! One thing to keep in mind - the value column is a protobuf BLOB, which means CQL can't filter on it (e.g., WHERE price > 100).

When #6121 (OpenAI-compatible search with metadata filtering) lands, this store will need value_num DOUBLE and value_text TEXT columns alongside the existing value BLOB - same pattern SQLite and Postgres use in that PR.

Not a blocker, just flagging so the schema is easy to extend later.

ref: #6121

zseta added 3 commits June 23, 2026 11:14
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
Signed-off-by: Attila Toth <attila.toth@scylladb.com>
@zseta

zseta commented Jun 23, 2026

Copy link
Copy Markdown
Contributor Author

@patelchaitany After discussion, I implemented the online store with a single table (with optional vector column). Let me know if anything else is missing.

@franciscojavierarceo

Copy link
Copy Markdown
Member

@zseta can you resolve conflicts? sorry

@zseta

zseta commented Jul 8, 2026

Copy link
Copy Markdown
Contributor Author

@franciscojavierarceo done

@franciscojavierarceo
franciscojavierarceo merged commit 1669661 into feast-dev:master Jul 8, 2026
24 checks passed
franciscojavierarceo pushed a commit that referenced this pull request Jul 20, 2026
# [0.65.0](v0.64.0...v0.65.0) (2026-07-20)

### Bug Fixes

* add debug logging for FIPS mode detection fallback ([6c1b24e](6c1b24e))
* Build embedded UI from local source ([#6525](#6525)) ([3500349](3500349))
* Bump decommissioned Snowflake Python UDF runtime from 3.9 to 3.10 ([#6606](#6606)) ([#6608](#6608)) ([10341e4](10341e4))
* configure FIPS-compliant gRPC cipher suites for offline server ([6bc80a2](6bc80a2))
* Correct Flink PyArrow dependency constraints ([#6604](#6604)) ([70a9751](70a9751))
* Fix ValueError in signal handling for Trino worker threads ([#6428](#6428)) ([506d919](506d919))
* Fixed monitoring page issues ([7946018](7946018))
* Make pytest config compatible with newer pytest ([#5779](#5779)) ([a57ea33](a57ea33))
* Replace comma with space in DynamoDB-incompatible label tag value ([51e3a16](51e3a16))
* Resolve UI build warnings ([#6529](#6529)) ([abe92af](abe92af))
* Unblock nightly UI build ([#6570](#6570)) ([f296d4b](f296d4b))
* Use LONGBLOB for SQL registry proto columns on MySQL ([#6566](#6566)) ([7e4beb2](7e4beb2))

### Features

* Add click-to-zoom lightbox for blog post images ([#6575](#6575)) ([1cb23fd](1cb23fd))
* Add dark mode support to website and blog ([#6589](#6589)) ([7358fb8](7358fb8))
* Add OnlineStore for Aerospike ([#6532](#6532)) ([9cd35e1](9cd35e1))
* Add OpenLineage Consumer to Feast - receive, store, and visualize cross-producer lineage ([#6549](#6549)) ([a834126](a834126))
* Add registry list feature views by updated since ([#6092](#6092)) ([#6093](#6093)) ([006c606](006c606))
* Add ScyllaDB online store with vector search ([#6508](#6508)) ([1669661](1669661))
* Added compute and jobs UI ([ba2c05c](ba2c05c))
* Added Iceberg REST Catalog data source support ([e0a8573](e0a8573))
* Bring Your Own Spark - SparkApplication ([#6550](#6550)) ([dcd496f](dcd496f))
* **cassandra:** Add multi-DC support via per-datacenter execution profiles ([#6434](#6434)) ([0de9196](0de9196))
* Enhanced data source creation as a visual catalog with type-specific forms ([#6557](#6557)) ([d6acbba](d6acbba))
* Enhanced datasets UI functionality ([de11152](de11152))
* Implement RegistryServer.Proto RPC with RBAC-filtered response ([#6558](#6558)) ([#6552](#6552)) ([0d02614](0d02614))
* New zoned timestamp feature type ([#6536](#6536)) ([#6537](#6537)) ([eb042f0](eb042f0))
* **operator:** Auto-create RBAC for spark_application batch engine ([#6597](#6597)) ([f487b37](f487b37))
* **operator:** integrate cluster TLS profile for OCP 5.0 compliance ([43263a6](43263a6))
* Permissions CRUD UI and OIDC auth integration in UI ([6511da1](6511da1))
* Retrieve historical features from BigQuery without entity_df ([#6569](#6569)) ([cd5f6bb](cd5f6bb)), closes [#6558](#6558) [#6552](#6552)
* **spark:** SparkSource query+path and pre-computed offline read for BatchFeatureView ([#6440](#6440)) ([4dc8757](4dc8757))

### BREAKING CHANGES

* total_timeout_ms is renamed to batch_total_timeout_ms. Config files using the old name must be updated. No default value change.

Docs updated (reference + perf-tuning guide) with a short explainer on the per-attempt vs total deadline distinction. Two new unit tests pin the policy wiring: socket_timeout_ms propagates to all three scopes, and is omitted (not injected as None) when unset.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* refactor(aerospike): use MAP_KEY_ORDERED, KEY_DIGEST, and instance-scoped client

Cheap-win cleanups flagged in review, all touching the same small patch of write-path and lifecycle code.

* Map CDTs are now created with MAP_KEY_ORDERED. map_get_by_key / map_remove_by_key on an ordered map are O(log N) in the map size instead of O(N); matters on reads of wide feature views and on the update() background scan (which walks every record in the project's set).

* Writes drop POLICY_KEY_SEND and rely on the client default (POLICY_KEY_DIGEST). The serialized entity key is no longer stored alongside each record, saving per-record storage the read path never consumes (batch_operate preserves request order; results are paired back by zip in online_read).

* _client moves from a class attribute to an instance attribute (set in __init__). Previously two AerospikeOnlineStore instances could share the cached client through class state until one wrote self._client. With the instance attribute the state is always per-instance from construction.

* Drop MongoDB references from class docstrings and comments (they referred to how the storage layout was derived rather than documenting current behavior). Also rewrite the _build_batch_writes docstring to describe the policies applied on the write path.

Unit test assertions for the write-path record are updated: bw.policy is now None (client default applies) and map ops carry map_policy={'map_order': MAP_KEY_ORDERED}. All three docker-backed integration tests still pass end-to-end (cross-FV upsert, update() background scan, full feature-store round-trip), so the read/write shape survives the ordering and policy changes against a real server.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* feat(aerospike): add per-FV namespace/set overrides and prewriting hook

Adds three configuration knobs to AerospikeOnlineStoreConfig:

- namespace_overrides: pin individual feature views to a different
  Aerospike namespace (e.g. RAM-only vs. SSD-backed) without splitting
  the project across stores.
- set_overrides: place a feature view in its own set so admin ops on
  it (truncate, scan-based deletes during `feast apply`) do not touch
  records of other views.
- prewriting_hook: import-string-resolved callable invoked once per
  online_write_batch with the rows about to be written, returning the
  rows that actually go on the wire. Resolved and cached on first use;
  returning [] short-circuits the wire call.

Read, write, update and teardown paths all honour the per-FV ns/set
resolution. update() groups dropped feature views by their resolved
(ns, set) pair and issues one background scan per group. teardown()
truncates every unique (ns, set) pair the project may have written to,
including the store-level default.

Adds 22 unit tests for the new behaviour and updates 3 existing call
sites of _build_batch_writes for the new namespace= parameter. Adds a
sample hook module under examples/online_store/aerospike_overrides_and_hooks/
and corresponding sections in docs/reference/online-stores/aerospike.md.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* test: update aerospike image tag

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* chore: sync README template and secrets baseline after master merge

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* chore: fix secrets baseline line number for v1 operator types

Adding aerospike to the feast-operator enum shifted the allowlisted
SecretRef entry in api/v1/featurestore_types.go by one line.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* docs: update aerospike docs

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* fix(aerospike): wire batch max_retries and fix empty projection handling

Copilot review feedback on PR #6532:

- Add max_retries to the batch client policy (batch_operate/batch_write path)
- Treat empty projected feature maps as present FV slots (is not None)
- Return {} from _normalize_projected_features([]) instead of None
- Fix projection unit test mock/assertions
- Correct prewriting_hook config docstring

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* style(aerospike): format online_read docs assignment for ruff

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* chore: update pixi.lock for aerospike optional extra

Regenerate the v6 lockfile with Pixi v0.63.1 after adding the aerospike extra to pyproject.toml.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* fix(aerospike): add client init lock and batch chunking

Guard lazy client creation with a lock to avoid connection leaks under concurrent first use, and chunk batch reads/writes by batch_max_records so large materializations stay under Aerospike server batch limits.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>
aniketpalu pushed a commit to aniketpalu/feast that referenced this pull request Jul 28, 2026
# [0.65.0](feast-dev/feast@v0.64.0...v0.65.0) (2026-07-20)

### Bug Fixes

* add debug logging for FIPS mode detection fallback ([6c1b24e](feast-dev@6c1b24e))
* Build embedded UI from local source ([feast-dev#6525](feast-dev#6525)) ([3500349](feast-dev@3500349))
* Bump decommissioned Snowflake Python UDF runtime from 3.9 to 3.10 ([feast-dev#6606](feast-dev#6606)) ([feast-dev#6608](feast-dev#6608)) ([10341e4](feast-dev@10341e4))
* configure FIPS-compliant gRPC cipher suites for offline server ([6bc80a2](feast-dev@6bc80a2))
* Correct Flink PyArrow dependency constraints ([feast-dev#6604](feast-dev#6604)) ([70a9751](feast-dev@70a9751))
* Fix ValueError in signal handling for Trino worker threads ([feast-dev#6428](feast-dev#6428)) ([506d919](feast-dev@506d919))
* Fixed monitoring page issues ([7946018](feast-dev@7946018))
* Make pytest config compatible with newer pytest ([feast-dev#5779](feast-dev#5779)) ([a57ea33](feast-dev@a57ea33))
* Replace comma with space in DynamoDB-incompatible label tag value ([51e3a16](feast-dev@51e3a16))
* Resolve UI build warnings ([feast-dev#6529](feast-dev#6529)) ([abe92af](feast-dev@abe92af))
* Unblock nightly UI build ([feast-dev#6570](feast-dev#6570)) ([f296d4b](feast-dev@f296d4b))
* Use LONGBLOB for SQL registry proto columns on MySQL ([feast-dev#6566](feast-dev#6566)) ([7e4beb2](feast-dev@7e4beb2))

### Features

* Add click-to-zoom lightbox for blog post images ([feast-dev#6575](feast-dev#6575)) ([1cb23fd](feast-dev@1cb23fd))
* Add dark mode support to website and blog ([feast-dev#6589](feast-dev#6589)) ([7358fb8](feast-dev@7358fb8))
* Add OnlineStore for Aerospike ([feast-dev#6532](feast-dev#6532)) ([9cd35e1](feast-dev@9cd35e1))
* Add OpenLineage Consumer to Feast - receive, store, and visualize cross-producer lineage ([feast-dev#6549](feast-dev#6549)) ([a834126](feast-dev@a834126))
* Add registry list feature views by updated since ([feast-dev#6092](feast-dev#6092)) ([feast-dev#6093](feast-dev#6093)) ([006c606](feast-dev@006c606))
* Add ScyllaDB online store with vector search ([feast-dev#6508](feast-dev#6508)) ([1669661](feast-dev@1669661))
* Added compute and jobs UI ([ba2c05c](feast-dev@ba2c05c))
* Added Iceberg REST Catalog data source support ([e0a8573](feast-dev@e0a8573))
* Bring Your Own Spark - SparkApplication ([feast-dev#6550](feast-dev#6550)) ([dcd496f](feast-dev@dcd496f))
* **cassandra:** Add multi-DC support via per-datacenter execution profiles ([feast-dev#6434](feast-dev#6434)) ([0de9196](feast-dev@0de9196))
* Enhanced data source creation as a visual catalog with type-specific forms ([feast-dev#6557](feast-dev#6557)) ([d6acbba](feast-dev@d6acbba))
* Enhanced datasets UI functionality ([de11152](feast-dev@de11152))
* Implement RegistryServer.Proto RPC with RBAC-filtered response ([feast-dev#6558](feast-dev#6558)) ([feast-dev#6552](feast-dev#6552)) ([0d02614](feast-dev@0d02614))
* New zoned timestamp feature type ([feast-dev#6536](feast-dev#6536)) ([feast-dev#6537](feast-dev#6537)) ([eb042f0](feast-dev@eb042f0))
* **operator:** Auto-create RBAC for spark_application batch engine ([feast-dev#6597](feast-dev#6597)) ([f487b37](feast-dev@f487b37))
* **operator:** integrate cluster TLS profile for OCP 5.0 compliance ([43263a6](feast-dev@43263a6))
* Permissions CRUD UI and OIDC auth integration in UI ([6511da1](feast-dev@6511da1))
* Retrieve historical features from BigQuery without entity_df ([feast-dev#6569](feast-dev#6569)) ([cd5f6bb](feast-dev@cd5f6bb)), closes [feast-dev#6558](feast-dev#6558) [feast-dev#6552](feast-dev#6552)
* **spark:** SparkSource query+path and pre-computed offline read for BatchFeatureView ([feast-dev#6440](feast-dev#6440)) ([4dc8757](feast-dev@4dc8757))

### BREAKING CHANGES

* total_timeout_ms is renamed to batch_total_timeout_ms. Config files using the old name must be updated. No default value change.

Docs updated (reference + perf-tuning guide) with a short explainer on the per-attempt vs total deadline distinction. Two new unit tests pin the policy wiring: socket_timeout_ms propagates to all three scopes, and is omitted (not injected as None) when unset.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* refactor(aerospike): use MAP_KEY_ORDERED, KEY_DIGEST, and instance-scoped client

Cheap-win cleanups flagged in review, all touching the same small patch of write-path and lifecycle code.

* Map CDTs are now created with MAP_KEY_ORDERED. map_get_by_key / map_remove_by_key on an ordered map are O(log N) in the map size instead of O(N); matters on reads of wide feature views and on the update() background scan (which walks every record in the project's set).

* Writes drop POLICY_KEY_SEND and rely on the client default (POLICY_KEY_DIGEST). The serialized entity key is no longer stored alongside each record, saving per-record storage the read path never consumes (batch_operate preserves request order; results are paired back by zip in online_read).

* _client moves from a class attribute to an instance attribute (set in __init__). Previously two AerospikeOnlineStore instances could share the cached client through class state until one wrote self._client. With the instance attribute the state is always per-instance from construction.

* Drop MongoDB references from class docstrings and comments (they referred to how the storage layout was derived rather than documenting current behavior). Also rewrite the _build_batch_writes docstring to describe the policies applied on the write path.

Unit test assertions for the write-path record are updated: bw.policy is now None (client default applies) and map ops carry map_policy={'map_order': MAP_KEY_ORDERED}. All three docker-backed integration tests still pass end-to-end (cross-FV upsert, update() background scan, full feature-store round-trip), so the read/write shape survives the ordering and policy changes against a real server.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* feat(aerospike): add per-FV namespace/set overrides and prewriting hook

Adds three configuration knobs to AerospikeOnlineStoreConfig:

- namespace_overrides: pin individual feature views to a different
  Aerospike namespace (e.g. RAM-only vs. SSD-backed) without splitting
  the project across stores.
- set_overrides: place a feature view in its own set so admin ops on
  it (truncate, scan-based deletes during `feast apply`) do not touch
  records of other views.
- prewriting_hook: import-string-resolved callable invoked once per
  online_write_batch with the rows about to be written, returning the
  rows that actually go on the wire. Resolved and cached on first use;
  returning [] short-circuits the wire call.

Read, write, update and teardown paths all honour the per-FV ns/set
resolution. update() groups dropped feature views by their resolved
(ns, set) pair and issues one background scan per group. teardown()
truncates every unique (ns, set) pair the project may have written to,
including the store-level default.

Adds 22 unit tests for the new behaviour and updates 3 existing call
sites of _build_batch_writes for the new namespace= parameter. Adds a
sample hook module under examples/online_store/aerospike_overrides_and_hooks/
and corresponding sections in docs/reference/online-stores/aerospike.md.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* test: update aerospike image tag

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* chore: sync README template and secrets baseline after master merge

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* chore: fix secrets baseline line number for v1 operator types

Adding aerospike to the feast-operator enum shifted the allowlisted
SecretRef entry in api/v1/featurestore_types.go by one line.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* docs: update aerospike docs

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* fix(aerospike): wire batch max_retries and fix empty projection handling

Copilot review feedback on PR feast-dev#6532:

- Add max_retries to the batch client policy (batch_operate/batch_write path)
- Treat empty projected feature maps as present FV slots (is not None)
- Return {} from _normalize_projected_features([]) instead of None
- Fix projection unit test mock/assertions
- Correct prewriting_hook config docstring

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* style(aerospike): format online_read docs assignment for ruff

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* chore: update pixi.lock for aerospike optional extra

Regenerate the v6 lockfile with Pixi v0.63.1 after adding the aerospike extra to pyproject.toml.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>

* fix(aerospike): add client init lock and batch chunking

Guard lazy client creation with a lock to avoid connection leaks under concurrent first use, and chunk batch reads/writes by batch_max_records so large materializations stay under Aerospike server batch limits.

Signed-off-by: Valentyn Kahamlyk <valentin.kagamlyk@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants