Skip to content

v0.2.0

Latest

Choose a tag to compare

@github-actions github-actions released this 18 Oct 12:20
· 1 commit to master since this release

OpenRec v0.2.0

Released: 2026-10-18

Coordinated source release of the refactored OpenRec recommendation stack, covering
standalone and cluster deployments. Exact component commits are recorded in
the manifest and component table.

Changes since v0.1.0

  • Online serving uses unified request triggers, aligned item/user graph channels,
    Elasticsearch-only recall and configurable virtual-thread graph execution.
  • Ranking separates model lifecycle, feature snapshots and inference execution,
    with batched scoring, cached projections and CPU/CUDA implementations.
  • Offline Spark jobs and streaming Flink/Spark jobs have explicit responsibility
    boundaries while preserving shared feature and versioned mutation contracts.
  • The console manages aligned recall publications and optional scheduled title
    embeddings. Java, Go and Python SDK diagnostics and error handling are aligned.
  • Initialization and regression tooling use the committed data/test fixture.
    Bootstrap build 19 organizes models by target/model family and makes language
    vectors optional for CPU-only initialization.

Installation and deployment modes

Check out v0.2.0, then run scripts/release/checkout-components.sh to obtain the
exact runtime/source-build dependencies. All OpenRec Java modules require Java 21.
Maven artifacts, the rec-algorithm wheel and Python SDK package are versioned
0.2.0; rebuild dependent services and update any external jar/wheel paths.

Standalone retains monitoring, entities and serving-graph operations and may
bypass ranking. Cluster enables Kafka projections, offline jobs, model/recall
lifecycle management and rank-engine. Recall storage is Elasticsearch-only in both
profiles; Redis remains the online entity/feature store. Remove obsolete
RECALL_STORE / recall.store settings.

Query encoding and scheduled title generation are opt-in. Enable the optional
vLLM service and matching model/preprocessing/dimensions before enabling title
vector publication or online query encoding. The default embedding configuration
uses Qwen3-Embedding-0.6B with 1024 dimensions; historical experiment vectors are
not interchangeable with these vectors.

Upgrade from v0.1.0

Upgrade the complete compatible component set. Back up persisted graph/DAG
configurations, release state, indexes, feature/model artifacts and checkpoints.
Source updates do not automatically migrate stored configuration overrides.

  • Graphs now expose requestTrigger, containing triggerItems, triggerUsers,
    triggerSelf and triggerQuery. Custom nodes using old standalone trigger ports
    must migrate with the graph implementation.
  • Persisted graph overrides must use item.search / user.search, SearchConfig,
    channels item_search / user_search and tables item-search / user-search.
    Republish text indexes; old sparse indexes are not reused. Persisted daily DAG
    configurations must replace sparse with text_search, including target_type=user
    for user search.
  • Item sequence embeddings use item-seq-emb and {scene}-item-seq-emb-index.
    The online query channel is query_title_emb, reading query-title-emb-index.
    Reinitialize or republish corresponding indexes before activating new graphs.
  • ALS neighbor edges use algorithm/channel user_als_u2u, table user-als-u2u
    and CSV user_als_u2u.csv (scene,left_user,right_user,score). ALS factors use
    user_als_emb.csv (scene,user,vector). Update saved DAG/graph overrides together.
  • Airflow publication task IDs use publish-<tableName>, such as
    publish-item-search and publish-item-seq-emb. Finish or stop old runs and
    reparse both DAGs before starting new runs. Historical task IDs are unchanged.
  • Bootstrap models live under rank/{item,user}/{lr,fm,lightgbm}/; recall tables
    live under recall/{item,user}/. Text CSVs are named text_search.csv.
    Build 19 defaults to 39 outputs; the companion model release retains verified
    optional title vectors for 40 checksummed outputs. The bundle manifest is the
    checksum authority. Always keep fitted feature sidecars with checkpoints.
  • Go/Python SDK callers must handle HTTP failures explicitly; they now return or
    raise errors rather than silently yielding empty responses. Go consumers use
    the additional go-client/v0.2.0 module tag.

Validation, publication and rollback

See the validation record for local release checks and
maintainer-reported prior CI/integration results. The maintainer triggers remote
acceptance workflows separately; prior successful runs are not presented as
verification of the new release commits.

This is a source release: GitHub Actions publishes source archives and SHA-256
checksums, not Maven Central/PyPI packages or versioned OCI images. The console
has no authentication/RBAC and the reference cluster is not a production HA
architecture. Existing documented deployment and latency boundaries still apply.

For rollback, restore the complete v0.1.0 component set together with its matching
saved graph/DAG configurations, artifact layout and index aliases. Retain old
indexes and backed-up state until acceptance completes; do not reuse a partially
migrated configuration with old binaries. Historical v0.1.0 tags are unchanged.