GeezerVison

Summary

Addresses several correctness and performance issues in the Redis online store
identified in #3596.

Changes

1. Batched get_online_features pipeline (read latency: O(N_fv) → O(1))

The base class get_online_features calls online_read once per feature view,
each opening and executing its own Redis pipeline. Since all feature views for
the same entity already share one Redis hash key, all HMGET commands across
every feature view can be batched into a single pipeline execution.

RedisOnlineStore now overrides get_online_features to pre-compute all
entity keys and hash field keys, issue one pipeline.execute() for everything,
then demultiplex results back per feature view. For a request spanning N feature
views this cuts Redis round trips from N to 1.

2. Fixed delete_table N+1 query pattern

The previous implementation called client.hgetall(key) synchronously inside
the scan_iter loop — one blocking round trip per entity key. Replaced with a
three-phase approach:

• Phase 1: collect all matching keys from SCAN (no per-key round trips)
• Phase 2: pipeline HKEYS for all collected keys (1 round trip)
• Phase 3: pipeline all HDEL/DEL commands (1 round trip)

3. skip_dedup config option for write throughput

Added skip_dedup: bool = False to RedisOnlineStoreConfig. When enabled,
online_write_batch skips the first timestamp-read pipeline and writes all
rows in a single pass, halving round trips. Intended for initial bulk loads or
append-only pipelines where concurrent out-of-order writes are not a concern.

4. online_write_batch_async implemented

Previously raised NotImplementedError. Now fully implemented using the
existing _get_client_async infrastructure, with skip_dedup support.

Files Changed

• sdk/python/feast/infra/online_stores/redis.py — only file modified; base
  class OnlineStore and all other store implementations are untouched.

Performance Impact

Related

• Fixes Redis Online Store adapter is not production-ready #3596