Redis Feature Store Demo
A small fraud-scoring feature store. Each user is one Redis hash
at __KEY_PREFIX__{id} with a batch-materialized
batch half (daily aggregates,
24-hour key-level EXPIRE) and a streaming
streaming half (real-time
signals, __STREAM_TTL__s per-field HEXPIRE).
Inference reads any subset with one HMGET; batch
scoring pipelines HMGET across N users.
Store state
Materialize batch features
Calls HSET + EXPIRE for each user
through one Predis pipeline — the whole batch
ships in one round trip.
Drop the TTL to e.g. 30 s and watch entities disappear on schedule — the same thing that happens if a daily refresher fails.
Streaming worker
Picks __USERS_PER_TICK__ users per tick, writes the
streaming features, applies HEXPIRE
__STREAM_TTL__s per field. Pause it and the
streaming fields drop out via per-field TTL while the batch
fields stay populated.
Inference read (HMGET)
Pick a user and a feature subset. One HMGET
round trip returns whatever the model needs.
Feature subset
Tick to include in the HMGET. Per-field TTL is
shown next to each field in the result table.
Pick a user and click Read features.
Batch scoring
Pipelined HMGET across N random users via
Predis pipelines. One network round trip for the
whole batch.
(no batch read yet)
Inspect one user
HGETALL plus per-field HTTL and
key-level TTL. Useful for spotting which
streaming fields have already expired.
(pick a user and click Inspect)