Feldera
Engineering Blog
Incremental computation, real-time feature engineering, streaming SQL, and lessons from building a feature engine from scratch.
Why Batch Feature Pipelines Fail Real-Time Fraud Models
Nightly batches calculate yesterday's features while fraud unfolds in seconds. Here is the structural gap and how incremental computation closes it.
Read article
SQL on Streams: How to Write Incremental Queries That Work
Standard SQL treats tables as static. Stream SQL changes that model. Learn to write queries that stay accurate as new data arrives.
Read more
When Is Your Feature Store Actually Fresh?
Feature stores often measure freshness from write time instead of event time. That gap is where staleness remains hidden.
Read more
CDC Integration Patterns for Real-Time Feature Engineering
CDC converts a Postgres replication slot or MySQL binlog into a continuous event stream for ML-ready features.
Read more
Streaming Joins at Scale: Lessons from Production
Joining unbounded streams becomes difficult when late arrivals, state size, and exactly-once semantics enter the picture. Production lessons.
Read more
Designing Low-Latency Feature Pipelines for Anomaly Detection
Build anomaly detection features that reflect current behavior, with pipeline choices that reduce update latency while preserving correctness.
Read more
The Hidden Cost of Recomputing Features from Scratch
Batch recomputation rereads the full history. As data grows, costs compound. Incremental maintenance changes the equation.
Read more
How Stale Features Kill Recommendation CTR
A recommendation model trained on fresh data can fly blind when served stale features, especially session-level features batch pipelines cannot supply.
Read more
Feldera Open Beta: Lessons from Building a Streaming Compute Engine
We built Feldera's incremental SQL engine from first principles, ran production workloads through beta and launch, and share what surprised us and what we rethought.
Read more
Kafka to Feature Store: Bridging the Stream-Batch Gap
Kafka streams events in real time, but feature stores need continuously updated aggregations without batch recompute jobs.
Read more
Incremental Computation 101: What It Is and Why ML Teams Need It
Incremental computation updates results from changes, not the full dataset, giving ML feature pipelines second-level instead of hour-level freshness.
Read more
Building a Fraud Score Pipeline with Sub-Second Feature Freshness
Stale velocity features can undermine fraud models. We built a pipeline that keeps transaction velocity counts current within seconds of every new event.
Read more
Incremental SQL: Computing Aggregations Without Reprocessing History
Incremental view maintenance updates SQL aggregation results as data changes, avoiding repeated processing of growing tables.
Read more
The Case for Incremental Computation in Modern ML Stacks
Why incremental computation fits modern ML pipelines, improves on batch recomputation, and belongs in today's data stack.
Read more