Product

A query engine for incremental data

Use standard SQL. Feldera turns it into an incremental operator graph, processing deltas rather than your full history.

Architecture

How the incremental engine works

As data arrives, Feldera computes only the output change, rather than rebuilding the result set. Operator nodes retain state incrementally.

Incremental SQL

The SQL you already know

Feldera accepts standard SELECT, GROUP BY, JOIN, and window functions, then builds and maintains the incremental plan for you.

Standard approach: full scan on every trigger

SELECT
  card_id,
  SUM(amount) AS spend_1h,
  COUNT(*) AS txn_count
FROM events
WHERE ts > NOW() - INTERVAL '1 hour'
GROUP BY card_id;
-- Scans full events table each run

Feldera incremental: delta-only updates

SELECT
  card_id,
  SUM(amount) AS spend_1h,
  COUNT(*) AS txn_count
FROM events
WHERE ts > NOW() - INTERVAL '1 hour'
GROUP BY card_id;
-- Feldera processes only new events

Same query. Feldera maintains incremental state and changes only affected output rows.

Data Flow

Connect, query, and serve features

Feldera connects event sources to model serving, keeping features current without custom glue code.

Source

Apache Kafka

Source

Postgres CDC

Source

Amazon S3

Engine

Feldera

Incremental SQL

Sink

Feature Store

Sink

REST API

Sink

Kafka Output

Connectors

Integration categories

Major streaming and database sources have built-in connectors. No plugins or separate ETL jobs.

Streaming

Apache Kafka / Redpanda

Confluent-compatible consumer with offset tracking. Handles Avro, JSON, and Protobuf schemas through standard consumer group properties.

Database CDC

PostgreSQL CDC

Integrates with logical replication slots and captures row-level INSERT, UPDATE, and DELETE events. No application changes needed.

Database CDC

MySQL Binlog

Binary log reader for MySQL 8.x and compatible forks. Uses row-based replication and supports schema evolution.

Object Storage

Amazon S3

Ingest Parquet, CSV, and JSON files with incremental file tracking. Read only new or changed objects. Supports Glue Data Catalog.

Performance

Sub-100ms feature update latency

Sub-100ms feature update latency came from our internal benchmark: five pipeline configurations were compared with a full-recompute baseline. Cases included single-key aggregation, multi-key GROUP BY, sliding-window joins, a CDC-fed feature table, and S3-batch bootstrap followed by stream incremental.

These results come from our internal test environment. Production speed varies with source event rate, query complexity, hardware configuration, and network topology.

Deploy in minutes

Start an incremental pipeline

Install the CLI, write a SQL query, and run your pipeline. No infrastructure setup beyond a Kafka topic or database.

Read the quickstart Get started free