All data processing in kamu is done by a set of plug-in engines. This allows us to integrate many mature data processing frameworks, while having kamu coordinate all processing to track provenance and ensure verifiability.

Known Engine Implementations

Name Technology Query Dialect Links Notes
spark Apache Spark Spark Streaming SQL Repository
Image		 Spark is used in kamu-cli for all data ingestion and is default (but not only) engine for SQL shell. Spark is also used in combination with Livy to query data from embedded Jupyter Notebooks. It’s currently the only engine that supports GIS data via Apache Sedona integration.
flink Apache Flink Flink Streaming SQL Repository
Image		 Flink has most mature support for stream processing, like stream-to-stream and stream-to-table joins, windowed aggregations, watermarks etc. It’s thus the recommended engine for most derivative datasets.
datafusion Apache Arrow DataFusion DataFusion Batch SQL Repository
Image		 Experimental engine that has limited functionality (due to being batch-oriented), but is extremely fast and low-footprint. There are ongoing attempts to add stream processing functionality. DataFusion is also embedded into kamu-cli and is used for executing ad-hoc SQL queries. We are aiming to make data ingest functionality use DataFusion for most cases instead of Spark.

Schema Support

Feature kamu Spark Flink DataFusion
Basic types ✔️ ✔️ ✔️ ✔️
Decimal type ✔️ ✔️ ✔️** ✔️
Nested types ✔️* ✔️ ❌ ❌
GIS types ✔️* ✔️ ❌ ❌

* There is currently no way to express nested and GIS data types when declaring root dataset schemas, but you still can use them through pre-processing queries

** Apache Flink has known issues with Decimal type and currently relies on our patches that have not been upstreamed yet, so stability is not guaranteed FLINK-17804.

Operation Types

Feature Spark Flink DataFusion
Filter ✔️ ✔️ ✔️
Map ✔️ ✔️ ✔️
Aggregations ❌* ✔️ ❌
Stream-to-Stream Joins ❌* ✔️ ❌
Projection / Temporal Table Joins ❌* ✔️ ❌
GIS extensions ✔️ ❌ ❌

* Spark Engine is capable of stream processing but temporarily we have to use it in the batch processing mode, so only row-level operations like map and filter are currently usable, as those do not require correct stream processing and watermark semantics.