Skip to content

Latest commit

 

History

History
115 lines (85 loc) · 3.09 KB

architecture.md

File metadata and controls

115 lines (85 loc) · 3.09 KB
Raw

Architecture

deltakernel crate architecture is still a work in progress!

Goals

in order of priority (this is placeholder and we need to redo them):

  1. simplicity and ease of use (probably too vague)
  2. query engine agnostic
  3. performance is explicitly secondary goal with the exception of operating in bounded memory

10,000-foot view

Two major API surface areas:

  1. Engine API
  2. Table API

Consider the usage pattern by example: if delta-rs wants to leverage deltakernel to read tables, it first must take a dependency on deltakernel and provide any of the traits it wishes (otherwise rely on defaults already provided in deltakernel) - this is API (1) above. Then the engine code can leverage the table API (2) in order to perform actual interaction with delta tables.

Engine API

The engine API aims to provide the least dependency surface as possible, this means largely using traits to dictate what behavior should be implemented "above" while placing as much "core" Delta Lake protocol implementation into the deltakernel crate as possible.

  • trait ObjectStore
  • trait JsonReader
    • arrow-json, simd-json, serde-json, etc
  • trait ParquetReader
    • Can DuckDB bring their own Parquer reader an in-memory format
  • trait ExpressionEvaluator
    • datafusion
    • duckdb's
    • predominantly used during data skipping, to scan for where x < 5 based on file skipping
  • struct DeltaLog/Segment/Snapshot/Replay
    • generic over the above traits to allow consistent log interactions across table API implementations.
classDiagram
    class ObjectStore {
    }

    %% DuckDB, Redshift, anything using Arrow C++, FFI?
    class ParquetReader {
    }

    %% arrow-json, simd-json, serde, FFI BYOJP
    class JsonReader {
    }

    class ExpressionEvaluator {
    }
Loading

Engine Integrations

Ideally there are some engines that would be able to support native Delta Lake integration on top of this API:

  • DuckDB
    • Has their own parquet reader and in-memory representation
  • ??

Table API

The Table API provides a little bit more opinions for handling some Delta Lake protocol nuance and should be incorporating more deppendencies to provide deltakernel users with a simpler path to building applications on top of the Delta Lake protocol..

Arrow

Sane defaults for the above traits with RecordBatch as the mode of interop between everything. This feature flag turns on the most sane default, parquet, json, some expression evaluator with arrow as its in-memory format.

classDiagram
    DeltaTable --> Snapshot : yields Snapshots
    Snapshot --> Scan
    Scan -->  ScanFileIterator
    ScanFileIterator -->  ScanDataIterator

    class StorageClient {
    }

    note for DeltaTable "Responsible for log storage"
    class DeltaTable {
        -Url location
        get_latest_snapshot(StorageClient) Snapshot
        get_latest_version(StorageClient) uint64
    }

    note for Snapshot "Responsible for log storage"
    class Snapshot {
        +uint64 version
    }

    class Scan {
        +object projection
        +object predicate
    }

    class StorageClient {
    }
Loading