An Ash data layer for ScyllaDB using Xandra (direct CQL driver).

This data layer implements the Ash.DataLayer behaviour to allow Ash resources to be backed by ScyllaDB/Cassandra.

Configuration

Configure your resource to use this data layer:

defmodule MyApp.MyResource do
  use Ash.Resource,
    data_layer: AshScylla.DataLayer

  attributes do
    uuid_primary_key :id
    attribute :name, :string
  end

  relationships do
    # Define relationships as needed
  end
end

Features Supported

• :create - Create records
• :read - Read records with filtering
• :update - Update records
• :destroy - Delete records
• :filter - Filter queries
• :limit - Limit results
• :select - Select specific fields
• :multitenancy - Keyspace-based multitenancy
• :upsert - Upsert records (INSERT IF NOT EXISTS with LWT)
• :update_query - Bulk update via filtered queries
• :destroy_query - Bulk delete via filtered queries
• :keyset - Token-based keyset pagination (the default pagination mode)
• :distinct - DISTINCT on partition key columns
• :sort / {:sort, _} - ORDER BY on clustering columns
• :boolean_filter - OR filter rewriting to IN where possible
• :nested_expressions - Nested filter expressions
• {:filter_expr, _} - Filter expression support
• :changeset_filter - Changeset-based filtering
• :calculate - In-memory calculations
• :action_select - Action-specific select
• :async_engine - Async engine support
• :bulk_create - Batch INSERT operations
• :transact - Transaction wrapper
• :composite_primary_key - Composite primary key support
• {:aggregate, :count} - Per-partition COUNT aggregates
• {:atomic, :update} - Atomic updates via LWT (IF clauses)
• {:atomic, :upsert} - Atomic upserts via LWT
• {:atomic, :create} - Atomic creates

Features NOT Supported

• :offset - ScyllaDB has no OFFSET; use keyset pagination
• :expr_error - Expression error handling not implemented
• :expression_calculation_sort - Not supported
• :aggregate_filter - Aggregate filtering not supported
• :aggregate_sort - Aggregate sorting not supported
• :bulk_create_with_partial_success - Bulk create is all-or-nothing
• :update_many - Update-many not implemented
• :composite_type - Composite types not supported
• :through_relationship - Through relationships not supported
• :bulk_upsert_return_skipped - Not supported
• :distinct_sort - Not supported
• {:combine, :union} / {:combine, :union_all} / {:combine, :intersection} - No combination queries
• {:lock, :for_update} - Locking is a no-op
• {:join, _} - No JOINs (use denormalization)
• {:lateral_join, _} - No lateral joins
• {:filter_relationship, _} - Relationship filtering not supported
• {:exists, :unrelated} - Exists queries not supported
• {:aggregate, :unrelated} - Unrelated aggregates not supported
• {:aggregate_relationship, _} - Aggregate relationships not supported
• {:query_aggregate, _} - Query aggregates not supported
• {:aggregate, :sum} / {:aggregate, :avg} / etc. - Only COUNT is supported

Ash Query Extensions

The following Ash 3.0+ query features are supported via Xandra:

• fragment/1+ — raw CQL injection: fragment("col = ?", value) passes through to Xandra directly
• now(), today(), ago(...), from_now(...) — evaluated client-side by Ash before reaching the data layer
• has(collection_col, value) — maps to CQL CONTAINS on collection/set/list columns
• overlaps(collection_col, [a, b]) — maps to col CONTAINS a OR col CONTAINS b with ALLOW FILTERING
• Arithmetic operators (+, -, *, /) — evaluated client-side by Ash
• String functions (concat, contains, starts_with, ends_with, string_length, etc.) — maps to CQL LIKE where applicable
• if/3, is_nil/1, length/1, round/1, string_downcase/1, string_trim/1 — evaluated client-side by Ash

Limitations

Since ScyllaDB/Cassandra is a wide-column store, not all SQL features are supported:

• No JOINs (use denormalization or multiple queries)
• Expression calculations are done in Elixir post-processing (not in-database)
• DISTINCT only works on partition key columns
• Limited aggregation support
• Combination queries (UNION/INTERSECT) are not supported
• No transactions across partitions (lightweight transactions only)
• Locking is a no-op (use LWT for conditional operations)
• No complex WHERE clauses on non-primary key columns without secondary indexes
• Cross-partition aggregates require materialized views
• CQL ORDER BY only works on clustering columns within a partition