DataFusion Federation allows DataFusion to execute (part of) a query plan by a remote execution engine.
┌────────────────┐
┌────────────┐ │ Remote DBMS(s) │
SQL Query ───> │ DataFusion │ ───> │ ( execution │
└────────────┘ │ happens here ) │
└────────────────┘
The goal is to allow resolving queries across remote query engines while pushing down as much compute as possible to the remote database(s). This allows execution to happen as close to the storage as possible. This concept is referred to as 'query federation'.
Tip
This repository implements the federation framework itself. If you want to connect to a specific database, check out the compatible providers available in datafusion-contrib/datafusion-table-providers.
Check out the examples to get a feel for how it works.
For a complete step-by-step example of how federation works, you can check the example here.
- Querying across SQLite, MySQL, PostgreSQL, ...
- Pushing down SQL or Substrait plans.
- DataFusion -> Flight SQL -> DataFusion
- ..
Say you have a query plan as follows:
┌────────────┐
│ Join │
└────────────┘
▲
┌───────┴────────┐
┌────────────┐ ┌────────────┐
│ Scan A │ │ Join │
└────────────┘ └────────────┘
▲
┌───────┴────────┐
┌────────────┐ ┌────────────┐
│ Scan B │ │ Scan C │
└────────────┘ └────────────┘
DataFusion Federation will identify the largest possible sub-plans that can be executed by an external database:
┌────────────┐ Optimizer recognizes
│ Join │ that B and C are
└────────────┘ available in an
▲ external database
┌──────────────┴────────┐
│ ┌ ─ ─ ─ ─ ─ ─ ┴ ─ ── ─ ─ ─ ─ ─┐
┌────────────┐ ┌────────────┐ │
│ Scan A │ │ │ Join │
└────────────┘ └────────────┘ │
│ ▲
┌───────┴────────┐ │
┌────────────┐ ┌────────────┐ │
││ Scan B │ │ Scan C │
└────────────┘ └────────────┘ │
─ ── ─ ─ ── ─ ─ ─ ─ ─ ─ ─ ── ─ ┘
The sub-plans are cut out and replaced by an opaque federation node in the plan:
┌────────────┐
│ Join │
└────────────┘ Rewritten Plan
▲
┌────────┴───────────┐
│ │
┌────────────┐ ┏━━━━━━━━━━━━━━━━━━┓
│ Scan A │ ┃ Scan B+C ┃
└────────────┘ ┃ (TableProvider ┃
┃ that can execute ┃
┃ sub-plan in an ┃
┃external database)┃
┗━━━━━━━━━━━━━━━━━━┛
Different databases may have different query languages and execution capabilities. To accommodate for this, we allow each 'federation provider' to self-determine what part of a sub-plan it will actually federate. This is done by letting each federation provider define its own optimizer rule. When a sub-plan is 'cut out' of the overall plan, it is first passed the federation provider's optimizer rule. This optimizer rule determines the part of the plan that is cut out, based on the execution capabilities of the database it represents.
A remote database is represented by the FederationProvider trait. To identify
table scans that are available in the same database, they implement
FederatedTableSource trait. This trait allows lookup of the corresponding
FederationProvider.
Identifying sub-plans to federate is done by the FederationOptimizerRule.
This rule needs to be registered in your DataFusion SessionState. One easy way
to do this is using default_session_state. To do its job, the
FederationOptimizerRule currently requires that all TableProviders that need
to be federated are FederatedTableProviderAdaptors. The
FederatedTableProviderAdaptor also has a fallback mechanism that allows
implementations to fallback to a 'vanilla' TableProvider in case the
FederationOptimizerRule isn't registered.
The FederationProvider can provide a compute_context. This allows it to
differentiate between multiple remote execution context of the same type. For
example two different mysql instances, database schemas, access level, etc. The
FederationProvider also returns the Optimizer that is allows it to
self-determine what part of a sub-plan it can federate.
The sql module implements a generic FederationProvider for SQL execution
engines. A specific SQL engine implements the SQLExecutor trait for its
engine specific execution. There are a number of compatible providers available
in
datafusion-contrib/datafusion-table-providers.
The project is in alpha status. Contributions welcome; land a PR = commit access.