Using extension libraries#

An extension library is a separate package that teaches a SessionContext something it does not know on its own — a new data source, extra functions, or a different way of executing your queries. You install it with pip, hand it to your session, and keep writing the same SQL and DataFrame code.

Examples in the wild include delta-rs, which exposes Delta Lake tables to DataFusion, and the two worked examples in this repository under examples/.

Two kinds of extension#

Which one you have determines how much setup you do.

Tables and functions register directly. If the library gives you a table or a function, register it the same way you would register a CSV file. No extra setup:

from datafusion import SessionContext
import my_tables

ctx = SessionContext()
ctx.register_table("events", my_tables.TableProvider("s3://bucket/events"))
ctx.sql("SELECT count(*) FROM events").show()

Libraries that change how queries run need to be installed on the session. A distributed engine, or anything that rewrites your query plan, has to be attached to the session before it can do its work. That is what with_extensions() is for. The library documents an object — often called Extension — that you pass to it:

from datafusion import SessionContext
import my_engine

ctx = SessionContext().with_extensions(my_engine.Extension("scheduler:50050"))
ctx.register_table("events", my_engine.TableProvider("s3://bucket/events"))
ctx.sql("SELECT count(*) FROM events").show()

with_extensions returns a context; use the returned one. It shares everything else with the context you called it on, so tables you registered before the call are still there.

Using more than one library#

Pass them all to a single call:

ctx = SessionContext().with_extensions(
    my_tables.Extension(),
    my_engine.Extension("scheduler:50050"),
)

One call rather than several is worth preferring: it lets the libraries see each other, which they cannot do if you install them one at a time. Order rarely matters. When a library needs a particular position — usually “list me last” for something that wraps the others — it says so in its own documentation.

Two things that will bite you#

Keep your context alive. A DataFrame or a plan does not keep its session alive on its own. If a context is garbage-collected while something built from it is still in use, the next query fails with:

TaskContextProvider went out of scope over FFI boundary

Almost always this is a helper that built a context locally and returned a DataFrame:

# Wrong — ctx is collected when the function returns.
def load():
    ctx = SessionContext().with_extensions(my_engine.Extension())
    return ctx.sql("SELECT * FROM events")

# Right — hand back the context too, or keep it on an object that lives
# as long as the frames derived from it.
def load():
    ctx = SessionContext().with_extensions(my_engine.Extension())
    return ctx, ctx.sql("SELECT * FROM events")

Versions have to match. An extension library is compiled against one DataFusion version. A mismatch raises an ImportError naming the version it found and the version expected, at the moment you register or install the library — not silently at query time. If you see one, upgrade or downgrade the extension library so its DataFusion version matches this package’s. See Mismatched extension libraries now fail loudly.

Checking what a session knows about#

logical_extension_codec_ids() and physical_extension_codec_ids() list which libraries a session has been taught about. Useful when a query fails and you want to confirm the library actually got installed:

ctx = SessionContext().with_extensions(my_engine.Extension())
ctx.logical_extension_codec_ids()
# ['my_engine.LogicalCodec']

An empty list means nothing extra is installed.

Next steps#