diff --git a/.gitignore b/.gitignore
@@ -6,6 +6,8 @@ rust-target/
 *.iml
 .DS_Store
 tpch-data/
+# Lockfile for the example ADBC driver crate (a demo cdylib, not a published lib).
+examples/adbc-datafusion-driver/Cargo.lock
 .claude
 docs/superpowers
 docs/build/

diff --git a/docs/source/user-guide/adbc-spark-connector.md b/docs/source/user-guide/adbc-spark-connector.md
@@ -0,0 +1,240 @@
+<!--
+Licensed to the Apache Software Foundation (ASF) under one
+or more contributor license agreements.  See the NOTICE file
+distributed with this work for additional information
+regarding copyright ownership.  The ASF licenses this file
+to you under the Apache License, Version 2.0 (the
+"License"); you may not use this file except in compliance
+with the License.  You may obtain a copy of the License at
+
+  http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing,
+software distributed under the License is distributed on an
+"AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+KIND, either express or implied.  See the License for the
+specific language governing permissions and limitations
+under the License.
+-->
+
+# DataFusion as a Spark data source (ADBC)
+
+This repository ships a Spark connector that lets a Spark job read from a
+DataFusion `TableProvider` as a native Spark `DataSourceV2`. Spark sees an
+ordinary external table; the data is produced by DataFusion in native code and
+handed back as Apache Arrow batches.
+
+## What this repository provides
+
+Two pieces work together:
+
+- **The `adbc-datafusion` Spark connector** (`spark/`) — a Spark
+  `DataSourceV2` registered under the format name `adbc-datafusion`. It builds
+  the scan, pushes projection / filters / limit down, maps DataFusion's scan
+  partitions onto Spark input partitions, and imports each partition's Arrow
+  data into Spark **zero-copy** as an `ArrowColumnVector`.
+- **An example DataFusion ADBC driver** (`examples/adbc-datafusion-driver/`) —
+  a small native library (cdylib) that exposes a DataFusion `TableProvider`
+  through the [ADBC](https://arrow.apache.org/adbc/) C API. It is a worked
+  example you can copy: your own driver registers whatever provider you want and
+  is loaded the same way.
+
+The connector is built on the standard Apache Arrow ADBC Java driver manager
+(`adbc-core` + `adbc-driver-jni`), so it talks to the native driver through a
+stable, public interface. Because that boundary is plain ADBC, the same
+connector can front **any** ADBC-speaking source; the DataFusion driver shown
+here is just one example, and you can point it at a different ADBC driver
+without changing the connector.
+
+```text
+Spark DataSourceV2 (adbc-datafusion)
+  → arrow-adbc Java driver manager (adbc-core + adbc-driver-jni)
+    → native ADBC driver (the DataFusion cdylib)
+      → your DataFusion TableProvider
+```
+
+## Reading from Spark
+
+Point the `driver` option at the built driver library and the `table` option at
+a table the provider exposes:
+
+```scala
+val df = spark.read
+  .format("adbc-datafusion")
+  .option("driver", "/abs/path/to/libadbc_datafusion_example_driver.so")
+  .option("entrypoint", "AdbcDatafusionExampleInit")
+  .option("table", "example")
+  .load()
+
+df.filter("id > 1").select("name").show()
+```
+
+The connector probes the schema once (on the driver) when the `DataFrame` is
+created, then plans the scan. `df.filter(...)` and `df.select(...)` are pushed
+into the DataFusion scan where possible (see below); the rest Spark evaluates
+itself.
+
+## Two partitioning models that must be reconciled
+
+The heart of the connector is mapping DataFusion's idea of a partitioned scan
+onto Spark's idea of tasks. The two engines size scans on **different
+principles**, and a provider author who ignores the difference will get a job
+that runs but performs badly.
+
+### DataFusion: parallelism-bound
+
+DataFusion partitions a scan to **saturate the cores of one machine**. Its
+built-in `ListingTable`, for example, packs input files into roughly
+`target_partitions` groups — where `target_partitions` defaults to the local
+core count — using a minimum-size floor and optional single-file splitting only
+to *reach* that count. The partition count is therefore approximately
+
+```text
+N ≈ target_partitions ≈ cores
+```
+
+independent of how much data there is. More data means **bigger** partitions,
+not more of them.
+
+### Spark: one task per partition, byte-bound
+
+Spark turns **each input partition into one task**, and a native Spark file
+source sizes partitions by **bytes**: it targets roughly one partition per
+`spark.sql.files.maxPartitionBytes` (default 128 MB), cutting files into splits
+and bin-packing them to that size. (Spark shrinks the target below 128 MB when
+the data is small, so that there are at least enough partitions to keep every
+core busy.)
+
+So large data means **many** partitions — typically far more than the cluster
+has cores — which run in waves. Each task processes a bounded slice (≈128 MB),
+which is what keeps memory bounded and lets Spark reschedule stragglers.
+
+### How the connector joins them
+
+For a scan that only pushes projection / filter / limit (no DataFusion-side
+join or aggregation), the physical plan's output partitioning **is** the
+provider's `scan()` partitioning, and the mapping to Spark is **1:1**:
+
+```text
+provider.scan() output partitions
+  → DataFusion physical plan partition count   N
+  → N ADBC partition descriptors
+  → N Spark input partitions = N Spark tasks (scan stage)
+```
+
+Spark never splits or merges a partition afterward — `df.rdd().getNumPartitions()
+== N`, and the cluster core count only bounds how many of the `N` run at once.
+
+**The consequence you must plan for:** suppose a provider keeps DataFusion's
+default of `N = target_partitions ≈ cores` partitions. On a large dataset that
+produces only a handful of partitions, each holding about `total_bytes / cores`
+of data — so Spark runs a few very large tasks. Because a task is the unit of
+work, those oversized tasks bring memory pressure, stragglers that hold up the
+whole stage, and no way for Spark to rebalance. To feed Spark well, a provider
+should size partitions by **bytes** — the way a Spark file source does — rather
+than by the `target_partitions` count.
+
+## Sizing the scan
+
+Aim for the Spark sweet spot:
+
+- **Floor:** `N` ≥ total executor cores, or cores sit idle. `~2–4× cores` gives
+  slack for skew and stragglers.
+- **Per-partition size:** big enough to amortize per-task overhead — the
+  canonical target is **128–256 MB** of data per partition (matching Spark's
+  `spark.sql.files.maxPartitionBytes` default of 128 MB).
+- **Ceiling — and ADBC lowers it.** Beyond Spark's generic per-task cost, this
+  path adds two per-partition costs: each of the `N` ADBC descriptors carries a
+  copy of the **whole serialized physical plan**, and each task **deserializes**
+  that plan when it reads its partition. (The driver caches a deserialized plan
+  per connection, but that does not help here: each Spark task uses its own
+  connection and reads a single partition, so the plan is deserialized once per
+  task and never reused.) Tens of thousands of tiny partitions is far costlier
+  here than for a native file source. Keep `N` in the hundreds; prefer bigger
+  partitions over a huge count.
+
+Rule of thumb, balanced by **bytes** rather than split count:
+
+```text
+N ≈ clamp(total_bytes / target_bytes, floor=cores, ceiling≈hundreds)
+```
+
+## Writing a provider that feeds Spark well
+
+Your provider's `scan()` receives the session, so it can read the connector's
+parallelism hint and then decide based on what it knows about the data:
+
+```text
+T            = state.config().target_partitions()   // connector hint ≈ desired parallelism
+target_bytes = ~128–256 MB                           // bias larger for ADBC to keep N down
+min_bytes    = floor so partitions never get tiny    // ~ tens of MB
+
+if total_bytes known AND splittable (files / row groups / key ranges):
+    N = clamp(ceil(total_bytes / target_bytes), 1, CEILING)   // byte-bound
+    bin-pack splits into N groups BALANCED BY BYTES           // not by count
+    do not split below min_bytes; merge small splits
+
+elif split_count known but not bytes (shards / external partitions):
+    N = min(split_count, CEILING)                             // natural partitioning
+    if split_count >> T: coalesce shards into ~T balanced groups
+
+else (size and splits unknown; opaque / streaming):
+    N = T                                                     // fall back to the hint
+    lean slightly high — stragglers hurt more than overhead
+```
+
+- **Balance by bytes, not by split count.** A task is atomic; Spark cannot
+  rebalance a fat partition. One 10 GB partition among 100 small ones stalls the
+  whole stage. Bin-pack so partition bytes are even.
+- **`target_partitions` is a hint, cap, and fallback** — not the data-bound
+  count. Use it when bytes are unknown or as a parallelism ceiling; when you know
+  the data size, let **bytes** drive `N`.
+
+### Planning happens once, on the driver
+
+Your provider's `scan()` runs **once**, on the driver, while the multi-partition
+descriptors are built. The driver plans the query, fixes the partitioning, and
+**serializes the whole physical plan** into each descriptor. Each executor task
+then deserializes that plan and executes its partition index. So partition
+`index i` always means the same slice: there is no second planning pass that
+could disagree with the first, and no need to make `N` stable across a
+driver/executor boundary.
+
+The one requirement this places on a provider is that its plan be
+**serializable**: built-in DataFusion nodes round-trip through the default codec,
+but a custom `ExecutionPlan` node needs a `PhysicalExtensionCodec` registered
+with the driver so it can be encoded on the driver and decoded on the executor.
+
+## Connector options
+
+| Option | Required | Meaning |
+| --- | --- | --- |
+| `driver` | yes | Path to (or manifest name of) the native ADBC driver library. |
+| `entrypoint` | depends on driver | Name of the driver's C init symbol (e.g. `AdbcDatafusionExampleInit`). |
+| `table` | yes | Table the provider exposes; its schema is probed on the driver. |
+| `target_partitions` | no | Parallelism hint. Defaults to `SparkContext.defaultParallelism` (total executor cores). The connector issues `SET datafusion.execution.target_partitions = N` on the planning session, so it influences the partition count when the driver builds the physical plan; that plan (with its partitioning fixed) is what gets serialized into the descriptors. Pass `k × cores` to raise parallelism. |
+| `manifest.path` | no | Extra search path for ADBC driver manifests when `driver` is a name rather than an absolute path. |
+| *(anything else)* | no | Forwarded verbatim as native ADBC database options, so provider-specific knobs pass straight through. |
+
+Two things are intentionally **not** connector options:
+
+- **`maxPartitionBytes`** — byte-sizing lives in the provider's `scan()`; the
+  connector does not know the data size and cannot derive it. Pass a sizing knob
+  as a provider-specific option if your provider supports one.
+- **Reported partitioning / storage-partitioned joins
+  (`SupportsReportPartitioning`)** — not available over ADBC today: the
+  descriptor carries no distribution metadata, and DataFusion's hash
+  partitioning does not match Spark's key-grouped / bucketed model.
+
+## Summary
+
+- The connector maps the provider's `scan()` partition count **1:1** onto Spark
+  tasks.
+- DataFusion sizes scans to **cores** (parallelism-bound); Spark sizes them to
+  **bytes** (one task per ~128 MB). To feed Spark well, size your provider's
+  partitions by bytes — typically 128–256 MB each, floored at executor cores and
+  capped in the hundreds because every descriptor carries the full serialized
+  plan and each task deserializes it.
+- Planning happens once on the driver; the physical plan is serialized into each
+  descriptor and executors run it as-is, so a custom `ExecutionPlan` node just
+  needs a registered `PhysicalExtensionCodec` to be serializable.
diff --git a/docs/source/user-guide/index.md b/docs/source/user-guide/index.md
@@ -26,7 +26,8 @@ DataFrame queries execute in native Rust; results return to the JVM as
 Data Interface.
 
 This guide covers installation, the `SessionContext` and `DataFrame` APIs,
-and Parquet ingestion.
+Parquet ingestion, table providers, and using DataFusion as a Spark data
+source over ADBC.
 
 ```{toctree}
 :maxdepth: 1
@@ -39,6 +40,7 @@ parquet
 proto-plans
 scalar-udf
 table-provider
+adbc-spark-connector
 api-reference
 ```
 

diff --git a/examples/adbc-datafusion-driver/Cargo.toml b/examples/adbc-datafusion-driver/Cargo.toml
@@ -0,0 +1,57 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+[package]
+name = "adbc-datafusion-example-driver"
+version = "0.1.0"
+edition = "2021"
+publish = false
+description = "End-to-end example: an ADBC driver that exposes a custom DataFusion TableProvider"
+
+# Standalone crate that stands in for a real, separate driver repo. The empty
+# [workspace] table makes it its own workspace root so it is NOT pulled into the
+# parent datafusion-java workspace (whose members are listed explicitly).
+[workspace]
+
+[dependencies]
+# The ready-made ADBC-over-DataFusion driver. We depend on it as a LIBRARY and
+# only customize the SessionContext via its new_with_context_init hook -- no
+# fork, no datafusion-ffi module loading. Pulled from its own repo (git) since
+# that is where it lives.
+adbc-driver-datafusion = { git = "https://github.com/adbc-drivers/datafusion" }
+adbc_core = "0.23"
+adbc_ffi = "0.23"
+datafusion = "53"
+async-trait = "0.1"
+# For building Float16 example data (arrow's half-precision element type).
+half = "2"
+
+# Build both an rlib (so the provider/driver can be unit-tested or embedded) and
+# a cdylib (the artifact the arrow-adbc Java driver manager loads at runtime).
+[lib]
+crate-type = ["cdylib", "lib"]
+
+# Temporary: redirect adbc-driver-datafusion to the PR branch that implements
+# execute_partitions / read_partition plus the shared (database-scoped) plan
+# cache (adbc-drivers/datafusion#32), so we can run a full multi-partition
+# system test before it merges. The crate is not on crates.io, so this patches
+# the git source (not [patch.crates-io]). Pinned to a rev (not the branch) so
+# builds are reproducible and the plan-cache assertion in AdbcSourceTest is
+# stable. Remove once the PR is merged and the canonical dependency carries the
+# change.
+[patch."https://github.com/adbc-drivers/datafusion"]
+adbc-driver-datafusion = { git = "https://github.com/timsaucer/adbc-driver-datafusion", rev = "9b51df2b13d20d81c563ef3ee818db1c02082fac" }