AGENT SKILL

ffi-capsule-protocol

TRIGGER — read before adding, changing, or reviewing any __datafusion_*__ capsule getter, any FFI_* export that asks for a TaskContextProvider or an extension codec, or any code that calls FFI_QueryPlanner::new / FFI_TableProvider::new / FFI_{Logical,Physical}ExtensionCodec::new. These methods are one protocol with a settled convention. Do not design it fresh; do not construct a SessionContext inside an extension library.

License 许可
Apache-2.0
最近更新
2026-11-09
Stars
★ 613
Forks
184

01 前言 Front Matter

name
ffi-capsule-protocol
description
TRIGGER — read before adding, changing, or reviewing any __datafusion_*__ capsule getter, any FFI_* export that asks for a TaskContextProvider or an extension codec, or any code that calls FFI_QueryPlanner::new / FFI_TableProvider::new / FFI_{Logical,Physical}ExtensionCodec::new. These methods are one protocol with a settled convention. Do not design it fresh; do not construct a SessionContext inside an extension library.
argument-hint
[getter name] (e.g., "__datafusion_query_planner__", "table provider", "codec", or omit to review the whole family)

02 资源文件 3

03 SKILL.md 正文

<!--- 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. -->

FFI Capsule Protocol

datafusion-python shares Rust objects with extension libraries through PyCapsules. Every hook is a dunder method named __datafusion_<thing>__ that returns a capsule wrapping an FFI-safe struct. They are one protocol, not a collection of unrelated methods, and they have a settled convention that has already been migrated once (see docs/source/user-guide/upgrade-guides.md, DataFusion 52.0.0 and 55.0.0).

Rule 1 — enumerate the family before you change a member

Do this first, every time. It takes one command and it is the whole point of this skill:

grep -rn "__datafusion_[a-z_]*__" --include="*.rs" crates/ examples/*/src/

Compare the signature you are about to write against what the others already do. If yours is shaped differently, that is a finding about your design, not about theirs.

Rule 2 — a getter takes the session it is being installed on

fn __datafusion_physical_extension_codec__<'py>(
    &self,
    py: Python<'py>,
    session: Bound<'py, PyAny>,
) -> PyResult<Bound<'py, PyCapsule>> { ... }

The host calls the getter and passes itself. That argument is how an extension library reaches things only the session has.

SessionContext implements the same getters and ignores the argument, so a session satisfies the protocol too — ctx.__datafusion_query_planner__() and ctx.__datafusion_query_planner__(ctx) are both valid.

__datafusion_session_planner__(ctx, fallback) is the exception to the shape above: it takes a second argument, the planner assembled so far. A session has one planner slot, so planners compose by nesting rather than by chaining, and the host hands each bundle the previous layer instead of letting it capture one. Wrap fallback and delegate to it; returning a planner that ignores it discards every layer beneath, including one the session already had. It runs after every bundle's codecs are installed, so ctx carries the final chains.

That is also the only hook where it does. __datafusion_session_components__ runs before anything is installed, so its ctx still carries the chains the receiver had — the same session, and the same task-context provider, but not this call's codecs, not even your own. Read the host's codec chains in the planner hook, never in the extension hook.

A codec must always be handed over as an object implementing its getter, never as the bare capsule the getter returns; with_extensions refuses a capsule. A codec's wire id — the string a payload names on decode, which has to mean the same thing in whichever process decodes — is read off the object it arrives as, and a capsule has no type to read one from. Deriving the id from the bundle that contributed the capsule is not the fix: the bundle is whatever object the caller passed, so an application packaging your library inside a bundle of its own would re-tag your payloads and they would stop decoding where they are read. If the object's class name is not the identity you want on the wire, declare __datafusion_codec_id__ on it. BundledLogicalCodec in examples/datafusion-ffi-query-planner-example/src/extension.rs is the shape. This applies only to codecs — a query planner carries no wire id.

Rule 3 — never construct a SessionContext in an extension library

The FFI constructors ask for things a library does not have:

Constructor Wants Take it from
FFI_{Logical,Physical}ExtensionCodec::new TaskContextProvider ffi_task_context_provider_from_pycapsule(&session)
FFI_TableProvider::new_with_ffi_codec logical codec ffi_logical_codec_from_pycapsule(session, None)
FFI_QueryPlanner::new_with_ffi_codecs both codecs ffi_{logical,physical}_codec_from_pycapsule(session, None)

Arc::new(SessionContext::new()) is the wrong answer to all three, for two independent reasons:

  1. It is the wrong registry. Decode callbacks resolve names against whatever provider the codec carries. An empty session resolves nothing, so a function the host registered with register_udf is invisible to a node that references it by name.
  2. It dangles. FFI_TaskContextProvider downgrades its provider to a Weak. A context built inline in the getter is dropped before the capsule is ever used, and every callback then fails with TaskContextProvider went out of scope over FFI boundary.

Prefer the *_with_ffi_codec(s) constructors when they exist. They take prebuilt codecs that already carry the host's provider, so there is no provider parameter to get wrong.

Rule 4 — the helpers live in crates/util/src/lib.rs

ffi_logical_codec_from_pycapsule, ffi_physical_codec_from_pycapsule, ffi_query_planner_from_pycapsule, ffi_task_context_provider_from_pycapsule, table_provider_from_pycapsule. Each takes the object and, where relevant, an Option<&Bound<PyAny>> session:

  • Some(session) — importing a foreign object; the getter needs the session.
  • None — the object already is a session and is being asked for what it holds.

Adding a getter means adding a helper here, not hand-rolling capsule extraction at the call site.

Rule 5 — changing a getter's signature is a breaking change

Extension libraries implement these methods. A signature change breaks every one of them, and the failure is a bare TypeError from a call1. So:

  • Add a section to docs/source/user-guide/upgrade-guides.md with before/after Rust, matching the 52.0.0 and 55.0.0 entries.
  • Add the api change label to the PR.
  • Map the TypeError to a diagnosable message. call_capsule_getter in crates/util/src/lib.rs already does this; reuse it.
  • Update python/datafusion/context.py and python/datafusion/user_defined.py, where the Protocol type hints for these methods live.

Changing what a codec puts on the wire is equally breaking, and easier to miss because no signature moves and nothing fails to compile. Serialized plans outlive the process that wrote them, so the same checklist applies: upgrade guide, api change label, and a statement of exactly which sessions produce different bytes.

Rule 6 — a session keeps one Arc<SessionContext> for life

FFI_TaskContextProvider holds its provider weakly, and every codec handed to a foreign object carries one. A registered catalog provider upgrades that handle on every supports_filters_pushdown and every scan. The handle is bound to an Arc<SessionContext> allocation, so anything that replaces the allocation orphans every handle bound to the old one: TaskContextProvider went out of scope over FFI boundary.

So mutate SessionState in place — *self.ctx.state_ref().write() = ..., the way add_physical_optimizer_rule and set_session_query_planner both do — rather than deriving a replacement SessionContext. Carry the session id across the rewrite; SessionStateBuilder::new_from_existing drops it and build mints a fresh one, which desyncs session_id() from every TaskContext the session hands out.

Do not try to repair it after the fact:

  • You cannot rebind what you cannot reach. A codec embedded in a registered FFI_CatalogProvider, and in every FFI_SchemaProvider and FFI_TableProvider minted from it, has no Python-side handle.
  • A codec must not retain its session. Codecs are routinely handed to a provider that is registered straight back into the session that built them, closing SessionContext -> catalog -> FFI provider -> FFI codec -> SessionContext.

test_registered_providers_survive_a_planner_install in examples/datafusion-ffi-query-planner-example/python/tests/_test_three_library_query_planner.py guards this. Its WHERE clause is load-bearing: filter pushdown upgrades the weak handle during logical optimization, before plan serialization could fail first for an unrelated reason.

SessionContext.with_extensions is where this rule is easiest to get wrong, because "bind the components to the context you are about to return" reads like an instruction to derive one first. It is not: the factories are handed the receiver, and the returned handle shares its allocation. There is nothing to keep alive separately and nothing to garbage-collect out from under a provider.

SessionContext.enable_url_table is the one method that mints a second allocation for a session. Its result must not outlive the receiver, and it also forks the session's SessionState while keeping its id, so two handles report one session_id() with divergent configuration. That is a bug rather than a design — tracked in apache/datafusion-python#1708 — so do not cite it as precedent for deriving a replacement context.

Rule 7 — installing a planner mutates the session, and says so

set_query_planner returns None, matching add_physical_optimizer_rule. The query planner lives in SessionState, so it belongs to the session and not to a handle on it; every context sharing that session plans through it. Do not reintroduce a with_query_planner that pretends otherwise — the only way to give a handle its own planner is a fresh Arc<SessionContext>, which is what Rule 6 forbids.

Installing a codec rebuilds the installed planner against it, and that rebuild reaches exactly one layer. FFI_QueryPlanner::new_with_ffi_codecs unwraps one ForeignQueryPlanner; a fallback that planner resolved at install time sits in its library's private data with no handle on this side, and cannot re-derive codecs itself because FFI_QueryPlanner holds them by value and Session exposes no accessor for the host's current ones. So do not promise that install order is free — for a layered planner it is not. The examples cannot show this: their fallback lives in the same cdylib as its wrapper, and datafusion-ffi short-circuits a same-library hop rather than serializing. A fix has to come from upstream; tracked in apache/datafusion#24762.

The session's planner also tracks whichever handle wrote it last, so re-installing a planner on the original handle rebinds the session back to that handle's codecs. test_reinstalling_a_planner_rebinds_the_session_to_that_handles_codecs pins that; changing it should be deliberate.

Where the truth is

  • docs/source/extension-guide/ — the protocol, for the library author. capsule-protocol.md has the hook convention and what the getter argument actually is; codecs.md, bundles.md, and query-planners.md have the per-component rules; index.md lists all 18 hooks.
  • docs/source/contributor-guide/ffi-internals.md — why the framing is shaped this way, including the weak-Arc scheme and the one-level rebind.
  • docs/source/user-guide/upgrade-guides.md — every past migration.
  • crates/core/src/codec.rs — the codec chain: the envelope, identity dispatch, and the two unframed cases from Rule 8.
  • examples/datafusion-ffi-example/src/ — provider, catalog, function, codec getters, all in current form. name_only_codec.rs is the codec that encodes nothing.
  • examples/datafusion-ffi-query-planner-example/src/planner.rs — planner getter.
  • examples/datafusion-ffi-query-planner-example/python/tests/_test_three_library_query_planner.py — require_udf_on_decode proves which session a decode callback resolves against. Extend these when touching the protocol.

04 同仓库更多 Skills