Is your feature request related to a problem or challenge? Please describe what you are trying to do.
In a distributed setup the scheduler plans a query and hands stages to executors; only the executors ever call a Python UDF. Decoding an inlined Python UDF unpickles the function, which requires a Python interpreter and every module the function closes over to be importable. Doing that on the scheduler costs work nobody needs and forces the scheduler image to carry Python and the full dependency set of user code it will never run. Raised in #1678 (review).
Describe the solution you'd like
An opaque decode path: a ScalarUDFImpl that holds the still-pickled blob rather than a live Python object, and a codec that produces it. A scheduler installs that codec, decodes a plan into something it can inspect, route, and re-encode, and never touches cloudpickle. The executor installs the ordinary codec and unpickles as it does now.
The wire format already allows this. An inlined UDF payload is DFPYUDF followed by a version byte and the cloudpickle blob (crates/core/src/codec.rs), so an opaque holder can carry those bytes verbatim and no format change is needed.
The part that needs design is re-encoding. A scheduler that forwards a stage has to emit the blob byte-identically, so the executor sees exactly what the client wrote. That also raises what such a UDF should report for the things DataFusion asks of a ScalarUDFImpl during planning — name, signature, and return type are all recoverable from the payload without unpickling, since they are stored alongside the function, but invoke has to be an error rather than a surprise.
Describe alternatives you've considered
Encoding Python UDFs by name only and registering them on every node. Already supported and appropriate when the function is available everywhere; it does not cover the case inlining exists for, which is a function the receiving process does not have.
Having the scheduler unpickle and immediately drop the object. Keeps the code simple, and still requires Python plus all user dependencies on the scheduler, which is the actual cost being avoided.
Additional context
Follow-up from #1678, which made extension codecs compose so a setup like this can install a scheduler-side codec alongside others. Likely also depends on #1703, gating pyo3/extension-module, if the consumer is a Rust crate rather than a Python process.
Is your feature request related to a problem or challenge? Please describe what you are trying to do.
In a distributed setup the scheduler plans a query and hands stages to executors; only the executors ever call a Python UDF. Decoding an inlined Python UDF unpickles the function, which requires a Python interpreter and every module the function closes over to be importable. Doing that on the scheduler costs work nobody needs and forces the scheduler image to carry Python and the full dependency set of user code it will never run. Raised in #1678 (review).
Describe the solution you'd like
An opaque decode path: a
ScalarUDFImplthat holds the still-pickled blob rather than a live Python object, and a codec that produces it. A scheduler installs that codec, decodes a plan into something it can inspect, route, and re-encode, and never touches cloudpickle. The executor installs the ordinary codec and unpickles as it does now.The wire format already allows this. An inlined UDF payload is
DFPYUDFfollowed by a version byte and the cloudpickle blob (crates/core/src/codec.rs), so an opaque holder can carry those bytes verbatim and no format change is needed.The part that needs design is re-encoding. A scheduler that forwards a stage has to emit the blob byte-identically, so the executor sees exactly what the client wrote. That also raises what such a UDF should report for the things DataFusion asks of a
ScalarUDFImplduring planning — name, signature, and return type are all recoverable from the payload without unpickling, since they are stored alongside the function, butinvokehas to be an error rather than a surprise.Describe alternatives you've considered
Encoding Python UDFs by name only and registering them on every node. Already supported and appropriate when the function is available everywhere; it does not cover the case inlining exists for, which is a function the receiving process does not have.
Having the scheduler unpickle and immediately drop the object. Keeps the code simple, and still requires Python plus all user dependencies on the scheduler, which is the actual cost being avoided.
Additional context
Follow-up from #1678, which made extension codecs compose so a setup like this can install a scheduler-side codec alongside others. Likely also depends on #1703, gating
pyo3/extension-module, if the consumer is a Rust crate rather than a Python process.