Tengfei Huang created SPARK-59824:
-------------------------------------
Summary: Guard against oversized rows during pickle Python UDF
input serialization
Key: SPARK-59824
URL: https://issues.apache.org/jira/browse/SPARK-59824
Project: Spark
Issue Type: Task
Components: PySpark
Affects Versions: 4.2.0
Reporter: Tengfei Huang
h3. Problem
For non-Arrow Python UDFs, Spark evaluates the UDF input expressions and
serializes the projected row using Java conversion followed by pickle
serialization.
A single row can contain a very large string or binary value. Because batching
cannot split an individual row, converting and pickling it may create several
large in-memory copies and cause an executor OOM before the row reaches the
Python worker.
Existing batch-size limits do not protect against this case.
h3. Proposed change
Add an optional row-size guard for pickle-based Python UDFs.
The guard should:
* Run after UDF input expressions are projected, so computed values such as
{{{}cast{}}}, {{{}concat{}}}, and {{substring}} are measured.
* Run before Java conversion and pickle serialization.
* Estimate the payload size of supported top-level variable-width arguments,
initially strings and binary values.
* Fail with a structured error when the estimated size exceeds a configurable
threshold based on executor heap.
* Introduce no per-row overhead when disabled or when the projected schema has
no supported variable-width arguments.
Nested arrays, maps, and structs are intentionally excluded initially to avoid
a recursive traversal of every input value.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
