cloud-fan commented on code in PR #58962: URL: https://github.com/apache/spark/pull/58962#discussion_r4068944306
########## sql/connect/common/src/main/scala/org/apache/spark/sql/connect/common/UdfSerialization.scala: ########## @@ -0,0 +1,108 @@ +/* + * 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 org.apache.spark.sql.connect.common + +import java.io.{ByteArrayInputStream, InputStream, ObjectInputStream, ObjectStreamClass} + +/** + * Java deserialization for Scala UDF payloads that tolerates the `serialVersionUID` drift of + * `org.apache.spark.sql.types` classes between Spark versions. + * + * A Scala UDF payload embeds `org.apache.spark.sql.types` classes (the UDF's input/output schema). + * Those classes carry no explicit `@SerialVersionUID`, so the JVM auto-computes it from the whole + * class shape, which folds in compiler-synthesized members that are irrelevant to serialization -- + * most notably the `$anonfun$` public static methods Scala emits for lambdas. A source change that + * only reshapes a lambda (e.g. rewriting a helper to use `existsRecursively { ... }`) changes the + * auto-computed `serialVersionUID` without changing any serialized field, which makes a plain + * [[ObjectInputStream]] reject a payload produced by a different Spark version with an + * `InvalidClassException`, even though the payload is field-compatible. + * + * The property that actually governs compatibility is the serialized field layout. When a + * `sql.types` class arrives with a mismatched SUID but an identical serialized field layout, this + * reader rebinds the stream descriptor to the local class; any field-layout difference is left + * untouched so the standard SUID check still fails fast rather than misreading the stream. + * + * Only `org.apache.spark.sql.types` descriptors are treated tolerantly; every other class keeps + * the standard `serialVersionUID` compatibility check. + */ +private[spark] object UdfSerialization { + + private val suidTolerantPackagePrefix = "org.apache.spark.sql.types." + + /** Deserialize `bytes` resolving classes with `loader`, tolerating `sql.types` SUID drift. */ + def deserialize[T](bytes: Array[Byte], loader: ClassLoader): T = { + val ois = new SuidTolerantObjectInputStream(new ByteArrayInputStream(bytes), loader) + try ois.readObject().asInstanceOf[T] + finally ois.close() + } + + /** Deserialize from `in` with default class resolution, tolerating `sql.types` SUID drift. */ + def deserialize[T](in: InputStream): T = { + new SuidTolerantObjectInputStream(in, null).readObject().asInstanceOf[T] + } + + /** + * The complete serialized field layout of a descriptor: the set of persistent field name + JVM + * type signature. This is what governs whether [[ObjectInputStream]] can consume the producer's + * class-data block through the local descriptor. Every persistent slot is included, in + * particular the Scala lazy-val init `bitmap$*` slots: rebinding to a local descriptor whose + * slot shape differs would misalign the stream, so a bitmap difference must also block rebinding. + */ + private[connect] def fieldSignature(desc: ObjectStreamClass): Set[String] = { + // getTypeString is null for primitives, where the single-char type code is the signature. + desc.getFields + .map(f => s"${f.getName}:${Option(f.getTypeString).getOrElse(f.getTypeCode.toString)}") + .toSet + } + + private class SuidTolerantObjectInputStream(in: InputStream, loader: ClassLoader) + extends ObjectInputStream(in) { + + override def resolveClass(desc: ObjectStreamClass): Class[_] = { + if (loader != null) { + // scalastyle:off classforname + Class.forName(desc.getName, false, loader) + // scalastyle:on classforname + } else { + super.resolveClass(desc) + } + } + + override def readClassDescriptor(): ObjectStreamClass = { + val streamDesc = super.readClassDescriptor() + if (!streamDesc.getName.startsWith(suidTolerantPackagePrefix)) { + return streamDesc + } + val localClass = + try { + resolveClass(streamDesc) + } catch { + case _: ClassNotFoundException => return streamDesc + } + val localDesc = ObjectStreamClass.lookup(localClass) + if (localDesc == null || + localDesc.getSerialVersionUID == streamDesc.getSerialVersionUID || + fieldSignature(streamDesc) != fieldSignature(localDesc)) { Review Comment: **Non-blocking (P2):** Persistent-field equality is not sufficient to replace the entire stream descriptor. `Metadata` is already a `sql.types` class with an explicit SUID, and `ObjectInputStream.readSerialData` also uses descriptor state such as `hasReadObjectMethod` and `hasWriteObjectData`. Returning `localDesc` here can therefore suppress a deliberate compatibility break or consume producer data under the local class's protocol, leading to a later stream error or incorrectly populated encoder state. This path should remain fail-closed unless the complete descriptor transition is known to be safe. **Recommended change:** Replace package-wide field-only eligibility with an exact audited class/from-SUID/to-SUID transition policy, retain the field-layout check as an additional guard, and add negative compatibility coverage for explicit-SUID classes and custom class-data protocols. **Why this works:** Key eligibility by class name plus the exact stream and local SUID pair for transitions whose producer and consumer are known to use the same default field-only Java serialization protocol. Require the existing persistent-field equality check as defense in depth. For any unlisted pair, including explicitly versioned production classes such as Metadata and classes with custom serialization hooks, return streamDesc so ObjectInputStream performs its standard compatibility rejection. **Scope:** Constrain whole-descriptor substitution to transitions whose SUID provenance and default serialization protocol have been audited, with regression-sensitive positive and negative fixtures. **Compatibility:** Classes outside sql.types, genuine field-layout mismatches, and authorized auto-computed-SUID drift retain their current helper outcomes. **Risks:** An incomplete transition set may reject a harmless drift until that exact pair is audited. A transition entry that omits either endpoint SUID could accidentally authorize a later incompatible version. **Constraints:** Do not depend on unsupported reflective access to private JDK descriptor flags. Keep standard ObjectInputStream rejection for every transition not affirmatively authorized. **Success:** A production sql.types class with an explicit SUID is not tolerated merely because its package and persistent fields match. A custom writeObject/readObject protocol difference cannot be hidden by local descriptor substitution. Each audited auto-SUID-only transition with an identical default field layout still deserializes successfully. Every unknown or incompatible transition retains streamDesc and uses standard ObjectInputStream checks. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
