diff --git a/.github/workflows/pr_build_linux.yml b/.github/workflows/pr_build_linux.yml
index 52e86f59ffd..55be2b93806 100644
--- a/.github/workflows/pr_build_linux.yml
+++ b/.github/workflows/pr_build_linux.yml
@@ -631,6 +631,8 @@ jobs:
               org.apache.comet.CometFuzzAggregateSuite
               org.apache.spark.sql.comet.execution.arrow.CometArrowStreamSuite
               org.apache.spark.sql.comet.execution.arrow.CachedBatchRowIteratorSuite
+              org.apache.spark.sql.comet.execution.arrow.CacheBatchCoalescerSuite
+              org.apache.spark.sql.comet.execution.arrow.ArrowCachedBatchStatsSuite
               org.apache.spark.sql.comet.execution.arrow.CometArrowWriterSuite
               org.apache.spark.sql.comet.execution.arrow.CometStringWriterSuite
               org.apache.spark.sql.CometSparkInternalFunctionsSuite
diff --git a/.github/workflows/pr_build_macos.yml b/.github/workflows/pr_build_macos.yml
index b1d89ec1240..29e0c34dd9b 100644
--- a/.github/workflows/pr_build_macos.yml
+++ b/.github/workflows/pr_build_macos.yml
@@ -297,6 +297,8 @@ jobs:
               org.apache.comet.CometFuzzAggregateSuite
               org.apache.spark.sql.comet.execution.arrow.CometArrowStreamSuite
               org.apache.spark.sql.comet.execution.arrow.CachedBatchRowIteratorSuite
+              org.apache.spark.sql.comet.execution.arrow.CacheBatchCoalescerSuite
+              org.apache.spark.sql.comet.execution.arrow.ArrowCachedBatchStatsSuite
               org.apache.spark.sql.comet.execution.arrow.CometArrowWriterSuite
               org.apache.spark.sql.comet.execution.arrow.CometStringWriterSuite
               org.apache.spark.sql.CometSparkInternalFunctionsSuite
diff --git a/docs/source/user-guide/latest/in-memory-cache.md b/docs/source/user-guide/latest/in-memory-cache.md
index c1bbb479c42..76cc565c8e1 100644
--- a/docs/source/user-guide/latest/in-memory-cache.md
+++ b/docs/source/user-guide/latest/in-memory-cache.md
@@ -77,6 +77,14 @@ metrics only through that scan.
 
 Each cached batch is stored as a single Arrow IPC record batch message and its body.
 
+A cache built from rows cuts its batches at `spark.sql.inMemoryColumnarStorage.batchSize` rows,
+10000 by default, as Spark's own format does. A cache built from the Arrow batches of Comet
+operators stores a batch of at least half that size as it stands, without copying it, and first
+combines smaller ones up to that size. A shuffle read, for example, returns at least one batch per
+map output block, so a relation cached from a shuffle over many map tasks would otherwise hold
+thousands of small batches, each with its own metadata and statistics. Batches of Spark's own
+column vectors have to be copied into Arrow anyway, so they are combined whatever their size.
+
 The message carries **no Arrow schema**. The reader already has one: `InMemoryRelation` knows the
 cached relation's attributes, and Comet maps them to exactly the Arrow fields the writer produced.
 Storing a schema in every batch would repeat the same bytes once per cached batch — for a wide
diff --git a/spark/src/main/scala/org/apache/spark/sql/comet/execution/arrow/ArrowCachedBatchSerializer.scala b/spark/src/main/scala/org/apache/spark/sql/comet/execution/arrow/ArrowCachedBatchSerializer.scala
index d34e07582ce..624215908ac 100644
--- a/spark/src/main/scala/org/apache/spark/sql/comet/execution/arrow/ArrowCachedBatchSerializer.scala
+++ b/spark/src/main/scala/org/apache/spark/sql/comet/execution/arrow/ArrowCachedBatchSerializer.scala
@@ -23,6 +23,7 @@ import java.lang.{Boolean => JBoolean, Byte => JByte, Double => JDouble, Float =
 
 import scala.collection.JavaConverters._
 
+import org.apache.arrow.vector.DecimalVector
 import org.apache.spark.TaskContext
 import org.apache.spark.rdd.RDD
 import org.apache.spark.sql.catalyst.InternalRow
@@ -35,11 +36,12 @@ import org.apache.spark.sql.internal.SQLConf
 import org.apache.spark.sql.types._
 import org.apache.spark.sql.vectorized.{ColumnarBatch, ColumnVector}
 import org.apache.spark.storage.StorageLevel
+import org.apache.spark.unsafe.Platform
 import org.apache.spark.unsafe.types.UTF8String
 import org.apache.spark.util.io.ChunkedByteBuffer
 
 import org.apache.comet.{CometArrowAllocator, CometConf, DataTypeSupport}
-import org.apache.comet.vector.{CometDecodedVector, NativeUtil}
+import org.apache.comet.vector.{CometDecodedVector, CometPlainVector, NativeUtil}
 
 /**
  * Cached batch format used when Comet writes Spark in-memory cache data.
@@ -298,6 +300,64 @@ class ArrowCachedBatchSerializer extends SimpleMetricsCachedBatchSerializer {
   }
 
   private def gatherDecimalStats(
+      col: ColumnVector,
+      numRows: Int,
+      dt: DecimalType): (Decimal, Decimal, Int) = col match {
+    case v: CometPlainVector if v.getValueVector.isInstanceOf[DecimalVector] =>
+      gatherDecimal128Stats(v, numRows, dt)
+    case _ => gatherDecimalStatsByValue(col, numRows, dt)
+  }
+
+  // Compares the 128-bit unscaled values in Arrow's data buffer rather than building a Decimal for
+  // every row, which for a precision past 18 digits means a BigInteger and a BigDecimal each. All
+  // the values of a column share one scale, so their unscaled order is their numeric order. A
+  // value is a signed high word over an unsigned low word, in little-endian order, and a narrow
+  // decimal is the sign extension of its low word, so one comparison covers every precision. Only
+  // the two bounds become Decimals, through the same getDecimal the by-value loop calls.
+  private def gatherDecimal128Stats(
+      col: CometPlainVector,
+      numRows: Int,
+      dt: DecimalType): (Decimal, Decimal, Int) = {
+    val base = col.getValueVector.getDataBuffer.memoryAddress()
+    var minRow = -1
+    var maxRow = -1
+    var minHi = 0L
+    var minLo = 0L
+    var maxHi = 0L
+    var maxLo = 0L
+    var nullCount = 0
+    var r = 0
+    while (r < numRows) {
+      if (col.isNullAt(r)) {
+        nullCount += 1
+      } else {
+        val lo = Platform.getLong(null, base + r * 16L)
+        val hi = Platform.getLong(null, base + r * 16L + 8L)
+        if (minRow < 0 || hi < minHi || (hi == minHi && JLong.compareUnsigned(lo, minLo) < 0)) {
+          minRow = r
+          minHi = hi
+          minLo = lo
+        }
+        if (maxRow < 0 || hi > maxHi || (hi == maxHi && JLong.compareUnsigned(lo, maxLo) > 0)) {
+          maxRow = r
+          maxHi = hi
+          maxLo = lo
+        }
+      }
+      r += 1
+    }
+    if (minRow < 0) {
+      (null, null, nullCount)
+    } else {
+      (
+        col.getDecimal(minRow, dt.precision, dt.scale),
+        col.getDecimal(maxRow, dt.precision, dt.scale),
+        nullCount)
+    }
+  }
+
+  // The fallback for a decimal column that is not a plain Arrow vector.
+  private def gatherDecimalStatsByValue(
       col: ColumnVector,
       numRows: Int,
       dt: DecimalType): (Decimal, Decimal, Int) = {
@@ -530,16 +590,23 @@ class ArrowCachedBatchSerializer extends SimpleMetricsCachedBatchSerializer {
   // Columnar Comet output is stored as one Arrow IPC record batch message per cached batch. Spark
   // only calls this when supportsColumnarInput returned true, so the schema is known to be
   // Comet-writable here.
+  //
+  // Small input batches are combined first, up to spark.sql.inMemoryColumnarStorage.batchSize
+  // rows, the size the row path cuts at; see CacheBatchCoalescer.
   override def convertColumnarBatchToCachedBatch(
       input: RDD[ColumnarBatch],
       schema: Seq[Attribute],
       storageLevel: StorageLevel,
       conf: SQLConf): RDD[CachedBatch] = {
 
+    val batchSize = conf.columnBatchSize
     val settings = writeSettings(conf)
 
     input.mapPartitions { batches =>
-      encodeBatches(batches, schema, settings)
+      encodeBatches(
+        new CacheBatchCoalescer(batches, schema, batchSize, CometArrowAllocator),
+        schema,
+        settings)
     }
   }
 
diff --git a/spark/src/main/scala/org/apache/spark/sql/comet/execution/arrow/CacheBatchCoalescer.scala b/spark/src/main/scala/org/apache/spark/sql/comet/execution/arrow/CacheBatchCoalescer.scala
new file mode 100644
index 00000000000..584d65e257f
--- /dev/null
+++ b/spark/src/main/scala/org/apache/spark/sql/comet/execution/arrow/CacheBatchCoalescer.scala
@@ -0,0 +1,227 @@
+/*
+ * 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.comet.execution.arrow
+
+import scala.collection.JavaConverters._
+import scala.collection.mutable
+import scala.util.control.NonFatal
+
+import org.apache.arrow.memory.BufferAllocator
+import org.apache.arrow.vector.VectorSchemaRoot
+import org.apache.spark.TaskContext
+import org.apache.spark.sql.catalyst.expressions.Attribute
+import org.apache.spark.sql.comet.util.Utils
+import org.apache.spark.sql.vectorized.ColumnarBatch
+
+import org.apache.comet.vector.NativeUtil
+
+/**
+ * The columnar input of a cache build, with its small batches combined.
+ *
+ * The columnar write stores each batch it is handed as one cached batch, and nothing upstream
+ * bounds how small a batch can be. A shuffle read returns at least one batch per map output
+ * block, so a relation cached from a shuffle over many map tasks would otherwise hold thousands
+ * of batches of a few hundred rows. Each carries its own message metadata, compression framing
+ * and statistics row, and every read of the cache pays for the batch count.
+ *
+ * A batch the writer can store as it stands, Arrow-backed and laid out the way the reader reads
+ * it, is passed through when it holds at least half of `targetRows`, so full native batches are
+ * still written without a copy. Every other batch is copied into batches of `targetRows` rows,
+ * the size the row path cuts at, and split where it crosses that size. A small batch is worth the
+ * copy, and a batch the writer cannot store as it stands has to be copied into Arrow anyway, so
+ * combining it costs nothing more. The copy reads each column through the `ColumnVector`
+ * interface, as `ArrowWriter.writeColumns` does for any input, so it handles dictionary-encoded
+ * and nested columns, and offsets that do not start at zero, as a native producer's sliced arrays
+ * can have. Arrow's `VectorAppender` assumes they start at zero.
+ *
+ * Row order is kept: the rows copied so far are emitted before a batch that passes through. A
+ * large input is copied one output batch at a time, as the consumer asks for them, so the batches
+ * open at once do not grow with its size. The input is not advanced until it has been copied in
+ * full, because a native producer releases the batch it returned last when it is asked for the
+ * next one.
+ *
+ * The batches built here are owned here. Each is released when the next batch is requested, and
+ * whatever is still open when the task completes is released then. A batch passed through still
+ * belongs to its producer.
+ */
+private[arrow] class CacheBatchCoalescer(
+    input: Iterator[ColumnarBatch],
+    attrs: Seq[Attribute],
+    targetRows: Int,
+    allocator: BufferAllocator)
+    extends Iterator[ColumnarBatch] {
+
+  require(targetRows > 0, s"The target batch size must be positive, but was $targetRows")
+
+  private val arrowSchema =
+    Utils.toArrowSchema(Utils.fromAttributes(attrs), CometArrowStream.NATIVE_TIMEZONE)
+  private val readerFields = arrowSchema.getFields.asScala.toIndexedSeq
+
+  // A batch at least this large that the writer can store as it stands is not copied.
+  private val passThroughRows = math.max(1, targetRows / 2)
+
+  // The batch being filled.
+  private var pending: VectorSchemaRoot = null
+  private var pendingWriter: ArrowWriter = null
+  private var pendingRows = 0
+
+  // The input batch being copied, and the first of its rows that has not been copied yet.
+  private var current: ColumnarBatch = null
+  private var currentStart = 0
+
+  // Batches built here and filled, in input order, ahead of `held`.
+  private val filled = mutable.Queue.empty[VectorSchemaRoot]
+
+  // A batch that passes through, waiting for the rows copied ahead of it to be emitted.
+  private var held: ColumnarBatch = null
+
+  // The batch the next call to `next` returns, and the root behind it when it was built here.
+  private var ready: ColumnarBatch = null
+  private var readyRoot: VectorSchemaRoot = null
+
+  // The root behind the batch `next` returned last, released once the consumer asks for more.
+  private var emitted: VectorSchemaRoot = null
+
+  private var closed = false
+
+  Option(TaskContext.get()).foreach(_.addTaskCompletionListener[Unit](_ => close()))
+
+  override def hasNext: Boolean = {
+    if (ready == null && !closed) {
+      release(emitted)
+      emitted = null
+      advance()
+    }
+    ready != null
+  }
+
+  override def next(): ColumnarBatch = {
+    if (!hasNext) {
+      throw new NoSuchElementException("No more cache input batches")
+    }
+    val batch = ready
+    emitted = readyRoot
+    ready = null
+    readyRoot = null
+    batch
+  }
+
+  /** Releases every batch built here that is still open. Idempotent. */
+  def close(): Unit = {
+    if (!closed) {
+      closed = true
+      release(emitted)
+      release(readyRoot)
+      release(pending)
+      filled.foreach(release)
+      filled.clear()
+      emitted = null
+      ready = null
+      readyRoot = null
+      pending = null
+      pendingWriter = null
+      current = null
+      held = null
+    }
+  }
+
+  private def advance(): Unit = {
+    while (filled.isEmpty && held == null && (current != null || input.hasNext)) {
+      if (current != null) {
+        copyRows()
+      } else {
+        val batch = input.next()
+        if (passesThrough(batch)) {
+          finishPending()
+          held = batch
+        } else if (batch.numRows() > 0) {
+          current = batch
+          currentStart = 0
+        }
+      }
+    }
+    if (filled.isEmpty && held == null) {
+      // The input is exhausted.
+      finishPending()
+    }
+    if (filled.nonEmpty) {
+      readyRoot = filled.dequeue()
+      ready = NativeUtil.rootAsBatch(readyRoot)
+    } else if (held != null) {
+      ready = held
+      held = null
+    }
+  }
+
+  private def passesThrough(batch: ColumnarBatch): Boolean =
+    batch.numRows() >= passThroughRows &&
+      Utils.isArrowBacked(batch) &&
+      CachedBatchIpc.matchesReaderLayout(batch, readerFields)
+
+  // Copies rows of `current` until the batch being filled is full or `current` is exhausted.
+  private def copyRows(): Unit = {
+    if (pending == null) {
+      startPending()
+    }
+    val rows = math.min(current.numRows() - currentStart, targetRows - pendingRows)
+    pendingWriter.writeColumns(current, currentStart, rows)
+    pendingRows += rows
+    currentStart += rows
+    if (currentStart == current.numRows()) {
+      current = null
+    }
+    if (pendingRows == targetRows) {
+      finishPending()
+    }
+  }
+
+  private def startPending(): Unit = {
+    val root = VectorSchemaRoot.create(arrowSchema, allocator)
+    try {
+      pendingWriter = ArrowWriter.create(root, targetRows)
+    } catch {
+      case NonFatal(e) =>
+        root.close()
+        throw e
+    }
+    pending = root
+    pendingRows = 0
+  }
+
+  private def finishPending(): Unit = {
+    if (pending != null) {
+      if (pendingRows > 0) {
+        pendingWriter.finish()
+        filled.enqueue(pending)
+      } else {
+        pending.close()
+      }
+      pending = null
+      pendingWriter = null
+      pendingRows = 0
+    }
+  }
+
+  private def release(root: VectorSchemaRoot): Unit = {
+    if (root != null) {
+      root.close()
+    }
+  }
+}
diff --git a/spark/src/test/scala/org/apache/comet/exec/CometInMemoryCacheSuite.scala b/spark/src/test/scala/org/apache/comet/exec/CometInMemoryCacheSuite.scala
index 2e1f4d56d33..e25e8b63d43 100644
--- a/spark/src/test/scala/org/apache/comet/exec/CometInMemoryCacheSuite.scala
+++ b/spark/src/test/scala/org/apache/comet/exec/CometInMemoryCacheSuite.scala
@@ -2248,6 +2248,125 @@ class CometInMemoryCacheSuite extends CometTestBase {
     }
   }
 
+  test("Comet in-memory cache combines small columnar batches in order") {
+    // At a batch size of 1000, an Arrow batch under 500 rows is combined with its neighbours, a
+    // larger one is stored as it stands after the rows combined ahead of it, and a batch of
+    // Spark's own vectors, which has to be copied into Arrow anyway, is combined at any size.
+    // Combined batches are cut at 1000 rows. The input releases each batch when it is asked for
+    // the next one, as a native producer does, so a batch held rather than copied would be read
+    // after it was released (#6950).
+    val input = Seq(
+      300 -> true,
+      0 -> true,
+      300 -> true,
+      9000 -> true,
+      700 -> true,
+      400 -> true,
+      2000 -> false,
+      400 -> true,
+      300 -> true)
+    val attrs = Seq(AttributeReference("id", LongType)(), AttributeReference("s", StringType)())
+    val serializer = new ArrowCachedBatchSerializer
+    withSQLConf(SQLConf.COLUMN_BATCH_SIZE.key -> "1000") {
+      val cached = serializer.convertColumnarBatchToCachedBatch(
+        spark.sparkContext.parallelize(Seq(0), 1).mapPartitions(_ => new ReleasingBatches(input)),
+        attrs,
+        StorageLevel.MEMORY_ONLY,
+        SQLConf.get)
+      assert(cached.map(_.numRows).collect().toSeq == Seq(600, 9000, 700, 1000, 1000, 1000, 100))
+
+      val rows = serializer
+        .convertCachedBatchToColumnarBatch(cached, attrs, attrs, SQLConf.get)
+        .mapPartitions(_.flatMap { batch =>
+          (0 until batch.numRows()).map(i =>
+            (batch.column(0).getLong(i), batch.column(1).getUTF8String(i).toString))
+        })
+        .collect()
+      assert(rows.toSeq == (0L until input.map(_._1).sum.toLong).map(i => (i, s"v$i")))
+    }
+  }
+
+  test("Comet in-memory cache combines the small batches of a shuffle read") {
+    // A shuffle read returns at least one batch per map output block, so a relation cached
+    // straight from a shuffle over 20 map tasks into 2 partitions arrives in 40 batches of about
+    // 500 rows. They are combined up to the default batch size of 10000 rows (#6950).
+    withTempPath { path =>
+      withSQLConf(CometConf.COMET_ENABLED.key -> "false") {
+        spark
+          .range(0, 20000, 1, 20)
+          .selectExpr(
+            "id",
+            "cast(id % 1000 as int) AS k",
+            "if(id % 11 = 0, NULL, concat('s', id)) AS s",
+            "named_struct('a', id, 'b', cast(id AS string)) AS st",
+            "if(id % 13 = 0, NULL, array(id, id + 1)) AS arr",
+            "map(cast(id % 7 AS string), id) AS mp")
+          .write
+          .parquet(path.toString)
+      }
+      withSQLConf(
+        SQLConf.ADAPTIVE_EXECUTION_ENABLED.key -> "false",
+        // One map task per file.
+        SQLConf.FILES_MAX_PARTITION_BYTES.key -> "1024") {
+        val query = s"SELECT /*+ REPARTITION(2, k) */ * FROM parquet.`${path.toString}`"
+        val expected = uncachedSparkAnswer(query)
+        val df = spark.sql(query)
+        assert(df.rdd.getNumPartitions == 2)
+        df.cache()
+        try {
+          assert(df.count() == 20000)
+          val relation = spark.sharedState.cacheManager.lookupCachedData(df).get
+          val cached = relation.cachedRepresentation.cacheBuilder.cachedColumnBuffers
+          assert(
+            cached.map(_.getClass.getName).distinct().collect().toSeq ==
+              Seq("org.apache.spark.sql.comet.execution.arrow.CometCachedBatch"))
+          val sizes = cached.map(_.numRows).collect()
+          assert(sizes.length <= 4 && sizes.forall(_ <= 10000), sizes.mkString(", "))
+          assert(QueryTest.sameRows(expected.toSeq, df.collect().toSeq).isEmpty)
+        } finally {
+          df.unpersist(blocking = true)
+        }
+      }
+    }
+  }
+
+  test("Comet in-memory cache keeps a sorted cache sorted when it combines its batches") {
+    // A cache reports the order of the plan it was built from, so the rows of the batches it
+    // combines have to keep their order. The native sort emits batches of 100 rows here.
+    withTempPath { path =>
+      withSQLConf(CometConf.COMET_ENABLED.key -> "false") {
+        spark.range(0, 20000, 1, 4).selectExpr("id", "id % 97 AS k").write.parquet(path.toString)
+      }
+      withSQLConf(
+        SQLConf.ADAPTIVE_EXECUTION_ENABLED.key -> "false",
+        CometConf.COMET_BATCH_SIZE.key -> "100") {
+        val df = spark.read
+          .parquet(path.toString)
+          .repartition(2, $"k")
+          .sortWithinPartitions($"id".desc)
+        df.cache()
+        try {
+          assert(df.count() == 20000)
+          val relation = spark.sharedState.cacheManager.lookupCachedData(df).get
+          val sizes =
+            relation.cachedRepresentation.cacheBuilder.cachedColumnBuffers
+              .map(_.numRows)
+              .collect()
+          assert(sizes.length <= 4, sizes.mkString(", "))
+          val sorted = df.rdd
+            .mapPartitions { rows =>
+              val ids = rows.map(_.getLong(0)).toArray
+              Iterator(ids.toSeq == ids.sorted.reverse.toSeq)
+            }
+            .collect()
+          assert(sorted.toSeq == Seq(true, true))
+        } finally {
+          df.unpersist(blocking = true)
+        }
+      }
+    }
+  }
+
   // Enough rows that every column's buffers are big enough for Arrow to actually compress them.
   // Arrow stores a buffer verbatim when compressing it would not make it smaller, and a boolean
   // column of a few hundred rows is a few dozen bytes, which takes that fallback -- leaving the
diff --git a/spark/src/test/scala/org/apache/comet/exec/ReleasingBatches.scala b/spark/src/test/scala/org/apache/comet/exec/ReleasingBatches.scala
new file mode 100644
index 00000000000..82095bed451
--- /dev/null
+++ b/spark/src/test/scala/org/apache/comet/exec/ReleasingBatches.scala
@@ -0,0 +1,89 @@
+/*
+ * 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.comet.exec
+
+import java.nio.charset.StandardCharsets
+
+import org.apache.arrow.vector.{BigIntVector, VarCharVector}
+import org.apache.spark.sql.execution.vectorized.OnHeapColumnVector
+import org.apache.spark.sql.types.{LongType, StringType}
+import org.apache.spark.sql.vectorized.{ColumnarBatch, ColumnVector}
+
+import org.apache.comet.CometArrowAllocator
+import org.apache.comet.vector.CometPlainVector
+
+/**
+ * Batches of `id`, counting from 0, and `s`, "v" followed by the id, one batch per element of
+ * `sizes`: a row count, and whether the batch holds Arrow vectors or Spark's own on-heap ones.
+ * Like a native producer, it releases each batch when it is asked for the next one.
+ */
+private class ReleasingBatches(sizes: Seq[(Int, Boolean)]) extends Iterator[ColumnarBatch] {
+  private var remaining = sizes.toList
+  private var nextId = 0L
+  private var previous: ColumnarBatch = null
+
+  override def hasNext: Boolean = {
+    if (previous != null) {
+      previous.close()
+      previous = null
+    }
+    remaining.nonEmpty
+  }
+
+  override def next(): ColumnarBatch = {
+    if (!hasNext) {
+      throw new NoSuchElementException("No more batches")
+    }
+    val (rows, arrow) = remaining.head
+    remaining = remaining.tail
+    val batch = if (arrow) arrowBatch(rows) else sparkBatch(rows)
+    nextId += rows
+    previous = batch
+    batch
+  }
+
+  private def value(i: Int): Array[Byte] = s"v${nextId + i}".getBytes(StandardCharsets.UTF_8)
+
+  private def arrowBatch(rows: Int): ColumnarBatch = {
+    val ids = new BigIntVector("id", CometArrowAllocator)
+    val strings = new VarCharVector("s", CometArrowAllocator)
+    ids.allocateNew(rows)
+    strings.allocateNew(rows.toLong * 8, rows)
+    (0 until rows).foreach { i =>
+      ids.set(i, nextId + i)
+      strings.setSafe(i, value(i))
+    }
+    ids.setValueCount(rows)
+    strings.setValueCount(rows)
+    new ColumnarBatch(
+      Array[ColumnVector](new CometPlainVector(ids), new CometPlainVector(strings)),
+      rows)
+  }
+
+  private def sparkBatch(rows: Int): ColumnarBatch = {
+    val ids = new OnHeapColumnVector(rows, LongType)
+    val strings = new OnHeapColumnVector(rows, StringType)
+    (0 until rows).foreach { i =>
+      ids.putLong(i, nextId + i)
+      strings.putByteArray(i, value(i))
+    }
+    new ColumnarBatch(Array[ColumnVector](ids, strings), rows)
+  }
+}
diff --git a/spark/src/test/scala/org/apache/spark/sql/benchmark/CometInMemoryCacheBenchmark.scala b/spark/src/test/scala/org/apache/spark/sql/benchmark/CometInMemoryCacheBenchmark.scala
index 187572c853f..b4aa3ee7aeb 100644
--- a/spark/src/test/scala/org/apache/spark/sql/benchmark/CometInMemoryCacheBenchmark.scala
+++ b/spark/src/test/scala/org/apache/spark/sql/benchmark/CometInMemoryCacheBenchmark.scala
@@ -39,7 +39,7 @@ import org.apache.spark.sql.execution.aggregate.HashAggregateExec
 import org.apache.spark.sql.execution.columnar.{CometInMemoryRelationHelper, DefaultCachedBatch, DefaultCachedBatchSerializer, InMemoryRelation, InMemoryTableScanExec}
 import org.apache.spark.sql.execution.vectorized.OnHeapColumnVector
 import org.apache.spark.sql.internal.{SQLConf, StaticSQLConf}
-import org.apache.spark.sql.types.{ArrayType, BinaryType, DataType, IntegerType, LongType, StringType, StructType}
+import org.apache.spark.sql.types.{ArrayType, BinaryType, DataType, Decimal, DecimalType, IntegerType, LongType, StringType, StructType}
 import org.apache.spark.sql.vectorized.{ColumnarBatch, ColumnVector}
 import org.apache.spark.unsafe.types.UTF8String
 
@@ -876,6 +876,42 @@ object CometInMemoryCacheBenchmark extends CometBenchmarkBase {
       addStatsCase(benchmark, "Arrow vectors", nestedBatch, nestedAttrs)
       benchmark.run()
     } finally nestedBatch.close()
+
+    // Decimals of one narrow and two wide precisions, one row in ten null and the values on both
+    // sides of zero. A decimal past 18 digits is the one the by-value loop paid a BigInteger and a
+    // BigDecimal for on every row.
+    val decimalTypes: Seq[DataType] =
+      Seq(DecimalType(10, 2), DecimalType(18, 2), DecimalType(38, 2))
+    val decimalAttrs = decimalTypes.zipWithIndex.map { case (dt, i) =>
+      AttributeReference(s"d$i", dt)()
+    }
+    val decimalRows = Iterator.tabulate(batchSize) { r =>
+      if (r % 10 == 0) {
+        new GenericInternalRow(decimalTypes.length)
+      } else {
+        val unscaled = BigInt(r - batchSize / 2) * 7919
+        InternalRow(
+          Decimal(unscaled.toLong, 10, 2),
+          Decimal(unscaled.toLong * 100000, 18, 2),
+          Decimal(BigDecimal(unscaled * BigInt(10).pow(20) + r, 2), 38, 2))
+      }
+    }
+    val decimalBatch = CometArrowConverters
+      .rowToArrowBatchIter(
+        decimalRows,
+        Utils.fromAttributes(decimalAttrs),
+        batchSize,
+        CometArrowStream.NATIVE_TIMEZONE,
+        CometArrowAllocator)
+      .next()
+    try {
+      val benchmark = new Benchmark(
+        "in-memory cache statistics, decimal columns",
+        numRows.toLong,
+        output = output)
+      addStatsCase(benchmark, "Arrow vectors", decimalBatch, decimalAttrs)
+      benchmark.run()
+    } finally decimalBatch.close()
   }
 
   private def runCacheBenchmark(
diff --git a/spark/src/test/scala/org/apache/spark/sql/comet/execution/arrow/ArrowCachedBatchStatsSuite.scala b/spark/src/test/scala/org/apache/spark/sql/comet/execution/arrow/ArrowCachedBatchStatsSuite.scala
new file mode 100644
index 00000000000..55eaacfc223
--- /dev/null
+++ b/spark/src/test/scala/org/apache/spark/sql/comet/execution/arrow/ArrowCachedBatchStatsSuite.scala
@@ -0,0 +1,159 @@
+/*
+ * 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.comet.execution.arrow
+
+import java.math.{BigDecimal => JBigDecimal, BigInteger}
+
+import scala.util.Random
+
+import org.scalatest.funsuite.AnyFunSuite
+
+import org.apache.arrow.c.CDataDictionaryProvider
+import org.apache.arrow.vector.DecimalVector
+import org.apache.spark.sql.catalyst.expressions.AttributeReference
+import org.apache.spark.sql.types.{Decimal, DecimalType}
+import org.apache.spark.sql.vectorized.{ColumnarBatch, ColumnVector}
+
+import org.apache.comet.CometArrowAllocator
+import org.apache.comet.vector.CometVector
+
+/** The bounds and null counts `ArrowCachedBatchSerializer` records for decimal columns. */
+class ArrowCachedBatchStatsSuite extends AnyFunSuite {
+
+  private val serializer = new ArrowCachedBatchSerializer
+
+  // Values that sit on either side of the boundaries between the two 64-bit words of a 128-bit
+  // unscaled value: the low word's sign bit, a carry into the high word, and both signs of each.
+  private def boundaryValues(precision: Int): Seq[BigInteger] = {
+    val two63 = BigInteger.ONE.shiftLeft(63)
+    val two64 = BigInteger.ONE.shiftLeft(64)
+    val limit = BigInteger.TEN.pow(precision).subtract(BigInteger.ONE)
+    val candidates = Seq(
+      BigInteger.ZERO,
+      BigInteger.ONE,
+      two63.subtract(BigInteger.ONE),
+      two63,
+      two63.add(BigInteger.ONE),
+      two64.subtract(BigInteger.ONE),
+      two64,
+      two64.add(BigInteger.ONE),
+      two64.multiply(BigInteger.valueOf(3)).add(two63))
+    val signed = candidates ++ candidates.map(_.negate())
+    (signed :+ limit :+ limit.negate()).filter(_.abs.compareTo(limit) <= 0).distinct
+  }
+
+  private def randomValues(precision: Int, count: Int, seed: Long): Seq[BigInteger] = {
+    val random = new Random(seed)
+    val limit = BigInteger.TEN.pow(precision)
+    Seq.fill(count) {
+      // A random width, so the values spread over the words instead of all filling the high one.
+      val bits = 1 + random.nextInt(limit.bitLength())
+      val magnitude = new BigInteger(bits, new java.util.Random(random.nextLong())).mod(limit)
+      if (random.nextBoolean()) magnitude else magnitude.negate()
+    }
+  }
+
+  // Gathers the statistics of one decimal column holding `values`, with a null before and after
+  // every `nullEvery`-th value when `nullEvery` is positive.
+  private def check(precision: Int, scale: Int, values: Seq[BigInteger], nullEvery: Int): Unit = {
+    val rows: Seq[Option[BigInteger]] =
+      if (nullEvery > 0) {
+        values.zipWithIndex.flatMap { case (v, i) =>
+          if (i % nullEvery == 0) Seq(None, Some(v), None) else Seq(Some(v))
+        }
+      } else {
+        values.map(Some(_))
+      }
+    val allocator = CometArrowAllocator.newChildAllocator("decimal-stats", 0, Long.MaxValue)
+    val vector = new DecimalVector("d", allocator, precision, scale)
+    try {
+      vector.allocateNew(rows.length)
+      rows.zipWithIndex.foreach {
+        case (None, i) => vector.setNull(i)
+        case (Some(v), i) => vector.setSafe(i, new JBigDecimal(v, scale))
+      }
+      vector.setValueCount(rows.length)
+
+      val column = CometVector.getVector(vector, new CDataDictionaryProvider)
+      val batch = new ColumnarBatch(Array[ColumnVector](column), rows.length)
+      val attrs = Seq(AttributeReference("d", DecimalType(precision, scale))())
+      val (lower, upper, nulls) =
+        serializer.gatherColumnStats(batch, attrs, serializer.boundsOrderings(attrs))
+
+      val present = rows.flatten
+      assert(nulls(0) == rows.length - present.length)
+      if (present.isEmpty) {
+        assert(lower(0) == null && upper(0) == null)
+      } else {
+        val expectedMin = new JBigDecimal(present.min, scale)
+        val expectedMax = new JBigDecimal(present.max, scale)
+        val actualMin = lower(0).asInstanceOf[Decimal]
+        val actualMax = upper(0).asInstanceOf[Decimal]
+        val context = s"decimal($precision, $scale), ${rows.length} rows"
+        assert(actualMin.toJavaBigDecimal.compareTo(expectedMin) == 0, s"min of $context")
+        assert(actualMax.toJavaBigDecimal.compareTo(expectedMax) == 0, s"max of $context")
+        assert(
+          actualMin.precision == precision && actualMin.scale == scale,
+          s"min type, $context")
+        assert(
+          actualMax.precision == precision && actualMax.scale == scale,
+          s"max type, $context")
+      }
+    } finally {
+      vector.close()
+      allocator.close()
+    }
+  }
+
+  // Both a narrow precision, whose unscaled value fits the low word, and a wide one, for which the
+  // high word decides the order and the low word is compared as unsigned.
+  private val precisions =
+    Seq((5, 2), (10, 2), (18, 0), (18, 6), (19, 0), (20, 4), (30, 10), (38, 2))
+
+  precisions.foreach { case (precision, scale) =>
+    test(s"decimal($precision, $scale) bounds across the word boundaries") {
+      val values = boundaryValues(precision)
+      check(precision, scale, values, nullEvery = 0)
+      // Every order of the same values must give the same bounds.
+      check(precision, scale, values.reverse, nullEvery = 0)
+      check(precision, scale, new Random(7).shuffle(values), nullEvery = 3)
+
+      // Without the largest magnitudes, which decide the bounds by their high word alone, the
+      // values are all near the word boundaries. Of one sign, the high words are equal for many of
+      // them, so the low word decides, and it is unsigned.
+      val inner = values.filter(_.abs.bitLength <= 66)
+      check(precision, scale, inner.filter(_.signum >= 0), nullEvery = 0)
+      check(precision, scale, inner.filter(_.signum <= 0), nullEvery = 0)
+      check(precision, scale, inner, nullEvery = 2)
+    }
+
+    test(s"decimal($precision, $scale) bounds of random values and nulls") {
+      (0 until 5).foreach { seed =>
+        check(precision, scale, randomValues(precision, 500, seed.toLong), nullEvery = 0)
+        check(precision, scale, randomValues(precision, 500, seed.toLong), nullEvery = 7)
+      }
+    }
+  }
+
+  test("decimal bounds of a column with a single value") {
+    check(38, 2, Seq(BigInteger.ONE.shiftLeft(100).negate()), nullEvery = 0)
+    check(10, 2, Seq(BigInteger.valueOf(-5)), nullEvery = 1)
+  }
+}
diff --git a/spark/src/test/scala/org/apache/spark/sql/comet/execution/arrow/CacheBatchCoalescerSuite.scala b/spark/src/test/scala/org/apache/spark/sql/comet/execution/arrow/CacheBatchCoalescerSuite.scala
new file mode 100644
index 00000000000..cf351fd42ce
--- /dev/null
+++ b/spark/src/test/scala/org/apache/spark/sql/comet/execution/arrow/CacheBatchCoalescerSuite.scala
@@ -0,0 +1,116 @@
+/*
+ * 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.comet.execution.arrow
+
+import java.nio.charset.StandardCharsets.UTF_8
+
+import org.scalatest.funsuite.AnyFunSuite
+
+import org.apache.arrow.memory.RootAllocator
+import org.apache.spark.sql.catalyst.expressions.AttributeReference
+import org.apache.spark.sql.execution.vectorized.OnHeapColumnVector
+import org.apache.spark.sql.types.StringType
+import org.apache.spark.sql.vectorized.{ColumnarBatch, ColumnVector}
+
+class CacheBatchCoalescerSuite extends AnyFunSuite {
+
+  private val numColumns = 4
+  private val attrs = (0 until numColumns).map(i => AttributeReference(s"s$i", StringType)())
+
+  // A batch of Spark's own on-heap vectors, which the coalescer always copies. Row `i` of column
+  // `c` holds "c<c>r<first + i>", so the output order can be checked.
+  private def sparkBatch(first: Int, rows: Int): ColumnarBatch = {
+    val columns = (0 until numColumns).map { c =>
+      val vector = new OnHeapColumnVector(rows, StringType)
+      (0 until rows).foreach(i => vector.putByteArray(i, s"c${c}r${first + i}".getBytes(UTF_8)))
+      vector: ColumnVector
+    }
+    new ColumnarBatch(columns.toArray, rows)
+  }
+
+  // Runs `f` on a coalescer over `inputs`, checks that closing it returns every byte it took, and
+  // returns the most the allocator held at once.
+  private def withCoalescer(limit: Long, inputs: Seq[ColumnarBatch], target: Int)(
+      f: CacheBatchCoalescer => Unit): Long = {
+    val allocator = new RootAllocator(limit)
+    val coalescer = new CacheBatchCoalescer(inputs.iterator, attrs, target, allocator)
+    try {
+      f(coalescer)
+    } finally {
+      coalescer.close()
+    }
+    assert(allocator.getAllocatedMemory == 0L)
+    val peak = allocator.getPeakMemoryAllocation
+    allocator.close()
+    peak
+  }
+
+  test("splitting one large input copies one output batch at a time") {
+    val inputRows = 4096
+    val target = 10
+    // The limit is far above what one output batch needs and far below what 410 of them need.
+    val limit = 4L * 1024 * 1024
+    val peak = withCoalescer(limit, Seq(sparkBatch(0, inputRows)), target) { coalescer =>
+      var rows = 0
+      var outputs = 0
+      while (coalescer.hasNext) {
+        val out = coalescer.next()
+        assert(out.numRows() <= target)
+        (0 until out.numRows()).foreach { i =>
+          (0 until numColumns).foreach { c =>
+            assert(out.column(c).getUTF8String(i).toString == s"c${c}r${rows + i}")
+          }
+        }
+        rows += out.numRows()
+        outputs += 1
+      }
+      assert(rows == inputRows)
+      assert(outputs == (inputRows + target - 1) / target)
+    }
+    // Only the batch being emitted is open, so the allocation does not grow with the number of
+    // outputs. Eagerly splitting the input would hold all of them at once.
+    assert(peak < 1024 * 1024, peak)
+  }
+
+  test("rows left over from one input are filled up from the next") {
+    val inputs = Seq(sparkBatch(0, 25), sparkBatch(25, 7), sparkBatch(32, 30))
+    withCoalescer(Long.MaxValue, inputs, 10) { coalescer =>
+      val sizes = Seq.newBuilder[Int]
+      var rows = 0
+      while (coalescer.hasNext) {
+        val out = coalescer.next()
+        (0 until out.numRows()).foreach { i =>
+          assert(out.column(0).getUTF8String(i).toString == s"c0r${rows + i}")
+        }
+        rows += out.numRows()
+        sizes += out.numRows()
+      }
+      assert(sizes.result() == Seq(10, 10, 10, 10, 10, 10, 2))
+    }
+  }
+
+  test("an empty input batch is skipped") {
+    val inputs = Seq(sparkBatch(0, 0), sparkBatch(0, 3), sparkBatch(3, 0))
+    withCoalescer(Long.MaxValue, inputs, 10) { coalescer =>
+      assert(coalescer.next().numRows() == 3)
+      assert(!coalescer.hasNext)
+    }
+  }
+}