perf: optimize JVM columnar-to-row conversion

.github/workflows/pr_build_linux.yml

@@ -326,6 +326,7 @@ jobs:
               org.apache.comet.exec.CometShuffleSuite
               org.apache.comet.exec.CometShuffle4_0Suite
               org.apache.comet.exec.CometNativeColumnarToRowSuite
+              org.apache.comet.exec.CometDirectColumnarToRowSuite
               org.apache.comet.exec.CometNativeShuffleSuite
               org.apache.spark.sql.comet.execution.shuffle.CometNativeShuffleInputRDDSuite
               org.apache.comet.exec.CometShuffleEncryptionSuite
@@ -344,6 +345,7 @@ jobs:
               org.apache.comet.exec.CometJoinSuite
               org.apache.spark.sql.comet.CometMapInBatchSuite
               org.apache.comet.CometNativeSuite
+              org.apache.comet.DirectColumnarToRowConverterSuite
               org.apache.comet.CometConfSuite
               org.apache.comet.CometPublicApiSuite
               org.apache.comet.QueryContextInternerSuite

.github/workflows/pr_build_macos.yml

@@ -142,6 +142,7 @@ jobs:
               org.apache.comet.exec.CometShuffleSuite
               org.apache.comet.exec.CometShuffle4_0Suite
               org.apache.comet.exec.CometNativeColumnarToRowSuite
+              org.apache.comet.exec.CometDirectColumnarToRowSuite
               org.apache.comet.exec.CometNativeShuffleSuite
               org.apache.spark.sql.comet.execution.shuffle.CometNativeShuffleInputRDDSuite
               org.apache.comet.exec.CometShuffleEncryptionSuite
@@ -160,6 +161,7 @@ jobs:
               org.apache.comet.exec.CometJoinSuite
               org.apache.spark.sql.comet.CometMapInBatchSuite
               org.apache.comet.CometNativeSuite
+              org.apache.comet.DirectColumnarToRowConverterSuite
               org.apache.comet.CometConfSuite
               org.apache.comet.CometPublicApiSuite
               org.apache.comet.QueryContextInternerSuite

docs/source/user-guide/latest/tuning.md

@@ -284,6 +284,21 @@ plan to fall back to Spark row-based execution — Comet removes its native oper
 mix of native and fallback operators joined by repeated conversions — which can be cheaper than paying the
 expensive conversions again and again.
 
+### Experimental: Direct Columnar-to-Row Conversion
+
+When the JVM columnar-to-row operator is in use (`spark.comet.exec.columnarToRow.native.enabled=false`),
+setting `spark.comet.exec.columnarToRow.direct.enabled=true` enables an experimental converter that writes
+values straight from Arrow buffers into Spark's row format without allocating an object per value. This is
+most beneficial for decimal-heavy schemas, where the default conversion allocates a `Decimal` object per
+value (and considerably more for decimals with precision above 18); microbenchmarks show up to 2x faster
+conversion and a large reduction in garbage creation for such schemas. Schemas containing data types the
+converter does not support fall back to the default conversion automatically.
+
+Batches with fewer rows than `spark.comet.exec.columnarToRow.direct.minBatchSize` (default `128`) also fall
+back to the default conversion, since the direct converter's per-batch setup does not pay off on very small
+batches. This optimization is experimental: it only affects the operator's non-codegen paths (including
+broadcast relation builds), and the default conversion remains enabled unless explicitly opted in.
+
 ## Metrics Overhead
 
 Comet exposes rich native operator metrics for observability (see [Metrics](metrics.md)), but they are