What causes "Columnar reads are not supported" in apache/iceberg?

Ensure supportColumnarReads() returns false for scans using this factory so Spark never requests columnar readers

apache/iceberg · error · UnsupportedOperationException

Columnar reads are not supported

Error message

Columnar reads are not supported

What it means

SparkRowReaderFactory explicitly does not support columnar reads: createColumnarReader() always throws UnsupportedOperationException('Columnar reads are not supported'), and supportColumnarReads() returns false. This factory exists solely for row-based reads (changelog/position-deletes).

Source

Thrown at spark/v4.2/spark/src/main/java/org/apache/iceberg/spark/source/SparkRowReaderFactory.java:61

    if (partition.allTasksOfType(FileScanTask.class)) {
      return new RowDataReader(partition);

    } else if (partition.allTasksOfType(ChangelogScanTask.class)) {
      return new ChangelogRowReader(partition);

    } else if (partition.allTasksOfType(PositionDeletesScanTask.class)) {
      return new PositionDeletesRowReader(partition);

    } else {
      throw new UnsupportedOperationException(
          "Unsupported task group for row-based reads: " + partition.taskGroup());
    }
  }

  @Override
  public PartitionReader<ColumnarBatch> createColumnarReader(InputPartition inputPartition) {
    throw new UnsupportedOperationException("Columnar reads are not supported");
  }

  @Override
  public boolean supportColumnarReads(InputPartition inputPartition) {
    return false;
  }
}

View on GitHub (pinned to 86d9c8fc54)

Solutions

Ensure supportColumnarReads() returns false for scans using this factory so Spark never requests columnar readers
Do not force columnar batch options on changelog/position-delete scans
Use the standard SparkTable/SparkBatchQueryScan path which selects the correct reader factory
If you need columnar reads, use the regular batch scan factory instead of the row reader factory

Example fix

// before
@Override
public boolean supportColumnarReads(InputPartition p) { return true; } // leads to throw
// after
@Override
public boolean supportColumnarReads(InputPartition p) { return false; }

Defensive patterns

Strategy: type-guard

Validate before calling

if (factory.supportColumnarReads(partition)) { reader = factory.createColumnarReader(partition); } else { reader = factory.createReader(partition); }

Type guard

boolean supportsColumnar = (factory instanceof SparkRowReaderFactory) ? false : factory.supportColumnarReads(partition);

Try / catch

try { return factory.createColumnarReader(p); } catch (UnsupportedOperationException e) { return factory.createReader(p); }

Prevention

Never advertise supportColumnarReads=true for factories that only produce rows
Don't force columnar batch options on changelog/position-delete scans
Use the standard batch scan path for columnar reads

When it happens

Trigger: Spark's DataSource V2 batch engine calls createColumnarReader because supportColumnarReads returned true or columnar read options were forced (e.g. spark.sql.sources.v2.batch.scan.enabled with columnar batching, or a custom reader wrapping this factory).

Common situations: Custom scan implementations advertising columnar support but delegating to SparkRowReaderFactory; misconfigured batch read options requesting ColumnarBatch output for changelog/position-delete reads.

Related errors

AI-assisted analysis of apache/iceberg@86d9c8fc54 (2026-09-12). Data as JSON: /api/errors/7986212070a4e922. Report an issue: GitHub.