[filesystem] HadoopDataInputStream.seek loops forever past EOF

[mattfaltyn]

Search before asking

• I searched in the issues and found nothing similar.

I searched titles, bodies, and comments for HadoopDataInputStream, skipFully, seek past EOF, HDFS seek hang, Hadoop skip EOF, and related filesystem terms. I also inspected the open Hadoop dependency-alignment PR #3699; it does not touch this code path.

Fluss version

• main (development) at 31621117db9b21e8e00a6e716b3a13b7b66e18be
• The same implementation is present in v0.9.1-incubating, v0.9.0-incubating, and v0.8.0-incubating

Please describe the bug 🐞

HadoopDataInputStream.seek(long) optimizes small forward seeks by calling skipFully(delta). If the requested position is past EOF, the underlying InputStream.skip(long) returns 0. The loop in skipFully then makes no progress and spins forever:

fluss/fluss-filesystems/fluss-fs-hadoop/src/main/java/org/apache/fluss/fs/hdfs/HadoopDataInputStream.java

Lines 132 to 135 in 3162111

	public void skipFully(long bytes) throws IOException {
	while (bytes > 0) {
	bytes -= fsDataInputStream.skip(bytes);
	}

This contradicts the public FSDataInputStream.seek contract, which says callers cannot seek past the end of the stream and reports seek errors through IOException:

fluss/fluss-common/src/main/java/org/apache/fluss/fs/FSDataInputStream.java

Lines 36 to 43 in 3162111

	/**
	* Seek to the given offset from the start of the file. The next read() will be from that
	* location. Can't seek past the end of the stream.
	*
	* @param desired the desired offset
	* @throws IOException Thrown if an error occurred while seeking inside the input stream.
	*/
	public abstract void seek(long desired) throws IOException;

Java's InputStream.skip contract permits returning fewer bytes than requested, including 0.

Reproduction

The existing SeekableByteArrayInputStream test fixture already returns 0 from skip at EOF. Adding this focused case to HadoopDataInputStreamTest reproduces the hang:

@Test
void testSeekPastEndOfStream() {
    FSDataInputStream input =
            new FSDataInputStream(new SeekableByteArrayInputStream(new byte[1]));
    HadoopDataInputStream stream = new HadoopDataInputStream(input);

    assertThatThrownBy(() -> stream.seek(2)).isInstanceOf(EOFException.class);
}

Run it with:

./mvnw -pl fluss-filesystems/fluss-fs-hadoop -am \
  -Dtest=HadoopDataInputStreamTest \
  -Dsurefire.failIfNoSpecifiedTests=false test

I also reproduced this twice against Hadoop's real local filesystem implementation using a one-byte file and seek(2). Both runs timed out after two seconds:

before seek target=2 length=1
TIMEOUT after 2s (seek did not return)

The nearby control seek(1) returned normally with position 1.

Expected behavior

The call should promptly throw IOException/EOFException because the requested position is past EOF.

Actual behavior

seek(2) never returns. The thread remains in the skipFully loop.

No special configuration or external service is required.

Solution

Delegate skipFully to Hadoop's existing org.apache.hadoop.io.IOUtils.skipFully(InputStream, long), or implement the same zero-progress handling: when skip returns 0, read one byte and throw EOFException if EOF has been reached.

A regression test can be added to the existing HadoopDataInputStreamTest. This should require no new dependency and no API or storage-format change.

Are you willing to submit a PR?

• I'm willing to submit a PR!

[wolfkill]

Opened a draft PR for this: #3738

Root cause: HadoopDataInputStream.seek() uses skipFully(delta) for small forward seeks, but skipFully() assumed each FSDataInputStream.skip(bytes) call makes progress. When the underlying stream returns 0 at EOF or no-progress conditions, the remaining byte count never decreases and the loop can spin forever.

The PR falls back to forceSeek(seekPos) when skip() returns 0 or a negative value, letting the Hadoop stream either reach the requested position or raise its normal out-of-bounds/EOF error.

Validation:

• RED: JAVA_HOME=/Users/cc/Library/Java/JavaVirtualMachines/corretto-18.0.2/Contents/Home mvn -pl fluss-filesystems/fluss-fs-hadoop -am -Dtest=HadoopDataInputStreamTest#testSeekFallsBackWhenSkipMakesNoProgress -DfailIfNoTests=false test failed before the fix because skip() was retried after making no progress.
• JAVA_HOME=/Users/cc/Library/Java/JavaVirtualMachines/corretto-18.0.2/Contents/Home mvn -pl fluss-filesystems/fluss-fs-hadoop -am -Dtest=HadoopDataInputStreamTest#testSeekFallsBackWhenSkipMakesNoProgress -DfailIfNoTests=false test
• JAVA_HOME=/Users/cc/Library/Java/JavaVirtualMachines/corretto-18.0.2/Contents/Home mvn -pl fluss-filesystems/fluss-fs-hadoop -am -Dtest=HadoopDataInputStreamTest -DfailIfNoTests=false test
• JAVA_HOME=/Users/cc/Library/Java/JavaVirtualMachines/corretto-18.0.2/Contents/Home mvn -pl fluss-filesystems/fluss-fs-hadoop -am -DfailIfNoTests=false test
• git diff --check