TFRecord Performance Tests doesn't work on hdfs

TFRecord have issue reading files from hdfs using filename pattern _"hdfs://...*"_
```

TFRecordIO.read().from(filenamePattern).withCompression(AUTO)
```

[link to github](https://github.com/apache/beam/blob/36257aba9054e664ebaafccfefb78bf54a162618/sdks/java/io/file-based-io-tests/src/test/java/org/apache/beam/sdk/io/tfrecord/TFRecordIOIT.java#L113) this is a blocker for running full set of filebased io tests on hdfs.

Steps to reproduce:
1. Create remote hadoop environment. This step asume you have local kubectl tool configured to use your GCP project.
```

pushd .test-infra/kubernetes/hadoop/SmallITCluster/ && /bin/bash ./setup-all.sh && popd

```

2. Update /etc/hosts file with the provided output from sctipt.
3. Confirm that it works and hadoop web interface is accessible on {color:#FF0000}http://hadoop-xxxxx:50070{color} where xxxxx is added in step2 sequence from your /etc/hosts entry. Please also substitute xxxxx in further usages of this.
4. Tell runner to use root as hadoop user.
```

export HADOOP_USER_NAME=root

```

5. Run TFRecord tests on this environment using DirectRunner:
```

mvn -e -Pio-it verify -pl sdks/java/io/file-based-io-tests/ -Dit.test=org.apache.beam.sdk.io.tfrecord.TFRecordIOIT
-Dfilesystem=hdfs -DintegrationTestPipelineOptions='["--filenamePrefix=hdfs://hadoop-xxxxx:9000/TFRecord",
"--hdfsConfiguration=[{\"fs.defaultFS\" : \"hdfs://hadoop-xxxxx:9000\", \"dfs.replication\": 1, \"dfs.client.use.datanode.hostname\":\"true\"}]"
]' -DforceDirectRunner=true

```

The error message is:
```

[ERROR] Tests run: 1, Failures: 0, Errors: 1, Skipped: 0, Time elapsed: 78.055 s <<< FAILURE! - in
org.apache.beam.sdk.io.tfrecord.TFRecordIOIT
[ERROR] writeThenReadAll(org.apache.beam.sdk.io.tfrecord.TFRecordIOIT)
Time elapsed: 78.055 s <<< ERROR!
java.lang.IllegalStateException: Invalid data
at org.apache.beam.sdk.repackaged.com.google.common.base.Preconditions.checkState(Preconditions.java:444)

at org.apache.beam.sdk.io.TFRecordIO$TFRecordCodec.read(TFRecordIO.java:642)
at org.apache.beam.sdk.io.TFRecordIO$TFRecordSource$TFRecordReader.readNextRecord(TFRecordIO.java:526)

at org.apache.beam.sdk.io.CompressedSource$CompressedReader.readNextRecord(CompressedSource.java:426)

at org.apache.beam.sdk.io.FileBasedSource$FileBasedReader.advanceImpl(FileBasedSource.java:473)

at org.apache.beam.sdk.io.OffsetBasedSource$OffsetBasedReader.advance(OffsetBasedSource.java:267)

at org.apache.beam.runners.direct.BoundedReadEvaluatorFactory$BoundedReadEvaluator.processElement(BoundedReadEvaluatorFactory.java:148)

at org.apache.beam.runners.direct.DirectTransformExecutor.processElements(DirectTransformExecutor.java:161)

at org.apache.beam.runners.direct.DirectTransformExecutor.run(DirectTransformExecutor.java:125)

at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
at java.util.concurrent.FutureTask.run(FutureTask.java:266)

at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
at
java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
at java.lang.Thread.run(Thread.java:745)

```

This results were also observed when running tests on jenkins. [Link to jenkins build](https://builds.apache.org/view/A-D/view/Beam/job/beam_PerformanceTests_TFRecordIOIT_HDFS/3/console)

When you open http://hadoop-xxxxx:50070/explorer.html#/ you will see TFRecord files that were created during write phase. Unable to be processed in reading phase.

{color:red}Important note{color}: if I copy files made by writing pipeline from hdfs directory to local directory and run reading pipeline over them, everything is working fine, so only reading from hdfs is a problem.

You can wipe out hdfs environment by runnning:
```

pushd .test-infra/kubernetes/hadoop/SmallITCluster/ && /bin/bash ./teardown-all.sh && popd

```

Imported from Jira [BEAM-3945](https://issues.apache.org/jira/browse/BEAM-3945). Original Jira may contain additional context.
Reported by: szewinho.