[fix](hdfs) Fix hdfsExists that return staled root cause (apache#27991)

The HDFS native client won't clear the last exception as expected so `hdfsGetLastExceptionRootCause` might return a staled root cause. This PR saves the last root cause here and verifies after hdfsExists returns a non-zero code.

be/src/io/fs/hdfs_file_system.cpp

@@ -227,15 +227,28 @@ Status HdfsFileSystem::delete_internal(const Path& path, int is_recursive) {
 Status HdfsFileSystem::exists_impl(const Path& path, bool* res) const {
  CHECK_HDFS_HANDLE(_fs_handle);
  Path real_path = convert_path(path, _fs_name);
+#ifdef USE_HADOOP_HDFS
+ // HACK: the HDFS native client won't clear the last exception as expected so
+ // `hdfsGetLastExceptionRootCause` might return a staled root cause. Save the
+ // last root cause here and verify after hdfsExists returns a non-zero code.
+ //
+ // See details:
+ // https://github.com/apache/hadoop/blob/5cda162a804fb0cfc2a5ac0058ab407662c5fb00/
+ // hadoop-hdfs-project/hadoop-hdfs-native-client/src/main/native/libhdfs/jni_helper.c#L795
+ char* former_root_cause = hdfsGetLastExceptionRootCause();
+#endif
  int is_exists = hdfsExists(_fs_handle->hdfs_fs, real_path.string().c_str());
 #ifdef USE_HADOOP_HDFS
  // when calling hdfsExists() and return non-zero code,
  // if root_cause is nullptr, which means the file does not exist.
  // if root_cause is not nullptr, which means it encounter other error, should return.
  // NOTE: not for libhdfs3 since it only runs on MaxOS, don't have to support it.
- char* root_cause = hdfsGetLastExceptionRootCause();
- if (root_cause != nullptr) {
- return Status::IOError("failed to check path existence {}: {}", path.native(), root_cause);
+ if (is_exists != 0) {
+ char* root_cause = hdfsGetLastExceptionRootCause();
+ if (root_cause != nullptr && root_cause != former_root_cause) {
+ return Status::IOError("failed to check path existence {}: {}", path.native(),
+ root_cause);
+ }
  }
 #endif
  *res = (is_exists == 0);