Filesystem Access Utilities

Overview

openLooKeng project includes a set of filesystem client utilities to help access and modifying files. Currently, two sets of filesystems are supported: HDFS and local filesystem. A HetuFileSystemClient interface is provided in the SPI, which defines the common file operations to be used in the project. The goal of this client is to provide unified interface, behaviors and exceptions across different filesystems. Therefore client code can easily reuse codes and transfer their logic without having to change the code.

A utility class FileBasedLock located in SPI can provide exclusive access on a given filesystem. It takes advantage of the unified filesystem client interface thus works for different filesystems.

The filesystem clients are implemented as openLooKeng Plugins. The implementations are placed in the hetu-filesystem-client module, where one factory extending HetuFileSystemClientFactory must be implemented and registered in HetuFileSystemClientPlugin.

Filesystem Profiles

A filesystem profile must contain a type field:

fs.client.type=<filesystem type>

Additional configs used by the corresponding filesystem client can be provided. For example, hdfs filesystem needs paths to config resource files core-site.xml and hdfs-site.xml. If the filesystem enables authentication, such as KERBEROS, credentials should be specified in the profile as well.

Multiple profiles defining filesystem clients can be placed in etc/filesystem. Similar to catalogs, they will be loaded upon server start by a filesystem manager which also produce filesystem clients to users in presto-main. Client modules can read profiles in this folder by themselves, but it is highly recommended that client modules obtain a client which is provided by the main module in order to prevent dependency issues.

A typical use case of having multiple profiles is when multiple hdfs are used and preset. In this case create a property file for each cluster and include their own authentication information in different profiles, for example hdfs1.properties and hdfs2.properties. Client code will be able to access them by specifying the profile name (file name), such as getFileSystemClient("hdfs1", <rootPath>) from the FileSystemClientManager.

For each local and hdfs filesystem client, it requires a root path to which the access to filesystem is limited. All access beyond this root dir will be denied. It is suggested to use the deepest directory that meets the need as the client root, in order to avoid the risk of modifying (or deleting) additional files by accident.

Filesystem profile properties

Property Name Mandatory Description Default Value
fs.client.type YES The type of filesystem profile. Accepted values: local, hdfs.
hdfs.config.resources NO Path to hdfs resource files (e.g. core-site.xml, hdfs-site.xml) Use Local hdfs
hdfs.authentication.type YES hdfs authentication Accepted values: KERBEROS, NONE
hdfs.krb5.conf.path YES if auth type set to KERBEROS Path to the krb5 config file
hdfs.krb5.keytab.path YES if auth type set to KERBEROS Path to the kerberos keytab file
hdfs.krb5.principal YES if auth type set to KERBEROS Principal of kerberos authentication
fs.hdfs.impl.disable.cache NO Disable cache in hdfs. false

openLooKengFileSystemClient

The unified HetuFileSystemClient sets the standard of filesystem access for:

  • Method signatures
  • Behaviors
  • Exceptions

The ultimate goal of doing all these unification is to increase the reusability of code that needs to perform file operations on multiple filesystem.

Method signatures and behaviors

The methods follow the same signature as those in java.nio.file.spi.FileSystemProvider and java.nio.file.Files in order to maximize compatibility. There are also additional methods that are not available in java.nio package to supply useful functionalities (e.g. deleteRecursively()).

Exceptions

Similar to method signatures, exception patterns strictly follow the same as those produced in Files. For other filesystems, the implementation translates exceptions to Java local java.nio.file.FileSystemException which is most semantically equivalent. This allows the client code to:

  1. handle IOException without worrying about the underlying filesystem.
  2. decouple from extra dependencies, such as hadoop

Here are examples of exceptions that are translated from hdfs:

  • org.apache.hadoop.fs.FileAlreadyExistsException -> java.nio.file.FileAlreadyExistsException
  • org.apache.hadoop.fs.PathIsNotEmptyDirectoryException -> java.nio.file.DirectoryNotEmptyException

FileBasedLock

A FileBasedLock utility is provided in io.prestosql.spi.filesystem to be used together with HetuFileSystemClient. This lock is file-based, and follows “try-and-back-off-on-exception” pattern to serialize operations in concurrent scenarios.

The lock design has a two-step file based lock schema: .lockFile to mark lock status and .lockInfo to record lock ownership:

  1. Before a client obtains the lock, it first checks if a valid (non-expire) .lockFile exists.

  2. If not it tries to access the information stored in .lockInfo to see if the lock is held by itself, or try to write its uuid into the .lockInfo file if a valid file does not exist.

  3. After all these it starts a background thread to keep updating .lockFile to inform others that the filesystem has been locked and renew it.

  4. If any of the check above indicates that the lock has been acquired by others, or any exception occurs during the process, the current process/thread gives up (back-off) locking and tries again in the next loop.