HDFS NFS Gateway

Overview

The NFS Gateway supports NFSv3 and allows HDFS to be mounted as part of the client's local file system. Currently NFS Gateway supports and enables the following usage patterns:

  • Users can browse the HDFS file system through their local file system on NFSv3 client compatible operating systems.
  • Users can download files from the the HDFS file system on to their local file system.
  • Users can upload files from their local file system directly to the HDFS file system.
  • Users can stream data directly to HDFS through the mount point. File append is supported but random write is not supported.

The NFS gateway machine needs the same thing to run an HDFS client like Hadoop JAR files, HADOOP_CONF directory. The NFS gateway can be on the same host as DataNode, NameNode, or any HDFS client.

Configuration

The user running the NFS-gateway must be able to proxy all the users using the NFS mounts. For instance, if user 'nfsserver' is running the gateway, and users belonging to the groups 'nfs-users1' and 'nfs-users2' use the NFS mounts, then in core-site.xml of the namenode, the following must be set (NOTE: replace 'nfsserver' with the user name starting the gateway in your cluster):

<property>
  <name>hadoop.proxyuser.nfsserver.groups</name>
  <value>nfs-users1,nfs-users2</value>
  <description>
         The 'nfsserver' user is allowed to proxy all members of the 'nfs-users1' and 
         'nfs-users2' groups. Set this to '*' to allow nfsserver user to proxy any group.
  </description>
</property>
<property>
  <name>hadoop.proxyuser.nfsserver.hosts</name>
  <value>nfs-client-host1.com</value>
  <description>
         This is the host where the nfs gateway is running. Set this to '*' to allow
         requests from any hosts to be proxied.
  </description>
</property>

The above are the only required configuration for the NFS gateway in non-secure mode. For Kerberized hadoop clusters, the following configurations need to be added to hdfs-site.xml:

  <property>
    <name>nfs.keytab.file</name>
    <value>/etc/hadoop/conf/nfsserver.keytab</value> <!-- path to the nfs gateway keytab -->
  </property>
  <property>
    <name>nfs.kerberos.principal</name>
    <value>nfsserver/_HOST@YOUR-REALM.COM</value>
  </property>

The AIX NFS client has a few known issues that prevent it from working correctly by default with the HDFS NFS Gateway. If you want to be able to access the HDFS NFS Gateway from AIX, you should set the following configuration setting to enable work-arounds for these issues:

<property>
  <name>nfs.aix.compatibility.mode.enabled</name>
  <value>true</value>
</property>

Note that regular, non-AIX clients should NOT enable AIX compatibility mode. The work-arounds implemented by AIX compatibility mode effectively disable safeguards to ensure that listing of directory contents via NFS returns consistent results, and that all data sent to the NFS server can be assured to have been committed.

It's strongly recommended for the users to update a few configuration properties based on their use cases. All the related configuration properties can be added or updated in hdfs-site.xml.

  • If the client mounts the export with access time update allowed, make sure the following property is not disabled in the configuration file. Only NameNode needs to restart after this property is changed. On some Unix systems, the user can disable access time update by mounting the export with "noatime". If the export is mounted with "noatime", the user doesn't need to change the following property and thus no need to restart namenode.
    <property>
      <name>dfs.namenode.accesstime.precision</name>
      <value>3600000</value>
      <description>The access time for HDFS file is precise upto this value.
        The default value is 1 hour. Setting a value of 0 disables
        access times for HDFS.
      </description>
    </property>
  • Users are expected to update the file dump directory. NFS client often reorders writes. Sequential writes can arrive at the NFS gateway at random order. This directory is used to temporarily save out-of-order writes before writing to HDFS. For each file, the out-of-order writes are dumped after they are accumulated to exceed certain threshold (e.g., 1MB) in memory. One needs to make sure the directory has enough space. For example, if the application uploads 10 files with each having 100MB, it is recommended for this directory to have roughly 1GB space in case if a worst-case write reorder happens to every file. Only NFS gateway needs to restart after this property is updated.
      <property>    
        <name>nfs.dump.dir</name>
        <value>/tmp/.hdfs-nfs</value>
      </property>
  • For optimal performance, it is recommended that rtmax be updated to 1MB. However, note that this 1MB is a per client allocation, and not from a shared memory pool, and therefore a larger value may adversely affect small reads, consuming a lot of memory. The maximum value of this property is 1MB.
    <property>
      <name>nfs.rtmax</name>
      <value>1048576</value>
      <description>This is the maximum size in bytes of a READ request
        supported by the NFS gateway. If you change this, make sure you
        also update the nfs mount's rsize(add rsize= # of bytes to the 
        mount directive).
      </description>
    </property>
    <property>
      <name>nfs.wtmax</name>
      <value>65536</value>
      <description>This is the maximum size in bytes of a WRITE request
        supported by the NFS gateway. If you change this, make sure you
        also update the nfs mount's wsize(add wsize= # of bytes to the 
        mount directive).
      </description>
    </property>
  • By default, the export can be mounted by any client. To better control the access, users can update the following property. The value string contains machine name and access privilege, separated by whitespace characters. The machine name format can be a single host, a Java regular expression, or an IPv4 address. The access privilege uses rw or ro to specify read/write or read-only access of the machines to exports. If the access privilege is not provided, the default is read-only. Entries are separated by ";". For example: "192.168.0.0/22 rw ; host.*\.example\.com ; host1.test.org ro;". Only the NFS gateway needs to restart after this property is updated.
    <property>
      <name>nfs.exports.allowed.hosts</name>
      <value>* rw</value>
    </property>
  • Customize log settings. To get NFS debug trace, users can edit the log4j.property file to add the following. Note, debug trace, especially for ONCRPC, can be very verbose.

    To change logging level:

        log4j.logger.org.apache.hadoop.hdfs.nfs=DEBUG

    To get more details of ONCRPC requests:

        log4j.logger.org.apache.hadoop.oncrpc=DEBUG

Start and stop NFS gateway service

Three daemons are required to provide NFS service: rpcbind (or portmap), mountd and nfsd. The NFS gateway process has both nfsd and mountd. It shares the HDFS root "/" as the only export. It is recommended to use the portmap included in NFS gateway package. Even though NFS gateway works with portmap/rpcbind provide by most Linux distributions, the package included portmap is needed on some Linux systems such as REHL6.2 due to an rpcbind bug. More detailed discussions can be found in HDFS-4763.

  1. Stop nfs/rpcbind/portmap services provided by the platform (commands can be different on various Unix platforms):
         service nfs stop
          
         service rpcbind stop
  2. Start package included portmap (needs root privileges):
         hadoop portmap
      
         OR
    
         hadoop-daemon.sh start portmap
  3. Start mountd and nfsd.

    No root privileges are required for this command. However, ensure that the user starting the Hadoop cluster and the user starting the NFS gateway are same.

         hadoop nfs3
    
         OR
    
         hadoop-daemon.sh start nfs3

    Note, if the hadoop-daemon.sh script starts the NFS gateway, its log can be found in the hadoop log folder.

  4. Stop NFS gateway services.
          hadoop-daemon.sh stop nfs3
    
          hadoop-daemon.sh stop portmap

Optionally, you can forgo running the Hadoop-provided portmap daemon and instead use the system portmap daemon on all operating systems if you start the NFS Gateway as root. This will allow the HDFS NFS Gateway to work around the aforementioned bug and still register using the system portmap daemon. To do so, just start the NFS gateway daemon as you normally would, but make sure to do so as the "root" user, and also set the "HADOOP_PRIVILEGED_NFS_USER" environment variable to an unprivileged user. In this mode the NFS Gateway will start as root to perform its initial registration with the system portmap, and then will drop privileges back to the user specified by the HADOOP_PRIVILEGED_NFS_USER afterward and for the rest of the duration of the lifetime of the NFS Gateway process. Note that if you choose this route, you should skip steps 1 and 2 above.

Verify validity of NFS related services

  1. Execute the following command to verify if all the services are up and running:
           rpcinfo -p $nfs_server_ip

    You should see output similar to the following:

           program vers proto   port
    
           100005    1   tcp   4242  mountd
    
           100005    2   udp   4242  mountd
    
           100005    2   tcp   4242  mountd
    
           100000    2   tcp    111  portmapper
    
           100000    2   udp    111  portmapper
    
           100005    3   udp   4242  mountd
    
           100005    1   udp   4242  mountd
    
           100003    3   tcp   2049  nfs
    
           100005    3   tcp   4242  mountd
  2. Verify if the HDFS namespace is exported and can be mounted.
            showmount -e $nfs_server_ip                         

    You should see output similar to the following:

            Exports list on $nfs_server_ip :
    
            / (everyone)

Mount the export “/”

Currently NFS v3 only uses TCP as the transportation protocol. NLM is not supported so mount option "nolock" is needed. It's recommended to use hard mount. This is because, even after the client sends all data to NFS gateway, it may take NFS gateway some extra time to transfer data to HDFS when writes were reorderd by NFS client Kernel.

If soft mount has to be used, the user should give it a relatively long timeout (at least no less than the default timeout on the host) .

The users can mount the HDFS namespace as shown below:

       mount -t nfs -o vers=3,proto=tcp,nolock,noacl $server:/  $mount_point

Then the users can access HDFS as part of the local file system except that, hard link and random write are not supported yet.

Allow mounts from unprivileged clients

In environments where root access on client machines is not generally available, some measure of security can be obtained by ensuring that only NFS clients originating from privileged ports can connect to the NFS server. This feature is referred to as "port monitoring." This feature is not enabled by default in the HDFS NFS Gateway, but can be optionally enabled by setting the following config in hdfs-site.xml on the NFS Gateway machine:

<property>
  <name>nfs.port.monitoring.disabled</name>
  <value>false</value>
</property>

User authentication and mapping

NFS gateway in this release uses AUTH_UNIX style authentication. When the user on NFS client accesses the mount point, NFS client passes the UID to NFS gateway. NFS gateway does a lookup to find user name from the UID, and then passes the username to the HDFS along with the HDFS requests. For example, if the NFS client has current user as "admin", when the user accesses the mounted directory, NFS gateway will access HDFS as user "admin". To access HDFS as the user "hdfs", one needs to switch the current user to "hdfs" on the client system when accessing the mounted directory.

The system administrator must ensure that the user on NFS client host has the same name and UID as that on the NFS gateway host. This is usually not a problem if the same user management system (e.g., LDAP/NIS) is used to create and deploy users on HDFS nodes and NFS client node. In case the user account is created manually on different hosts, one might need to modify UID (e.g., do "usermod -u 123 myusername") on either NFS client or NFS gateway host in order to make it the same on both sides. More technical details of RPC AUTH_UNIX can be found in RPC specification.

Optionally, the system administrator can configure a custom static mapping file in the event one wishes to access the HDFS NFS Gateway from a system with a completely disparate set of UIDs/GIDs. By default this file is located at "/etc/nfs.map", but a custom location can be configured by setting the "nfs.static.mapping.file" property to the path of the static mapping file. The format of the static mapping file is similar to what is described in the exports(5) manual page, but roughly it is:

# Mapping for clients accessing the NFS gateway
uid 10 100 # Map the remote UID 10 the local UID 100
gid 11 101 # Map the remote GID 11 to the local GID 101