gitlab-org--gitlab-foss/doc/administration/high_availability/nfs.md

15 KiB

type
reference

NFS

You can view information and options set for each of the mounted NFS file systems by running nfsstat -m and cat /etc/fstab.

CAUTION: Caution: From GitLab 13.0, using NFS for Git repositories is deprecated. In GitLab 14.0, support for NFS for Git repositories is scheduled to be removed. Upgrade to Gitaly Cluster as soon as possible.

NOTE: Note: Filesystem performance has a big impact on overall GitLab performance, especially for actions that read or write to Git repositories. See Filesystem Performance Benchmarking for steps to test filesystem performance.

Known kernel version incompatibilities

RedHat Enterprise Linux (RHEL) and CentOS v7.7 and v7.8 ship with kernel version 3.10.0-1127, which contains a bug that causes uploads to fail to copy over NFS. The following GitLab versions include a fix to work properly with that kernel version:

  1. 12.10.12
  2. 13.0.7
  3. 13.1.1
  4. 13.2 and up

If you are using that kernel version, be sure to upgrade GitLab to avoid errors.

NFS Server features

Required features

File locking: GitLab requires advisory file locking, which is only supported natively in NFS version 4. NFSv3 also supports locking as long as Linux Kernel 2.6.5+ is used. We recommend using version 4 and do not specifically test NFSv3.

When you define your NFS exports, we recommend you also add the following options:

  • no_root_squash - NFS normally changes the root user to nobody. This is a good security measure when NFS shares will be accessed by many different users. However, in this case only GitLab will use the NFS share so it is safe. GitLab recommends the no_root_squash setting because we need to manage file permissions automatically. Without the setting you may receive errors when the Omnibus package tries to alter permissions. Note that GitLab and other bundled components do not run as root but as non-privileged users. The recommendation for no_root_squash is to allow the Omnibus package to set ownership and permissions on files, as needed. In some cases where the no_root_squash option is not available, the root flag can achieve the same result.
  • sync - Force synchronous behavior. Default is asynchronous and under certain circumstances it could lead to data loss if a failure occurs before data has synced.

Due to the complexities of running Omnibus with LDAP and the complexities of maintaining ID mapping without LDAP, in most cases you should enable numeric UIDs and GIDs (which is off by default in some cases) for simplified permission management between systems:

Disable NFS server delegation

We recommend that all NFS users disable the NFS server delegation feature. This is to avoid a Linux kernel bug which causes NFS clients to slow precipitously due to excessive network traffic from numerous TEST_STATEID NFS messages.

To disable NFS server delegation, do the following:

  1. On the NFS server, run:

    echo 0 > /proc/sys/fs/leases-enable
    sysctl -w fs.leases-enable=0
    
  2. Restart the NFS server process. For example, on CentOS run service nfs restart.

Important notes

The kernel bug may be fixed in more recent kernels with this commit.

Red Hat Enterprise 7 shipped a kernel update on August 6, 2019 that may also have resolved this problem.

You may not need to disable NFS server delegation if you know you are using a version of the Linux kernel that has been fixed. That said, GitLab still encourages instance administrators to keep NFS server delegation disabled.

Improving NFS performance with GitLab

Improving NFS performance with Unicorn

NOTE: Note: From GitLab 12.1, it will automatically be detected if Rugged can and should be used per storage.

If you previously enabled Rugged using the feature flag, you will need to unset the feature flag by using:

sudo gitlab-rake gitlab:features:unset_rugged

If the Rugged feature flag is explicitly set to either true or false, GitLab will use the value explicitly set.

Improving NFS performance with Puma

NOTE: Note: From GitLab 12.7, Rugged auto-detection is disabled if Puma thread count is greater than 1.

If you want to use Rugged with Puma, it is recommended to set Puma thread count to 1.

If you want to use Rugged with Puma thread count more than 1, Rugged can be enabled using the feature flag

If the Rugged feature flag is explicitly set to either true or false, GitLab will use the value explicitly set.

Known issues

Avoid using AWS's Elastic File System (EFS)

GitLab strongly recommends against using AWS Elastic File System (EFS). Our support team will not be able to assist on performance issues related to file system access.

Customers and users have reported that AWS EFS does not perform well for GitLab's use-case. Workloads where many small files are written in a serialized manner, like git, are not well-suited for EFS. EBS with an NFS server on top will perform much better.

If you do choose to use EFS, avoid storing GitLab log files (e.g. those in /var/log/gitlab) there because this will also affect performance. We recommend that the log files be stored on a local volume.

For more details on another person's experience with EFS, see this Commit Brooklyn 2019 video.

Avoid using CephFS and GlusterFS

GitLab strongly recommends against using CephFS and GlusterFS. These distributed file systems are not well-suited for GitLab's input/output access patterns because Git uses many small files and access times and file locking times to propagate will make Git activity very slow.

Avoid using PostgreSQL with NFS

GitLab strongly recommends against running your PostgreSQL database across NFS. The GitLab support team will not be able to assist on performance issues related to this configuration.

Additionally, this configuration is specifically warned against in the PostgreSQL Documentation:

PostgreSQL does nothing special for NFS file systems, meaning it assumes NFS behaves exactly like locally-connected drives. If the client or server NFS implementation does not provide standard file system semantics, this can cause reliability problems. Specifically, delayed (asynchronous) writes to the NFS server can cause data corruption problems.

For supported database architecture, please see our documentation on Configuring a Database for GitLab HA.

NFS Client mount options

Here is an example snippet to add to /etc/fstab:

10.1.0.1:/var/opt/gitlab/.ssh /var/opt/gitlab/.ssh nfs4 defaults,vers=4.1,hard,rsize=1048576,wsize=1048576,noatime,nofail,lookupcache=positive 0 2
10.1.0.1:/var/opt/gitlab/gitlab-rails/uploads /var/opt/gitlab/gitlab-rails/uploads nfs4 defaults,vers=4.1,hard,rsize=1048576,wsize=1048576,noatime,nofail,lookupcache=positive 0 2
10.1.0.1:/var/opt/gitlab/gitlab-rails/shared /var/opt/gitlab/gitlab-rails/shared nfs4 defaults,vers=4.1,hard,rsize=1048576,wsize=1048576,noatime,nofail,lookupcache=positive 0 2
10.1.0.1:/var/opt/gitlab/gitlab-ci/builds /var/opt/gitlab/gitlab-ci/builds nfs4 defaults,vers=4.1,hard,rsize=1048576,wsize=1048576,noatime,nofail,lookupcache=positive 0 2
10.1.0.1:/var/opt/gitlab/git-data /var/opt/gitlab/git-data nfs4 defaults,vers=4.1,hard,rsize=1048576,wsize=1048576,noatime,nofail,lookupcache=positive 0 2

Note there are several options that you should consider using:

Setting Description
vers=4.1 NFS v4.1 should be used instead of v4.0 because there is a Linux NFS client bug in v4.0 that can cause significant problems due to stale data.
nofail Don't halt boot process waiting for this mount to become available
lookupcache=positive Tells the NFS client to honor positive cache results but invalidates any negative cache results. Negative cache results cause problems with Git. Specifically, a git push can fail to register uniformly across all NFS clients. The negative cache causes the clients to 'remember' that the files did not exist previously.
hard Instead of soft. Further details.

soft mount option

We recommend that you use hard in your mount options, unless you have a specific reason to use soft.

On GitLab.com, we use soft because there were times when we had NFS servers reboot and soft improved availability, but everyone's infrastructure is different. If your NFS is provided by on-premise storage arrays with redundant controllers, for example, you shouldn't need to worry about NFS server availability.

The NFS man page states:

"soft" timeout can cause silent data corruption in certain cases

Read the Linux man page to understand the difference, and if you do use soft, ensure that you've taken steps to mitigate the risks.

If you experience behavior that might have been caused by writes to disk on the NFS server not occurring, such as commits going missing, use the hard option, because (from the man page):

use the soft option only when client responsiveness is more important than data integrity

Other vendors make similar recommendations, including SAP and NetApp's knowledge base, they highlight that if the NFS client driver caches data, soft means there is no certainty if writes by GitLab are actually on disk.

Mount points set with the option hard may not perform as well, and if the NFS server goes down, hard will cause processes to hang when interacting with the mount point. Use SIGKILL (kill -9) to deal with hung processes. The intr option stopped working in the 2.6 kernel.

A single NFS mount

It's recommended to nest all GitLab data directories within a mount, that allows automatic restore of backups without manually moving existing data.

mountpoint
└── gitlab-data
    ├── builds
    ├── git-data
    ├── shared
    └── uploads

To do so, we'll need to configure Omnibus with the paths to each directory nested in the mount point as follows:

Mount /gitlab-nfs then use the following Omnibus configuration to move each data location to a subdirectory:

git_data_dirs({"default" => { "path" => "/gitlab-nfs/gitlab-data/git-data"} })
gitlab_rails['uploads_directory'] = '/gitlab-nfs/gitlab-data/uploads'
gitlab_rails['shared_path'] = '/gitlab-nfs/gitlab-data/shared'
gitlab_ci['builds_directory'] = '/gitlab-nfs/gitlab-data/builds'

Run sudo gitlab-ctl reconfigure to start using the central location. Please be aware that if you had existing data you will need to manually copy/rsync it to these new locations and then restart GitLab.

Bind mounts

Alternatively to changing the configuration in Omnibus, bind mounts can be used to store the data on an NFS mount.

Bind mounts provide a way to specify just one NFS mount and then bind the default GitLab data locations to the NFS mount. Start by defining your single NFS mount point as you normally would in /etc/fstab. Let's assume your NFS mount point is /gitlab-nfs. Then, add the following bind mounts in /etc/fstab:

/gitlab-nfs/gitlab-data/git-data /var/opt/gitlab/git-data none bind 0 0
/gitlab-nfs/gitlab-data/.ssh /var/opt/gitlab/.ssh none bind 0 0
/gitlab-nfs/gitlab-data/uploads /var/opt/gitlab/gitlab-rails/uploads none bind 0 0
/gitlab-nfs/gitlab-data/shared /var/opt/gitlab/gitlab-rails/shared none bind 0 0
/gitlab-nfs/gitlab-data/builds /var/opt/gitlab/gitlab-ci/builds none bind 0 0

Using bind mounts will require manually making sure the data directories are empty before attempting a restore. Read more about the restore prerequisites.

Multiple NFS mounts

When using default Omnibus configuration you will need to share 4 data locations between all GitLab cluster nodes. No other locations should be shared. The following are the 4 locations need to be shared:

Location Description Default configuration
/var/opt/gitlab/git-data Git repository data. This will account for a large portion of your data git_data_dirs({"default" => { "path" => "/var/opt/gitlab/git-data"} })
/var/opt/gitlab/gitlab-rails/uploads User uploaded attachments gitlab_rails['uploads_directory'] = '/var/opt/gitlab/gitlab-rails/uploads'
/var/opt/gitlab/gitlab-rails/shared Build artifacts, GitLab Pages, LFS objects, temp files, etc. If you're using LFS this may also account for a large portion of your data gitlab_rails['shared_path'] = '/var/opt/gitlab/gitlab-rails/shared'
/var/opt/gitlab/gitlab-ci/builds GitLab CI/CD build traces gitlab_ci['builds_directory'] = '/var/opt/gitlab/gitlab-ci/builds'

Other GitLab directories should not be shared between nodes. They contain node-specific files and GitLab code that does not need to be shared. To ship logs to a central location consider using remote syslog. Omnibus GitLab packages provide configuration for UDP log shipping.

Having multiple NFS mounts will require manually making sure the data directories are empty before attempting a restore. Read more about the restore prerequisites.


Read more on high-availability configuration:

  1. Configure the database
  2. Configure Redis
  3. Configure the GitLab application servers
  4. Configure the load balancers