Troubleshooting Geo replication

DETAILS: Tier: Premium, Ultimate Offering: Self-managed

Fixing PostgreSQL database replication errors

The following sections outline troubleshooting steps for fixing replication error messages (indicated by Database replication working? ... no in the geo:check output. The instructions present here mostly assume a single-node Geo Linux package deployment, and might need to be adapted to different environments.

Removing an inactive replication slot

Replication slots are marked as 'inactive' when the replication client (a secondary site) connected to the slot disconnects. Inactive replication slots cause WAL files to be retained, because they are sent to the client when it reconnects and the slot becomes active once more. If the secondary site is not able to reconnect, use the following steps to remove its corresponding inactive replication slot:

1. Start a PostgreSQL console session on the Geo primary site's database node:

   sudo gitlab-psql -d gitlabhq_production

   NOTE: Using gitlab-rails dbconsole does not work, because managing replication slots requires superuser permissions.

2. View the replication slots and remove them if they are inactive:

   SELECT * FROM pg_replication_slots;

   Slots where active is f are inactive.

   • When this slot should be active, because you have a secondary site configured using that slot, look for the PostgreSQL logs for the secondary site, to view why the replication is not running.

   • If you are no longer using the slot (for example, you no longer have Geo enabled), or the secondary site is no longer able to reconnect, you should remove it using the PostgreSQL console session:

     SELECT pg_drop_replication_slot('<name_of_inactive_slot>');

3. Follow either the steps to remove that Geo site if it's no longer required, or re-initiate the replication process, which recreates the replication slot correctly.

Message: WARNING: oldest xmin is far in the past and pg_wal size growing

If a replication slot is inactive, the pg_wal logs corresponding to the slot are reserved forever (or until the slot is active again). This causes continuous disk usage growth and the following messages appear repeatedly in the PostgreSQL logs:

WARNING: oldest xmin is far in the past
HINT: Close open transactions soon to avoid wraparound problems.
You might also need to commit or roll back old prepared transactions, or drop stale replication slots.

To fix this, you should remove the inactive replication slot and re-initiate the replication.

Message: ERROR: replication slots can only be used if max_replication_slots > 0?

This means that the max_replication_slots PostgreSQL variable needs to be set on the primary database. This setting defaults to 1. You may need to increase this value if you have more secondary sites.

Be sure to restart PostgreSQL for this to take effect. See the PostgreSQL replication setup guide for more details.

Message: FATAL: could not start WAL streaming: ERROR: replication slot "geo_secondary_my_domain_com" does not exist?

This occurs when PostgreSQL does not have a replication slot for the secondary site by that name.

You may want to rerun the replication process on the secondary site .

Message: "Command exceeded allowed execution time" when setting up replication?

This may happen while initiating the replication process on the secondary site, and indicates your initial dataset is too large to be replicated in the default timeout (30 minutes).

Re-run gitlab-ctl replicate-geo-database, but include a larger value for --backup-timeout:

sudo gitlab-ctl \
   replicate-geo-database \
   --host=<primary_node_hostname> \
   --slot-name=<secondary_slot_name> \
   --backup-timeout=21600

This gives the initial replication up to six hours to complete, rather than the default 30 minutes. Adjust as required for your installation.

Message: "PANIC: could not write to file pg_xlog/xlogtemp.123: No space left on device"

Determine if you have any unused replication slots in the primary database. This can cause large amounts of log data to build up in pg_xlog.

Removing the inactive slots can reduce the amount of space used in the pg_xlog.

Message: "ERROR: canceling statement due to conflict with recovery"

This error message occurs infrequently under typical usage, and the system is resilient enough to recover.

However, under certain conditions, some database queries on secondaries may run excessively long, which increases the frequency of this error message. This can lead to a situation where some queries never complete due to being canceled on every replication.

These long-running queries are planned to be removed in the future, but as a workaround, we recommend enabling hot_standby_feedback. This increases the likelihood of bloat on the primary site as it prevents VACUUM from removing recently-dead rows. However, it has been used successfully in production on GitLab.com.

To enable hot_standby_feedback, add the following to /etc/gitlab/gitlab.rb on the secondary site:

postgresql['hot_standby_feedback'] = 'on'

Then reconfigure GitLab:

sudo gitlab-ctl reconfigure

To help us resolve this problem, consider commenting on the issue.

Message: FATAL: could not connect to the primary server: server certificate for "PostgreSQL" does not match host name

This happens because the PostgreSQL certificate that the Linux package automatically creates contains the Common Name PostgreSQL, but the replication is connecting to a different host and GitLab attempts to use the verify-full SSL mode by default.

To fix this issue, you can either:

• Use the --sslmode=verify-ca argument with the replicate-geo-database command.
• For an already replicated database, change sslmode=verify-full to sslmode=verify-ca in /var/opt/gitlab/postgresql/data/gitlab-geo.conf and run gitlab-ctl restart postgresql.
• Configure SSL for PostgreSQL with a custom certificate (including the host name that's used to connect to the database in the CN or SAN) instead of using the automatically generated certificate.

Message: LOG: invalid CIDR mask in address

This happens on wrongly-formatted addresses in postgresql['md5_auth_cidr_addresses'].

2020-03-20_23:59:57.60499 LOG:  invalid CIDR mask in address "***"
2020-03-20_23:59:57.60501 CONTEXT:  line 74 of configuration file "/var/opt/gitlab/postgresql/data/pg_hba.conf"

To fix this, update the IP addresses in /etc/gitlab/gitlab.rb under postgresql['md5_auth_cidr_addresses'] to respect the CIDR format (for example, 10.0.0.1/32).

Message: LOG: invalid IP mask "md5": Name or service not known

This happens when you have added IP addresses without a subnet mask in postgresql['md5_auth_cidr_addresses'].

2020-03-21_00:23:01.97353 LOG:  invalid IP mask "md5": Name or service not known
2020-03-21_00:23:01.97354 CONTEXT:  line 75 of configuration file "/var/opt/gitlab/postgresql/data/pg_hba.conf"

To fix this, add the subnet mask in /etc/gitlab/gitlab.rb under postgresql['md5_auth_cidr_addresses'] to respect the CIDR format (for example, 10.0.0.1/32).

Message: Found data in the gitlabhq_production database! when running gitlab-ctl replicate-geo-database

This happens if data is detected in the projects table. When one or more projects are detected, the operation is aborted to prevent accidental data loss. To bypass this message, pass the --force option to the command.

In GitLab 13.4, a seed project is added when GitLab is first installed. This makes it necessary to pass --force even on a new Geo secondary site. There is an issue to account for seed projects when checking the database.

Message: FATAL: could not map anonymous shared memory: Cannot allocate memory

If you see this message, it means that the secondary site's PostgreSQL tries to request memory that is higher than the available memory. There is an issue that tracks this problem.

Example error message in Patroni logs (located at /var/log/gitlab/patroni/current for Linux package installations):

2023-11-21_23:55:18.63727 FATAL:  could not map anonymous shared memory: Cannot allocate memory
2023-11-21_23:55:18.63729 HINT:  This error usually means that PostgreSQL's request for a shared memory segment exceeded available memory, swap space, or huge pages. To reduce the request size (currently 17035526144 bytes), reduce PostgreSQL's shared memory usage, perhaps by reducing shared_buffers or max_connections.

The workaround is to increase the memory available to the secondary site's PostgreSQL nodes to match the memory requirements of the primary site's PostgreSQL nodes.

Fixing non-PostgreSQL replication failures

If you notice replication failures in Admin > Geo > Sites or the Sync status Rake task, you can try to resolve the failures with the following general steps:

1. Geo automatically retries failures. If the failures are new and few in number, or if you suspect the root cause is already resolved, then you can wait to see if the failures go away.
2. If failures were present for a long time, then many retries have already occurred, and the interval between automatic retries has increased to up to 4 hours depending on the type of failure. If you suspect the root cause is already resolved, you can manually retry replication or verification.
3. If the failures persist, use the following sections to try to resolve them.

Manually retry replication or verification

A Geo data type is a specific class of data that is required by one or more GitLab features to store relevant information and is replicated by Geo to secondary sites.

The following Geo data types exist:

• Blob types:
  • Ci::JobArtifact
  • Ci::PipelineArtifact
  • Ci::SecureFile
  • LfsObject
  • MergeRequestDiff
  • Packages::PackageFile
  • PagesDeployment
  • Terraform::StateVersion
  • Upload
  • DependencyProxy::Manifest
  • DependencyProxy::Blob
• Repository types:
  • ContainerRepositoryRegistry
  • DesignManagement::Repository
  • ProjectRepository
  • ProjectWikiRepository
  • SnippetRepository
  • GroupWikiRepository

The main kinds of classes are Registry, Model, and Replicator. If you have an instance of one of these classes, you can get the others. The Registry and Model mostly manage PostgreSQL DB state. The Replicator knows how to replicate/verify (or it can call a service to do it):

model_record = Packages::PackageFile.last
model_record.replicator.registry.replicator.model_record # just showing that these methods exist

With all this information, you can:

• Manually resync and reverify individual components
• Manually resync and reverify multiple components

Resync and reverify individual components

You can force a resync and reverify individual items for all component types managed by the self-service framework using the UI. On the secondary site, visit Admin > Geo > Replication.

However, if this doesn't work, you can perform the same action using the Rails console. The following sections describe how to use internal application commands in the Rails console to cause replication or verification for individual records synchronously or asynchronously.

WARNING: Commands that change data can cause damage if not run correctly or under the right conditions. Always run commands in a test environment first and have a backup instance ready to restore.

Start a Rails console session to enact the following, basic troubleshooting steps:

• For Blob types (using the Packages::PackageFile component as an example)

  • Find registry records that failed to sync:

    Geo::PackageFileRegistry.failed

  • Find registry records that are missing on the primary site:

    Geo::PackageFileRegistry.where(last_sync_failure: 'The file is missing on the Geo primary site')

  • Resync a package file, synchronously, given an ID:

    model_record = Packages::PackageFile.find(id)
    model_record.replicator.send(:download)

  • Resync a package file, synchronously, given a registry ID:

    registry = Geo::PackageFileRegistry.find(registry_id)
    registry.replicator.send(:download)

  • Resync a package file, asynchronously, given a registry ID. Since GitLab 16.2, a component can be asynchronously replicated as follows:

    registry = Geo::PackageFileRegistry.find(registry_id)
    registry.replicator.enqueue_sync

  • Reverify a package file, asynchronously, given a registry ID. Since GitLab 16.2, a component can be asynchronously reverified as follows:

    registry = Geo::PackageFileRegistry.find(registry_id)
    registry.replicator.verify_async

• For Repository types (using the SnippetRepository component as an example)

  • Resync a snippet repository, synchronously, given an ID:

    model_record = Geo::SnippetRepositoryRegistry.find(id)
    model_record.replicator.sync_repository

  • Resync a snippet repository, synchronously, given a registry ID

    registry = Geo::SnippetRepositoryRegistry.find(registry_id)
    registry.replicator.sync_repository

  • Resync a snippet repository, asynchronously, given a registry ID. Since GitLab 16.2, a component can be asynchronously replicated as follows:

    registry = Geo::SnippetRepositoryRegistry.find(registry_id)
    registry.replicator.enqueue_sync

  • Reverify a snippet repository, asynchronously, given a registry ID. Since GitLab 16.2, a component can be asynchronously reverified as follows:

    registry = Geo::SnippetRepositoryRegistry.find(registry_id)
    registry.replicator.verify_async

Resync and reverify multiple components

NOTE: There is an issue to implement this functionality in the Admin Area UI.

WARNING: Commands that change data can cause damage if not run correctly or under the right conditions. Always run commands in a test environment first and have a backup instance ready to restore.

The following sections describe how to use internal application commands in the Rails console to cause bulk replication or verification.

Reverify all components (or any SSF data type which supports verification)

For GitLab 16.4 and earlier:

1. SSH into a GitLab Rails node in the primary Geo site.

2. Open the Rails console.

3. Mark all uploads as pending verification:

   Upload.verification_state_table_class.each_batch do |relation|
     relation.update_all(verification_state: 0)
   end

4. This causes the primary to start checksumming all Uploads.

5. When a primary successfully checksums a record, then all secondaries recalculate the checksum as well, and they compare the values.

For other SSF data types replace Upload in the command above with the desired model class.

Verify blob files on the secondary manually

This iterates over all package files on the secondary, looking at the verification_checksum stored in the database (which came from the primary) and then calculate this value on the secondary to check if they match. This does not change anything in the UI.

For GitLab 14.4 and later:

# Run on secondary
status = {}

Packages::PackageFile.find_each do |package_file|
  primary_checksum = package_file.verification_checksum
  secondary_checksum = Packages::PackageFile.sha256_hexdigest(package_file.file.path)
  verification_status = (primary_checksum == secondary_checksum)

  status[verification_status.to_s] ||= []
  status[verification_status.to_s] << package_file.id
end

# Count how many of each value we get
status.keys.each {|key| puts "#{key} count: #{status[key].count}"}

# See the output in its entirety
status

For GitLab 14.3 and earlier:

# Run on secondary
status = {}

Packages::PackageFile.find_each do |package_file|
  primary_checksum = package_file.verification_checksum
  secondary_checksum = Packages::PackageFile.hexdigest(package_file.file.path)
  verification_status = (primary_checksum == secondary_checksum)

  status[verification_status.to_s] ||= []
  status[verification_status.to_s] << package_file.id
end

# Count how many of each value we get
status.keys.each {|key| puts "#{key} count: #{status[key].count}"}

# See the output in its entirety
status

Failed verification of Uploads on the primary Geo site

If verification of some uploads is failing on the primary Geo site with verification_checksum = nil and with the verification_failure = Error during verification: undefined method `underscore' for NilClass:Class, this can be due to orphaned Uploads. The parent record owning the Upload (the upload's model) has somehow been deleted, but the Upload record still exists. These verification failures are false.

You can find these errors in the geo.log file on the primary Geo site.

To confirm that model records are missing, you can run a Rake task on the primary Geo site:

sudo gitlab-rake gitlab:uploads:check

You can delete these Upload records on the primary Geo site to get rid of these failures by running the following script from the Rails console:

# Look for uploads with the verification error
# or edit with your own affected IDs
uploads = Geo::UploadState.where(
  verification_checksum: nil,
  verification_state: 3,
  verification_failure: "Error during verification: undefined method  `underscore' for NilClass:Class"
).pluck(:upload_id)

uploads_deleted = 0
begin
    uploads.each do |upload|
    u = Upload.find upload
    rescue => e
        puts "checking upload #{u.id} failed with #{e.message}"
      else
        uploads_deleted=uploads_deleted + 1
        p u                            ### allow verification before destroy
        # p u.destroy!                 ### uncomment to actually destroy
  end
end
p "#{uploads_deleted} remote objects were destroyed."

Investigate causes of database replication lag

If the output of sudo gitlab-rake geo:status shows that Database replication lag remains significantly high over time, the primary node in database replication can be checked to determine the status of lag for different parts of the database replication process. These values are known as write_lag, flush_lag, and replay_lag. For more information, see the official PostgreSQL documentation.

Run the following command from the primary Geo node's database to provide relevant output:

gitlab-psql -xc 'SELECT write_lag,flush_lag,replay_lag FROM pg_stat_replication;'

-[ RECORD 1 ]---------------
write_lag  | 00:00:00.072392
flush_lag  | 00:00:00.108168
replay_lag | 00:00:00.108283

If one or more of these values is significantly high, this could indicate a problem and should be investigated further. When determining the cause, consider that:

• write_lag indicates the time since when WAL bytes have been sent by the primary, then received to the secondary, but not yet flushed or applied.
• A high write_lag value may indicate degraded network performance or insufficient network speed between the primary and secondary nodes.
• A high flush_lag value may indicate degraded or sub-optimal disk I/O performance with the secondary node's storage device.
• A high replay_lag value may indicate long running transactions in PostgreSQL, or the saturation of a needed resource like the CPU.
• The difference in time between write_lag and flush_lag indicates that WAL bytes have been sent to the underlying storage system, but it has not reported that they were flushed. This data is most likely not fully written to a persistent storage, and likely held in some kind of volatile write cache.
• The difference between flush_lag and replay_lag indicates WAL bytes that have been successfully persisted to storage, but could not be replayed by the database system.

Resetting Geo secondary site replication

If you get a secondary site in a broken state and want to reset the replication state, to start again from scratch, there are a few steps that can help you:

1. Stop Sidekiq and the Geo LogCursor.

   It's possible to make Sidekiq stop gracefully, but making it stop getting new jobs and wait until the current jobs to finish processing.

   You need to send a SIGTSTP kill signal for the first phase and them a SIGTERM when all jobs have finished. Otherwise just use the gitlab-ctl stop commands.

   gitlab-ctl status sidekiq
   # run: sidekiq: (pid 10180) <- this is the PID you will use
   kill -TSTP 10180 # change to the correct PID
   
   gitlab-ctl stop sidekiq
   gitlab-ctl stop geo-logcursor

   You can watch the Sidekiq logs to know when Sidekiq jobs processing has finished:

   gitlab-ctl tail sidekiq

2. Rename repository storage folders and create new ones. If you are not concerned about possible orphaned directories and files, you can skip this step.

   mv /var/opt/gitlab/git-data/repositories /var/opt/gitlab/git-data/repositories.old
   mkdir -p /var/opt/gitlab/git-data/repositories
   chown git:git /var/opt/gitlab/git-data/repositories

   NOTE: You may want to remove the /var/opt/gitlab/git-data/repositories.old in the future as soon as you confirmed that you don't need it anymore, to save disk space.

3. Optional. Rename other data folders and create new ones.

   WARNING: You may still have files on the secondary site that have been removed from the primary site, but this removal has not been reflected. If you skip this step, these files are not removed from the Geo secondary site.

   Any uploaded content (like file attachments, avatars, or LFS objects) is stored in a subfolder in one of these paths:

   • /var/opt/gitlab/gitlab-rails/shared
   • /var/opt/gitlab/gitlab-rails/uploads

   To rename all of them:

   gitlab-ctl stop
   
   mv /var/opt/gitlab/gitlab-rails/shared /var/opt/gitlab/gitlab-rails/shared.old
   mkdir -p /var/opt/gitlab/gitlab-rails/shared
   
   mv /var/opt/gitlab/gitlab-rails/uploads /var/opt/gitlab/gitlab-rails/uploads.old
   mkdir -p /var/opt/gitlab/gitlab-rails/uploads
   
   gitlab-ctl start postgresql
   gitlab-ctl start geo-postgresql

   Reconfigure to recreate the folders and make sure permissions and ownership are correct:

   gitlab-ctl reconfigure

4. Reset the Tracking Database.

   WARNING: If you skipped the optional step 3, be sure both geo-postgresql and postgresql services are running.

   gitlab-rake db:drop:geo DISABLE_DATABASE_ENVIRONMENT_CHECK=1   # on a secondary app node
   gitlab-ctl reconfigure     # on the tracking database node
   gitlab-rake db:migrate:geo # on a secondary app node

5. Restart previously stopped services.

   gitlab-ctl start