The Server Deployment Package (SDP) is the implementation of Perforce’s recommendations for operating and managing a production Perforce Helix Core Version Control System. It is intended to provide the Helix Core administration team with tools to help:

  • High Availability (HA)

  • Disaster Recovery (DR)

This guide is intended to provide instructions for failover in an SDP environment using built in Helix Core features.

For more details see:

Please Give Us Feedback

Perforce welcomes feedback from our users. Please send any suggestions for improving this document or the SDP to

1. Overview

We need to consider planned vs unplanned failover. Planned may be due to upgrading the core Operating System or some other dependency in your infrastructure, or a similar activity.

Unplanned covers risks you are seeking to mitigate with failover:

  • loss of a machine, or some machine related hardware failure

  • loss of a VM cluster

  • failure of storage

  • loss of a data center or machine room

  • etc…​

Please refer to the following sections:

1.1. Planning

HA failover should not require a P4PORT change for end users. Depending on your topology, you can avoid changing P4PORT by having users set their P4PORT to a DNS name that is centrally managed.

For example:

  • bos-helix-01, a master/commit-server in Boston, pointed to by a DNS name like perforce or

  • bos-helix-02, a standby replica in Boston, not pointed to by a DNS until failover. In event of failover, it gets pointed to by perforce/ via a DNS change.

Other advanced networking options might be possible if you talk to your local networking gurus (virtual IPs etc).

2. Planned Failover

In this instance you can run p4 failover with the active participation of its upstream server.

We are going to provide examples with the following assumptions:

  • ServerID master_1 is current commit, running on machine p4d-bos-01

  • ServerID p4d_ha_bos is HA server

  • DNS alias perforce is set to p4d-bos-01

2.1. Prerequisites

You need to ensure:

  1. you are running p4d 2018.2 or later for your commit and all replica instances, preferably 2020.1+

    source /p4/common/bin/p4_vars 1
    p4 info | grep version
  2. your failover target server instance is of type standby or forwarding-standby

    On HA machine:

    p4 info
    ServerID: p4d_ha_bos
    Server services: standby
    Replica of: perforce:1999
  3. it has Options mandatory set in its server spec

    p4 server -o p4d_ha_bos | grep Options
    Options:  mandatory
  4. you have a valid license installed in /p4/1/root (<instance> root)

    On HA machine:

    cat /p4/1/root/license
  5. Monitoring is enabled - so the following works:

    p4 monitor show -al
  6. DNS changes are possible so that downstream replicas can seamlessly connect to HA server

  7. Current pull status is valid

    p4 pull -lj
  8. You have a valid offline_db for the HA instance

    Check that the sizes of the db.* are similar - compare output:

    ls -lhSr /p4/1/offline_db/db.* | tail
    ls -lhSr /p4/1/root/db.* | tail

    Check the current journal counter and compare against live journal counter:

    /p4/1/bin/p4d_1 -r /p4/1/offline_db -jd - db.counters | grep journal
    p4 counters | grep journal
  9. Check all defined triggers will work (see next section) - including Swarm triggers

  10. Check authentication will work (e.g. LDAP configuration)

  11. Check firewall for HA host - ensure that other replicas will be able to connect on the appropriate port to the instance (using the DNS alias)

2.1.1. Pre Failover Checklist

It is important to perform these checks before any planned failover, and also to make sure they are considered prior any unplanned failover. Confirm Trust Established with Replicas

If using ssl on the commit server, then downstream replicas with have already set up P4TRUST entries with that server and its fingerprint.

After failover, you need to check that all replicas have appropriate p4 trust setup with the HA server (which is now the commit server).

If you are using background submit between edge and commit servers, then you need to confirm the following:

  • p4 trust is setup on the HA (new commit) with all edge servers

  • the commit service user is correctly logged on to all edge servers (make sure the configurable is appropriately set to avoid IP-only tickets) on the HA server before failover Swarm Triggers

If Swarm is installed, ensure:

  1. Swarm trigger is installed on HA machine (could be executed from a checked in depot file)

    Typically installed (via package) to /opt/perforce/swarm-triggers/bin/

    But can be installed anywhere on the filesystem

    Execute the trigger to ensure that any required Perl modules are installed:


    Note that things like can often be installed with:

    sudo apt-get install libjson-perl


    sudo yum install perl-JSON
  2. Swarm trigger configuration file has been copied over from commit server to appropriate place.

    This defaults to one of:

    swarm-trigger.conf (in the same directory as trigger script) Other Triggers


  • Make sure that the appropriate version of perl, python, ruby etc are installed in locations as referenced by p4 triggers entries.

  • Make sure that any relevant modules and dependencies have also been installed (e.g. P4Python or P4Perl) Other Replica’s P4TARGET

Review the settings for other replicas and also to check live replicas on the source server of the failover (p4 servers -J)

p4 configure show allservers | grep P4TARGET

Make sure the above settings are using correct DNS alias (which will be redirected). Proxies

These are typically configured via /p4/common/bin/p4_1.vars settings:


Ensure the target is the correct DNS alias (which will be redirected). HA Server OS Configuration

Check to make sure any performance configuration such as turning off THP (transparent huge pages), and putting serverlocks.dir into a RAM filesystem have also been made to your HA Failover system. See SDP Guide: Maximizing Server Performance

2.2. Failing over

The basic actions are the same for Unix and Windows, but there are extra steps required as noted in Section 2.4, “Failing over on Windows”

  1. Rotate the active journal on the master:

    p4 admin journal
    p4 servers -J       # Check all is up-to-date
  2. Run p4 failover in reporting mode on HA machine:

    p4 failover

    Successful output looks like:

    Checking if failover might be possible ...
    Checking for archive file content not transferred ...
    Verifying content of recently update archive files ...
    After addressing any reported issues that might prevent failover, use --yes or -y to execute the failover.
  3. Perform failover:

    p4 failover --yes

    Output should be something like:

    Starting failover process ...
    Refusing new commands on server from which failover is occurring ...
    Giving commands already running time to complete ...
    Stalling commands on server from which failover is occurring ...
    Waiting for 'journalcopy' to complete its work ...
    Waiting for 'pull -L' to complete its work ...
    Waiting for 'pull -u' to complete its work ...
    Checking for archive file content not transferred ...
    Verifying content of recently updated archive files ...
    Stopping server from which failover is occurring ...
    Moving latest journalcopy'd journal into place as the active journal ...
    Updating configuration of the failed-over server ...
    Restarting this server ...

    During this time, if you run commands against the master, you may see:

    Server currently in failover mode, try again after failover has completed
  4. Change the DNS entries so downstream replicas (and users) will connect to the new master (that was previously HA)

  5. Validate that your downstream replicas are communicating with your new master

    On each replica machine:

    p4 pull -lj

    Or against the new master:

    p4 servers -J

    Check output of p4 info:

    Server address: p4d-bos-02
    ServerID: master_1
    Server services: commit-server
  6. Make sure the old server spec (p4d_ha_bos) has correctly had its Options: field set to nomandatory (otherwise all replication would stop!)

2.3. Post Failover

2.3.1. Validation of SDP Config

  1. Run the script:

    /p4/common/bin/ 1

(or specify other instance ID).

Ensure that all output is valid and that there are no errors.

Ensure that the following returns 'yes':

/p4/common/bin/ 1 echo yes

If it does not, then run: bash -xv /p4/common/bin/ 1 echo yes and check the output. The most likely problems are things like the settings of the following variables in /p4/common/config/p4_1.vars:

P4MASTERHOST   <== Set to DNS name of new master host
P4MASTER_ID    <== Make sure this is same as value in /p4/1/root/ on the master server machine.

2.3.2. Moving of Checkpoints

After failing over, on Unix there will be journals which may need to be copied/moved and renamed due to the SDP structure.

For example, an HA server might have stored its journals in /p4/1/checkpoints.ha_bos (assuming it was create by with serverid p4d_ha_bos):


As a result of failover, these files need to be copied/moved to:

Easiest way is to install prename on your Linux. Then p4rename -n 's/.ha_bos//' p4_1.* - and re-run without the -n if it looks good.

The reason different directories are sometimes used is because in some installations the /hxdepots filesystem is shared on NFS between commit server and HA server. When /hxdepots is shared, the journalPrefix must be set differently to avoid path conflicts on the shared storage.

If these files are not present, then normal SDP crontab tasks such as will fail as they won’t be able to find the required journals to be applied to the offline_db.

The following command will rotate journal and replay any missing ones to offline_db (it is both fairly quick and safe to run without placing much load on the server host as it doesn’t do any checkpointing):

/p4/common/bin/ 1

If it fails, then check /p4/1/logs/checkpoint.log for details - it may have an error such as:

Replay journal /p4/1/checkpoints/p4_1.jnl.123 to offline db.
Perforce server error:
     open for read: /p4/1/checkpoints/p4_1.jnl.123: No such file or directory

This indicates missing journals which will need to be moved/copied as above.

2.3.3. Check runs successfully

This is an important script for a master server, and it is good to make sure ASAP after failover that it works as expected.

nohup /p4/common/bin/ 1 &

And check output in: /p4/1/logs/checkpoint.log

This script can take a while on big repositories. It is OK to have it run in the crontab over night.

2.3.4. Check on Replication

We recommend that you connect to all your replicas/proxies/brokers and make sure that they are successfully working after failover.

It is surprisingly common to find forgotten configuration details meaning that they are attempting to connect to old server for example!

For proxies and brokers - you probably just need to run:

p4 info

For downstream replicas of any type, we recommend logging on to the host and running:

p4 pull -lj

and checking for any errors.

We also recommend the following is executed on both HA server and all replicas and the output examined for any unexpected errors:

grep -A4 'Perforce server error:' /p4/1/logs/log

Or you can review the contents of /p4/1/logs/errors.csv if you have enabled structured logging.

2.4. Failing over on Windows

The basic steps are the same as for on Unix, but with some extra steps at the end.

After the p4 failover --yes command has completed its work (on the HA server machine):

  1. Review the settings for the Windows service (examples are for instance 1) - note below -S is uppercase:

    p4 set -S p4_1

    Example results:

    C:\p4\1>p4 set -S p4_1
    P4JOURNAL=c:\p4\1\logs\journal (set -S)
    P4LOG=c:\p4\1\logs\p4d_ha_aws.log (set -S)
    P4NAME=p4d_ha_aws (set -S)
    P4PORT=1666 (set -S)
    P4ROOT=c:\p4\1\root (set -S)
  2. Change the value of P4NAME and P4LOG to correct value for master:

    p4 set -S p4_1 P4NAME=master
    p4 set -S p4_1 P4LOG=c:\p4\1\logs\master.log

    And re-check the output of p4 set -S p4_1

  3. Restart the service:

    c:\p4\common\bin\svcinst stop -n p4_1
    c:\p4\common\bin\svcinst start -n p4_1
  4. Run p4 configure show to check that the output is as expected for the above values.

2.4.1. Post Failover on Windows

This is slightly different for the Windows and Linux SDP, since there is currently no equivalent of for Windows, and replication topologies on Windows are typically smaller and simpler.

Thus it is Windows HA instances are likely to have journals rotated into c:\p4\1\checkpoints instead of something like c:\p4\1\checkpoints.ha_bos.

However, it is still worth ensuring things like:

  • offline_db on HA is up-to-date

  • all triggers (including any Swarm triggers) are appropriately configured

  • daily_backup.bat will work appropriate after failover

2.5. Preparation for Failing Back - Setup original commit server as new Failover replica

After failover it is quite common to want to to "failback" to the original machine. This is fairly straight forwards if you are going to use the same server ids (master and failover replica).

The steps are:

  • Reset the original master to be the failover replica

  • Once it is up and running as a replica, then you can just use p4 failover on it, and then use these same instructions to reset the original HA replica to be running as a failover target.

The original master p4d service will have been stopped as part of failover, and will not be able to restart without some configuration changes.

Note that for convenience below, we assume SDP instance 1 and failover serverid p4d_ha_bos.

The commands below are run on the on original master server which will now become the HA failover replica.
  1. Change the serverid to replica serverid:

    cd /p4/1/root
    mv save/
    mv state* save/
    echo p4d_ha_bos>
  2. Reset the journal and logs:

    cd /p4/1/logs
    mkdir -p save
    mv journal* save/
    mv log log.old
  3. Restart the service and ensure the serviceUser is logged in:

    sudo systemctl restart p4d_1
    /p4/common/bin/p4login -v -service 1
    p4 pull -lj
  4. You may wish to ensure that verification of recent archive/depot files has been done, especially if the service was stopped for hours or days since the original failover.

    nohup /p4/common/bin/ -o MISSING -recent &

Please note details for definition of "recent" and how to extend that:

At this stage, the replica should be replicating as normal. If it is not, then please follow standard replication trouble shooting procedures.

The Server spec will have had its Options: set to nomandatory - and this should be set back to mandatory when the replica is working correctly.

3. Unplanned Failover

In this case there is no active participation of upstream server, so there is an increased risk of lost data.

We assume we are still failing over to the HA machine, so:

  • Failover target is standby or forwarding-standby

  • Server spec still has Options: set to mandatory

  • Original master is not running

The output of p4 failover on the DR machine might be:

Checking if failover might be possible ...
Server ID must be specified in the '-s' or --serverid' argument for a failover without the participation of the server from which failover is occurring.
Checking for archive file content not transferred ...
Verifying content of recently update archive files ...
After addressing any reported issues that might prevent failover, use --yes or -y to execute the failover.
  1. Execute p4 failover with the extra parameter to specify server we are failing over from:

    p4 failover --serverid master_1 --yes

Expected output is somewhat shorter than for planned failover:

Starting failover process ...
Waiting for 'pull -L' to complete its work ...
Checking for archive file content not transferred ...
Verifying content of recently updated archive files ...
Moving latest journalcopy'd journal into place as the active journal ...
Updating configuration of the failed-over server ...
Restarting this server ...

3.1. Post Unplanned Failover

This is similar to Section 2.3, “Post Failover” with the exception of the next section below.

3.1.1. Resetting Downstream Replicas

In an unplanned failover scenario, if the standby server is not a 'mandatory' standby, it is possible that there is a journal synchronization problem with downstream replicas.

The output of p4 pull -lj may indicate an error, and/or there may be errors in the log:

grep -A4 error: /p4/1/logs/log | less

If you need to reset the replica to restart from the beginning of the current journal it is attempting to pull, then the process is:

  1. Stop the replica:

    sudo systemctl stop p4d_1
  2. Remove the state file:

    cd /p4/1/root
    mv state save/.
  3. Restart the replica

    sudo systemctl start p4d_1
  4. Recheck the log for errors as above.

3.2. Unplanned Failover on Windows

The extra steps required are basically the same as in Section 2.4, “Failing over on Windows” as well as the steps in Section 3.1, “Post Unplanned Failover”