cluster_databases.diviner

@title Cluster: Databases
@group cluster

Configuring Phabricator to use multiple database hosts.

Overview
========

You can deploy Phabricator with multiple database hosts, configured as a master
and a set of replicas. The advantages of doing this are:

  - faster recovery from disasters by promoting a replica;
  - graceful degradation if the master fails; and
  - some tools to help monitor and manage replica health.

This configuration is complex, and many installs do not need to pursue it.

If you lose the master, Phabricator can degrade automatically into read-only
mode and remain available, but can not fully recover without operational
intervention unless the master recovers on its own.

Phabricator will not currently send read traffic to replicas unless the master
has failed, so configuring a replica will not currently spread any load away
from the master. Future versions of Phabricator are expected to be able to
distribute some read traffic to replicas.

Phabricator can not currently be configured into a multi-master mode, nor can
it be configured to automatically promote a replica to become the new master.
There are no current plans to support multi-master mode or autonomous failover,
although this may change in the future.

Phabricator applications //can// be partitioned across multiple database
masters. This does not provide redundancy and generally does not increase
resilience or resistance to data loss, but can help you scale and operate
Phabricator. For details, see
@{article:Cluster: Partitioning and Advanced Configuration}.


Setting up MySQL Replication
============================

To begin, set up a replica database server and configure MySQL replication.

If you aren't sure how to do this, refer to the MySQL manual for instructions.
The MySQL documentation is comprehensive and walks through the steps and
options in good detail. You should understand MySQL replication before
deploying it in production: Phabricator layers on top of it, and does not
attempt to abstract it away.

Some useful notes for configuring replication for Phabricator:

**Binlog Format**: Phabricator issues some queries which MySQL will detect as
unsafe if you use the `STATEMENT` binlog format (the default). Instead, use
`MIXED` (recommended) or `ROW` as the `binlog_format`.

**Grant `REPLICATION CLIENT` Privilege**: If you give the user that Phabricator
will use to connect to the replica database server the `REPLICATION CLIENT`
privilege, Phabricator's status console can give you more information about
replica health and state.

**Copying Data to Replicas**: Phabricator currently uses a mixture of MyISAM
and InnoDB tables, so it can be difficult to guarantee that a dump is wholly
consistent and suitable for loading into a replica because MySQL uses different
consistency mechanisms for the different storage engines.

An approach you may want to consider to limit downtime but still produce a
consistent dump is to leave Phabricator running but configured in read-only
mode while dumping:

  - Stop all the daemons.
  - Set `cluster.read-only` to `true` and deploy the new configuration. The
    web UI should now show that Phabricator is in "Read Only" mode.
  - Dump the database. You can do this with `bin/storage dump --for-replica`
    to add the `--master-data` flag to the underlying command and include a
    `CHANGE MASTER ...` statement in the dump.
  - Once the dump finishes, turn `cluster.read-only` off again to restore
    service. Continue loading the dump into the replica normally.

**Log Expiration**: You can configure MySQL to automatically clean up old
binary logs on startup with the `expire_logs_days` option. If you do not
configure this and do not explicitly purge old logs with `PURGE BINARY LOGS`,
the binary logs on disk will grow unboundedly and relatively quickly.

Once you have a working replica, continue below to tell Phabricator about it.


Configuring Replicas
====================

Once your replicas are in working order, tell Phabricator about them by
configuring the `cluster.databases` option. This option must be configured from
the command line or in configuration files because Phabricator needs to read
it //before// it can connect to databases.

This option value will list all of the database hosts that you want Phabricator
to interact with: your master and all your replicas. Each entry in the list
should have these keys:

  - `host`: //Required string.// The database host name.
  - `role`: //Required string.// The cluster role of this host, one of
    `master` or `replica`.
  - `port`: //Optional int.// The port to connect to. If omitted, the default
    port from `mysql.port` will be used.
  - `user`: //Optional string.// The MySQL username to use to connect to this
    host. If omitted, the default from `mysql.user` will be used.
  - `pass`: //Optional string.// The password to use to connect to this host.
    If omitted, the default from `mysql.pass` will be used.
  - `disabled`: //Optional bool.// If set to `true`, Phabricator will not
    connect to this host. You can use this to temporarily take a host out
    of service.

When `cluster.databases` is configured the `mysql.host` option is not used.
The other MySQL connection configuration options (`mysql.port`, `mysql.user`,
`mysql.pass`) are used only to provide defaults.

Once you've configured this option, restart Phabricator for the changes to take
effect, then continue to "Monitoring Replicas" to verify the configuration.


Monitoring Replicas
===================

You can monitor replicas in {nav Config > Database Servers}. This interface
shows you a quick overview of replicas and their health, and can detect some
common issues with replication.

The table on this page shows each database and current status.

NOTE: This page runs its diagnostics //from the web server that is serving the
request//. If you are recovering from a disaster, the view this page shows
may be partial or misleading, and two requests served by different servers may
see different views of the cluster.

**Connection**: Phabricator tries to connect to each configured database, then
shows the result in this column. If it fails, a brief diagnostic message with
details about the error is shown. If it succeeds, the column shows a rough
measurement of latency from the current webserver to the database.

**Replication**: This is a summary of replication status on the database. If
things are properly configured and stable, the replicas should be actively
replicating and no more than a few seconds behind master, and the master
should //not// be replicating from another database.

To report this status, the user Phabricator is connecting as must have the
`REPLICATION CLIENT` privilege (or the `SUPER` privilege) so it can run the
`SHOW SLAVE STATUS` command. The `REPLICATION CLIENT` privilege only enables
the user to run diagnostic commands so it should be reasonable to grant it in
most cases, but it is not required. If you choose not to grant it, this page
can not show any useful diagnostic information about replication status but
everything else will still work.

If a replica is more than a second behind master, this page will show the
current replication delay. If the replication delay is more than 30 seconds,
it will report "Slow Replication" with a warning icon.

If replication is delayed, data is at risk: if you lose the master and can not
later recover it (for example, because a meteor has obliterated the datacenter
housing the physical host), data which did not make it to the replica will be
lost forever.

Beyond the risk of data loss, any read-only traffic sent to the replica will
see an older view of the world which could be confusing for users: it may
appear that their data has been lost, even if it is safe and just hasn't
replicated yet.

Phabricator will attempt to prevent clients from seeing out-of-date views, but
sometimes sending traffic to a delayed replica is the best available option
(for example, if the master can not be reached).

**Health**: This column shows the result of recent health checks against the
server. After several checks in a row fail, Phabricator will mark the server
as unhealthy and stop sending traffic to it until several checks in a row
later succeed.

Note that each web server tracks database health independently, so if you have
several servers they may have different views of database health. This is
normal and not problematic.

For more information on health checks, see "Unreachable Masters" below.

**Messages**: This column has additional details about any errors shown in the
other columns. These messages can help you understand or resolve problems.


Testing Replicas
================

To test that your configuration can survive a disaster, turn off the master
database. Do this with great ceremony, making a cool explosion sound as you
run the `mysqld stop` command.

If things have been set up properly, Phabricator should degrade to a temporary
read-only mode immediately. After a brief period of unresponsiveness, it will
degrade further into a longer-term read-only mode. For details on how this
works internally, see "Unreachable Masters" below.

Once satisfied, turn the master back on. After a brief delay, Phabricator
should recognize that the master is healthy again and recover fully.

Throughout this process, the {nav Database Servers} console will show a
current view of the world from the perspective of the web server handling the
request. You can use it to monitor state.

You can perform a more narrow test by enabling `cluster.read-only` in
configuration. This will put Phabricator into read-only mode immediately
without turning off any databases.

You can use this mode to understand which capabilities will and will not be
available in read-only mode, and make sure any information you want to remain
accessible in a disaster (like wiki pages or contact information) is really
accessible.

See the next section, "Degradation to Read Only Mode", for more details about
when, why, and how Phabricator degrades.

If you run custom code or extensions, they may not accommodate read-only mode
properly. You should specifically test that they function correctly in
read-only mode and do not prevent you from accessing important information.


Degradation to Read-Only Mode
=============================

Phabricator will degrade to read-only mode when any of these conditions occur:

  - you turn it on explicitly;
  - you configure cluster mode, but don't set up any masters;
  - the master can not be reached while handling a request; or
  - recent attempts to connect to the master have consistently failed.

When Phabricator is running in read-only mode, users can still read data and
browse and clone repositories, but they can not edit, update, or push new
changes. For example, users can still read disaster recovery information on
the wiki or emergency contact information on user profiles.

You can enable this mode explicitly by configuring `cluster.read-only`. Some
reasons you might want to do this include:

  - to test that the mode works like you expect it to;
  - to make sure that information you need will be available;
  - to prevent new writes while performing database maintenance; or
  - to permanently archive a Phabricator install.

You can also enable this mode implicitly by configuring `cluster.databases`
but disabling the master, or by not specifying any host as a master. This may
be more convenient than turning it on explicitly during the course of
operations work.

If Phabricator is unable to reach the master database, it will degrade into
read-only mode automatically. See "Unreachable Masters" below for details on
how this process works.

If you end up in a situation where you have lost the master and can not get it
back online (or can not restore it quickly) you can promote a replica to become
the new master. See the next section, "Promoting a Replica", for details.


Promoting a Replica
===================

If you lose access to the master database, Phabricator will degrade into
read-only mode. This is described in greater detail below.

The easiest way to get out of read-only mode is to restore the master database.
If the database recovers on its own or operations staff can revive it,
Phabricator will return to full working order after a few moments.

If you can't restore the master or are unsure you will be able to restore the
master quickly, you can promote a replica to become the new master instead.

Before doing this, you should first assess how far behind the master the
replica was when the link died. Any data which was not replicated will either
be lost or become very difficult to recover after you promote a replica.

For example, if some `T1234` had been created on the master but had not yet
replicated and you promote the replica, a new `T1234` may be created on the
replica after promotion. Even if you can recover the master later, merging
the data will be difficult because each database may have conflicting changes
which can not be merged easily.

If there was a significant replication delay at the time of the failure, you
may wait to try harder or spend more time attempting to recover the master
before choosing to promote.

If you have made a choice to promote, disable replication on the replica and
mark it as the `master` in `cluster.databases`. Remove the original master and
deploy the configuration change to all surviving hosts.

Once write service is restored, you should provision, deploy, and configure a
new replica by following the steps you took the first time around. You are
critically vulnerable to a second disruption until you have restored the
redundancy.


Unreachable Masters
===================

This section describes how Phabricator determines that a master has been lost,
marks it unreachable, and degrades into read-only mode.

Phabricator degrades into read-only mode automatically in two ways: very
briefly in response to a single connection failure, or more permanently in
response to a series of connection failures.

In the first case, if a request needs to connect to the master but is not able
to, Phabricator will temporarily degrade into read-only mode for the remainder
of that request. The alternative is to fail abruptly, but Phabricator can
sometimes degrade successfully and still respond to the user's request, so it
makes an effort to finish serving the request from replicas.

If the request was a write (like posting a comment) it will fail anyway, but
if it was a read that did not actually need to use the master it may succeed.

This temporary mode is intended to recover as gracefully as possible from brief
interruptions in service (a few seconds), like a server being restarted, a
network link becoming temporarily unavailable, or brief periods of load-related
disruption. If the anomaly is temporary, Phabricator should recover immediately
(on the next request once service is restored).

This mode can be slow for users (they need to wait on connection attempts to
the master which fail) and does not reduce load on the master (requests still
attempt to connect to it).

The second way Phabricator degrades is by running periodic health checks
against databases, and marking them unhealthy if they fail over a longer period
of time. This mechanism is very similar to the health checks that most HTTP
load balancers perform against web servers.

If a database fails several health checks in a row, Phabricator will mark it as
unhealthy and stop sending all traffic (except for more health checks) to it.
This improves performance during a service interruption and reduces load on the
master, which may help it recover from load problems.

You can monitor the status of health checks in the {nav Database Servers}
console. The "Health" column shows how many checks have run recently and
how many have succeeded.

Health checks run every 3 seconds, and 5 checks in a row must fail or succeed
before Phabricator marks the database as healthy or unhealthy, so it will
generally take about 15 seconds for a database to change state after it goes
down or comes up.

If all of the recent checks fail, Phabricator will mark the database as
unhealthy and stop sending traffic to it. If the master was the database that
was marked as unhealthy, Phabricator will actively degrade into read-only mode
until it recovers.

This mode only attempts to connect to the unhealthy database once every few
seconds to see if it is recovering, so performance will be better on average
(users rarely need to wait for bad connections to fail or time out) and the
database will receive less load.

Once all of the recent checks succeed, Phabricator will mark the database as
healthy again and continue sending traffic to it.

Health checks are tracked individually for each web server, so some web servers
may see a host as healthy while others see it as unhealthy. This is normal, and
can accurately reflect the state of the world: for example, the link between
datacenters may have been lost, so hosts in one datacenter can no longer see
the master, while hosts in the other datacenter still have a healthy link to
it.


Backups
======

Even if you configure replication, you should still retain separate backup
snapshots. Replicas protect you from data loss if you lose a host, but they do
not let you recover from data mutation mistakes.

If something issues `DELETE` or `UPDATE` statements and destroys data on the
master, the mutation will propagate to the replicas almost immediately and the
data will be gone forever. Normally, the only way to recover this data is from
backup snapshots.

Although you should still have a backup process, your backup process can
safely pull dumps from a replica instead of the master. This operation can
be slow, so offloading it to a replica can make the performance of the master
more consistent.

To dump from a replica, you can use `bin/storage dump --host <host>` to
control which host the command connects to. (You may still want to execute
this command //from// that host, to avoid sending the whole dump over the
network).

With the `--for-replica` flag, the `bin/storage dump` command creates dumps
with `--master-data`, which includes a `CHANGE MASTER` statement in the output.
This may be helpful when initially setting up new replicas, as it can make it
easier to change the binlog coordinates to the correct position for the dump.

With recent versions of MySQL, it is also possible to configure a //delayed//
replica which intentionally lags behind the master (say, by 12 hours). In the
event of a bad mutation, this could give you a larger window of time to
recognize the issue and recover the lost data from the delayed replica (which
might be quick) without needing to restore backups (which might be very slow).

Delayed replication is outside the scope of this document, but may be worth
considering as an additional data security step on top of backup snapshots
depending on your resources and needs. If you configure a delayed replica, do
not add it to the `cluster.databases` configuration: Phabricator should never
send traffic to it, and does not need to know about it.


Next Steps
==========

Continue by:

  - returning to @{article:Clustering Introduction}.