cluster_partitioning.diviner

@title Cluster: Partitioning and Advanced Configuration
@group cluster

Guide to partitioning Phabricator applications across multiple database hosts.

Overview
========

You can partition Phabricator's applications across multiple databases. For
example, you can move an application like Files or Maniphest to a dedicated
database host.

The advantages of doing this are:

  - moving heavily used applications to dedicated hardware can help you
    scale; and
  - you can match application workloads to hardware or configuration to make
    operating the cluster easier.

This configuration is complex, and very few installs will benefit from pursuing
it. Phabricator will normally run comfortably with a single database master
even for large organizations.

Partitioning generally does not do much to increase resilience or make it
easier to recover from disasters, and is primarily a mechanism for scaling and
operational convenience.

If you are considering partitioning, you likely want to configure replication
with a single master first. Even if you choose not to deploy replication, you
should review and understand how replication works before you partition. For
details, see @{article:Cluster: Databases}.

Databases also support some advanced configuration options. Briefly:

  - `persistent`: Allows use of persistent connections, reducing pressure on
    outbound ports.

See "Advanced Configuration", below, for additional discussion.


What Partitioning Does
======================

When you partition Phabricator, you move all of the data for one or more
applications (like Maniphest) to a new master database host. This is possible
because Phabricator stores data for each application in its own logical
database (like `phabricator_maniphest`) and performs no joins between databases.

If you're running into scale limits on a single master database, you can move
one or more of your most commonly-used applications to a second database host
and continue adding users. You can keep partitioning applications until all
heavily used applications have dedicated database servers.

Alternatively or additionally, you can partition applications to make operating
the cluster easier. Some applications have unusual workloads or requirements,
and moving them to separate hosts may make things easier to deal with overall.

For example: if Files accounts for most of the data on your install, you might
move it to a different host to make backing up everything else easier.


Configuration Overview
======================

To configure partitioning, you will add multiple entries to `cluster.databases`
with the `master` role. Each `master` should specify a new `partition` key,
which contains a list of application databases it should host.

One master may be specified as the `default` partition. Applications not
explicitly configured to be assigned elsewhere will be assigned here.

When you define multiple `master` databases, you must also specify which master
each `replica` database follows. Here's a simple example config:

```lang=json
...
"cluster.databases": [
  {
    "host": "db001.corporation.com",
    "role": "master",
    "user": "phabricator",
    "pass": "hunter2!trustno1",
    "port": 3306,
    "partition": [
      "default"
    ]
  },
  {
    "host": "db002.corporation.com",
    "role": "replica",
    "user": "phabricator",
    "pass": "hunter2!trustno1",
    "port": 3306,
    "master": "db001.corporation.com:3306"
  },
  {
    "host": "db003.corporation.com",
    "role": "master",
    "user": "phabricator",
    "pass": "hunter2!trustno1",
    "port": 3306,
    "partition": [
      "file",
      "passphrase",
      "slowvote"
    ]
  },
  {
    "host": "db004.corporation.com",
    "role": "replica",
    "user": "phabricator",
    "pass": "hunter2!trustno1",
    "port": 3306,
    "master": "db003.corporation.com:3306"
  }
],
...
```

In this configuration, `db001` is a master and `db002` replicates it.
`db003` is a second master, replicated by `db004`.

Applications have been partitioned like this:

  - `db003`/`db004`: Files, Passphrase, Slowvote
  - `db001`/`db002`: Default (all other applications)

Not all of the database partition names are the same as the application
names. You can get a list of databases with `bin/storage databases` to identify
the correct database names.

After you have configured partitioning, it needs to be committed to the
databases. This writes a copy of the configuration to tables on the databases,
preventing errors if a webserver accidentally starts with an old or invalid
configuration.

To commit the configuration, run this command:

```
phabricator/ $ ./bin/storage partition
```

Run this command after making any partition or clustering changes. Webservers
will not serve traffic if their configuration and the database configuration
differ.


Launching a new Partition
=========================

To add a new partition, follow these steps:

  - Set up the new database host or hosts.
  - Add the new database to `cluster.databases`, but keep its "partition"
    configuration empty (just an empty list). If this is the first time you
    are partitioning, you will need to configure your existing master as the
    new "default". This will let Phabricator interact with it, but won't send
    any traffic to it yet.
  - Run `bin/storage partition`.
  - Run `bin/storage upgrade` to initialize the schemata on the new hosts.
  - Stop writes to the applications you want to move by putting Phabricator
    in read-only mode, or shutting down the webserver and daemons, or telling
    everyone not to touch anything.
  - Dump the data from the application databases on the old master.
  - Load the data into the application databases on the new master.
  - Reconfigure the "partition" setup so that Phabricator knows the databases
    have moved.
  - Run `bin/storage partition`.
  - While still in read-only mode, check that all the data appears to be
    intact.
  - Resume writes.

You can do this with a small, rarely-used application first (on most installs,
Slowvote might be a good candidate) if you want to run through the process
end-to-end before performing a larger, higher-stakes migration.


How Partitioning Works
======================

If you have multiple masters, Phabricator keeps the entire set of schemata up
to date on all of them. When you run `bin/storage upgrade` or other storage
management commands, they generally affect all masters (if they do not, they
will prompt you to be more specific).

When the application goes to read or write normal data (for example, to query a
list of tasks) it only connects to the master which the application it is
acting on behalf of is assigned to.

In most cases, a masters will not have any data in most the databases which are
not assigned to it. If they do (for example, because they previously hosted the
application) the data is ignored. This approach (of maintaining all schemata on
all hosts) makes it easier to move data and to quickly revert changes if a
configuration mistake occurs.

There are some exceptions to this rule. For example, all masters keep track
of which patches have been applied to that particular master so that
`bin/storage upgrade` can upgrade hosts correctly.

Phabricator does not perform joins across logical databases, so there are no
meaningful differences in runtime behavior if two applications are on the same
physical host or different physical hosts.


Advanced Configuration
======================

Separate from partitioning, some advanced configuration is supported. These
options must be set on database specifications in `cluster.databases`. You can
configure them without actually building a cluster by defining a cluster with
only one master.

`persistent` //(bool)// Enables persistent connections. Defaults to off.

With persistent connections enabled, Phabricator will keep a pool of database
connections open between web requests and reuse them when serving subsequent
requests.

The primary benefit of using persistent connections is that it will greatly
reduce pressure on how quickly outbound TCP ports are opened and closed. After
a TCP port closes, it normally can't be used again for about 60 seconds, so
rapidly cycling ports can cause resource exhaustion. If you're seeing failures
because requests are unable to bind to an outbound port, enabling this option
is likely to fix the issue. This option may also slightly increase performance.

The cost of using persistent connections is that you may need to raise the
MySQL `max_connections` setting: although Phabricator will make far fewer
connections, the connections it does make will be longer-lived. Raising this
setting will increase MySQL memory requirements and may run into other limits,
like `open_files_limit`, which may also need to be raised.

Persistent connections are enabled per-database. If you always want to use
them, set the flag on each configured database in `cluster.databases`.


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

Continue by:

  - returning to @{article:Clustering Introduction}.