1
0
Fork 0
arangodb/Documentation/UserManual/UserManualReplication.md

622 lines
30 KiB
Markdown

Replication {#UserManualReplication}
====================================
@NAVIGATE_UserManualReplication
@EMBEDTOC{UserManualReplicationTOC}
Introduction to Replication {#UserManualReplicationIntro}
=========================================================
Starting with ArangoDB 1.4, ArangoDB comes with optional asynchronous master-slave
replication. Replication is configured on a per-database level, meaning that
different databases in the same ArangoDB instance can have different replication
settings. Replication must be turned on explicitly before it becomes active for a
database.
In a typical master-slave replication setup, clients direct *all* their write
operations for a specific database to the master. The master database is the only
place to connect to when making any insertions/updates/deletions.
The master database will log all write operations in its so-called *event log*.
The event log can be considered as an ordered stream of changes for the database.
Any number of slaves can then connect to the master database and fetch data from the
master database's event log. The slaves then can apply all the events from the log in
the same order locally. After that, they will have the same state of data as the master
database.
In this setup, write operations are applied first in the master database, and applied
in the slave database(s) afterwards.
For example, let's assume a write operation is applied and logged in the master database
at point in time t0. To make a slave database apply the same operation, it must first
fetch the write operation's data from master database's event log, then parse it and
apply it locally. This will happen at some point in time after t0, let's say t1.
The difference between t1 and t0 is called the *replication lag*, and it is unavoidable
in asychronous replication. The amount of replication lag depends on many factors, a
few of which are:
- the network capacity between the slaves and the master
- the load of the master and the slaves
- the frequency in which slaves poll the master for updates
Between t0 and t1, the state of data on the master is newer than the state of data
on the slave(s). At point in time t1, the state of data on the master and slave(s)
is consistent again (provided no new data modifications happened on the master in
between). Thus, the replication will lead to an *eventually consistent* state of data.
Transactions are honored in replication, i.e. transactional write operations will
become visible on slaves atomically.
As all write operations will be logged to a master database's event log, the replication
in ArangoDB 1.4 cannot be used for write-scaling. The main purposes of the replication
in ArangoDB 1.4 are to provide read-scalability and "hot backups" for specific databases.
It is possible to connect multiple slave databaes to the same master database. Slave
databases should be used as read-only instances, and no user-initiated write operations
should be carried out on them. Otherwise data conflicts may occur that cannot be solved
automatically, and that will make the replication stop. Master-master (or multi-master)
replication is not supported in ArangoDB 1.4.
The replication in ArangoDB 1.4 is asychronous, meaning that slaves will *pull*
changes from the master database. Slaves need to know to which master database they should
connect to, but a master database is not aware of the slaves that replicate from it.
When the network connection between the master database and a slave goes down, write
operations on the master can continue normally. When the network is up again, slaves
can reconnect to the master database and transfer the remaining changes. This will
happen automatically provided slaves are configured appropriately.
The replication is turned off by default. In order to create a master-slave setup,
the replication features need to be enabled on both the master database and the slave
databases.
Replication is configured on a per-database level. If multiple database are to be
replicated, the replication must be set up individually per database.
Components {#UserManualReplicationComponents}
=============================================
The replication architecture in ArangoDB 1.4 consists of two main components, which
can be used together or in isolation: the *replication logger* and the *replication applier*.
Both are available in ArangoDB 1.4 and can be administered via the command line or
a REST API (see @ref HttpReplication).
In most cases, the *replication logger* will be run inside a master database, and the
*replication applier* will be executed in a slave database.
As there can be multiple database inside an ArangoDB instance, there can be multiple
replication loggers and multiple replication appliers per ArangoDB instance.
Replication Logger {#UserManualReplicationLogger}
-------------------------------------------------
###Purpose
The purpose of the replication logger is to log all changes that modify the state
of data in a specific database. This includes document insertions, updates, and deletions.
It also includes creating, dropping, renaming and changing collections and indexes.
When the replication logger is used, it will log all these write operations for the
database in its *event log*, in the same order in which the operations were carried out
originally.
Reading the event log sequentially provides a list of all write operations carried out
in the database. Replication clients can request certain log events from the logger.
For example, ArangoDB's replication applier will permanently query the latest events
from the replication logger. The applier of the slave database will then apply all changes
locally to get to the same state of data as the master database. It will keep track of
which parts of the event log it already synchronised, meaning that it will perform an
incremental synchronisation.
Technically, the event log is a system collection named `_replication` inside the database
the applier runs in. The event log is persisted and will still be present after a server
shutdown or crash. The event log's underlying collection should not be modified by users
directly. It should only be accessed using the special API methods offered by ArangoDB.
###Starting and Stopping
ArangoDB will only log changes if the replication logger is turned on for the specific
database. Should there be any write operations in the database while the replication logger
is turned off, these events will be lost for replication.
To turn on the replication logger for a database once, the following command can be executed:
require("org/arangodb/replication").logger.start();
Note that starting the replication logger does not necessarily mean it will be started
automatically on all following server startups. This can be configured seperately (keep on
reading for this).
Please note that the above command turns on the replication logger for the current database
only, but not for all databases. To turn on the replication logger for multiple databases,
the above command must be executed separately for each database.
To turn the replication logger off for the current database, execute the `stop` command:
require("org/arangodb/replication").logger.stop();
This has turned off the logger, but still the logger may be started again automatically
the next time the ArangoDB server is started. This can be configured separately (again,
keep on reading).
To determine the current state of the replication logger in the current database (including
whether it is currently running or not), use the `state` command:
require("org/arangodb/replication").logger.state();
The result might look like this:
{
"state" : {
"running" : false,
"lastLogTick" : "255376126918573",
"totalEvents" : 0,
"time" : "2013-08-02T11:01:28Z"
},
"server" : {
"version" : "1.4.devel",
"serverId" : "53904504772335"
},
"clients" : [ ]
}
The `running` attribute indicates whether the logger is currently enabled and will
log any events. The `totalEvents` attribute will indicate how many log events have been
logged since the start of the ArangoDB server. The value will not be reset between
multiple stops and restarts of the logger. Finally, the `lastLogTick` value will indicate
the id of the last event that was logged. It can be used to determine whether new
events were logged, and is also used by the replication applier for incremental
fetching of data.
Note: the replication logger state can also be queried via the HTTP API (see @ref HttpReplication).
###Configuration
To determine whether the current database's replication logger is automatically started
when the ArangoDB server is started, the logger has a separate configuration. The
configuration is stored in a file `REPLICATION-LOGGER-CONFIG` inside the current database's
directory. If it does not exist, ArangoDB will use default values.
To check and adjust the configuration of the replication logger for the current database,
use the `properties` command. To view the current configuration, use `properties` without
any arguments:
require("org/arangodb/replication").logger.properties();
The result might look like this:
{
"autoStart" : false,
"logRemoteChanges" : false,
"maxEvents" : 1048576,
"maxEventsSize" : 134217728
}
The `autoStart` attribute indicates whether the current database's replication logger is
automatically started whenever the ArangoDB server is started. You may want to set it to
`true` for a proper master setup. To do so, supply the updated attributes to the `properties`
command, e.g.:
require("org/arangodb/replication").logger.properties({
autoStart: true
});
This will ensure the replication logger for the current database is automatically started
on all following ArangoDB starts unless `autoStart` is set to `false` again.
Note that only those attributes will be changed that you supplied in the argument to `properties`.
All other configuration values will remain unchanged. Also note that the replication logger
can be reconfigured while it is running.
It a replication logger is turned on, the event log may be allowed to grow indefinitely.
It may be sensible to set a maximum size or a maximum number of events to keep.
If one of these thresholds is reached during logging, the replication logger will start
removing the oldest events from the event log automatically.
The thresholds are set by adjusting the `maxEvents` and `maxEventsSize` attributes via the
`properties` command. The `maxEvents` attribute is the maximum number of events to keep
in the event log before starting to remove oldest events. The `maxEventsSize` is the maximum
cumulated size of event log data (in bytes) that is kept in the log before starting to remove
oldest events. If both are set to a value of `0`, it means that the number and size
of log events to keep is unrestricted. If both are set to a non-zero value, it means
that there are restrictions on both the number and cumulated size of events to keep,
and if either of the restrictions is hit, the deletion will be triggered.
The following command will set the threshold to 5000 events to keep, irrespective of the
cumulated event data sizes:
require("org/arangodb/replication").logger.properties({
maxEvents: 5000,
maxEventsSize: 0
});
Replication Applier {#UserManualReplicationApplier}
---------------------------------------------------
###Purpose
The purpose of the replication applier is to read data from a master database's event log,
and apply them locally. The applier will check the master database for new events periodically.
It will perform an incremental synchronisation, i.e. only asking the master for events
that occurred after the last synchronisation.
The replication applier does not get notified by the master database when there are "new"
events, but uses the pull principle. It might thus take some time (the *replication lag*)
before an event from the master database gets shipped to and applied in a slave database.
The replication applier of a database is run in a separate thread. It may encounter problems
when a log event from the master cannot be applied safely, or when the connection to the master
database goes down (network outage, master database is down or unavailable etc.). In this case,
the database's replication applier thread might terminate itself. It is then up to the
administrator to fix the problem and restart the database's replication applier.
If the replication applier cannot connect to the master database, or the communication fails at
some point during the synchronisation, the replication applier will try to reconnect to
the master database. It will give up reconnecting only after a configurable amount of connection
attempts.
The replication applier state is queryable at any time by using the `state` command of the
applier. This will return the state of the applier of the current database:
require("org/arangodb/replication").applier.state();
The result might look like this:
{
"state" : {
"running" : false,
"lastAppliedContinuousTick" : "231848832948633",
"lastProcessedContinuousTick" : "231848832948633",
"lastAvailableContinuousTick" : null,
"progress" : {
"time" : "2013-08-02T11:40:08Z",
"message" : "applier created",
"failedConnects" : 0
},
"totalRequests" : 0,
"totalFailedConnects" : 0,
"totalEvents" : 0,
"lastError" : {
"errorNum" : 0
},
"time" : "2013-08-02T11:40:22Z"
},
"server" : {
"version" : "1.4.devel",
"serverId" : "53904504772335"
},
"endpoint" : "tcp://master.domain.org:8529"
}
The `running` attribute indicates whether the replication applier of the current database
is currently running and polling the server at `endpoint` for new events.
The `failedConnects` attribute shows how many failed connection attempts the replication
applier currently has encountered in a row. In contrast, the `totalFailedConnects` attribute
indicates how many failed connection attempts the applier has made in total. The
`totalRequest` attribute shows how many requests the applier has sent to the master database
in total. The `totalEvents` attribute shows how many log events the applier has read from the
master.
The `message` sub-attribute of the `progress` sub-attribute gives a brief hint of what the
applier currently does (if it is running). The `lastError` attribute also has an optional
`errorMessage` sub-attribute, showing the latest error message. The `errorNum` sub-attribute of the
`lastError` attribute can be used by clients to programmatically check for errors. It should be `0`
if there is no error, and it should be non-zero when the applier terminated itself due to
a problem.
Here is an example of the state after the replication applier terminated itself due to
(repeated) connection problems:
{
"state" : {
"running" : false,
"progress" : {
"time" : "2013-08-02T12:40:25Z",
"message" : "applier stopped",
"failedConnects" : 6
},
"totalRequests" : 18,
"totalFailedConnects" : 11,
"totalEvents" : 0,
"lastError" : {
"time" : "2013-08-02T12:40:25Z",
"errorMessage" : "could not connect to master at tcp://master.example.org:8529: Could not connect to 'tcp:/...",
"errorNum" : 1400
},
...
}
}
Note: the state of a database's replication applier is queryable via the HTTP API, too.
Please refer to @ref HttpReplication for more details.
###Starting and Stopping
To start and stop the applier in the current database, the `start` and `stop` commands can
be used:
require("org/arangodb/replication").applier.start(<tick>);
require("org/arangodb/replication").applier.stop();
Note that starting a replication applier without setting up an initial configuration will
fail. The replication applier will look for its configuration in a file named
`REPLICATION-APPLIER-CONFIG` in the current database's directory. If the file is not present,
ArangoDB will use some default configuration, but it cannot guess the endpoint (the address
of the master database) the applier should connect to. Thus starting the applier without
configuration will fail.
Note that starting a database's replication applier via the `start` command will not necessarily
start the applier on the next and following ArangoDB server restarts. Additionally, stopping a
database's replication applier manually will not necessarily prevent the applier from being
started again on the next server start. All of this is configurable seperately (hang on reading).
Please note that when starting the replication applier of database, it will resume where it stopped.
This is sensible because replication log events should be applied incrementally. If the
replication applier of a database has never been started before, it needs some `tick` value from the
master database's event log from which to start fetching events.
####Configuration
To configure the replication applier of a specific database, use the `properties` command. Using
it without any arguments will return the current configuration:
require("org/arangodb/replication").applier.properties();
The result might look like this:
{
"requestTimeout" : 300,
"connectTimeout" : 10,
"ignoreErrors" : 0,
"maxConnectRetries" : 10,
"chunkSize" : 0,
"autoStart" : false,
"adaptivePolling" : true
}
Note that there is no `endpoint` attribute configured yet. The `endpoint` attribute is required
for the replication applier to be startable. You may also want to configure a username and password
for the connection via the `username` and `password` attributes.
require("org/arangodb/replication").applier.properties({
endpoint: "tcp://master.domain.org:8529",
username: "root",
password: "secret"
});
This will re-configure the replication applier for the current database. The configuration will be
used from the next start of the replication applier. The replication applier cannot be re-configured
while it is running. It must be stopped first to be re-configured.
To make the replication applier of the current database start automatically when the ArangoDB server
starts, use the `autoStart` attribute.
Setting the `adaptivePolling` attribute to `true` will make the replication applier poll the
master database for changes with a variable frequency. The replication applier will then lower the
frequency when the master is idle, and increase it when the master can provide new events).
Otherwise the replication applier will poll the master database for changes with a constant frequency
of 5 seconds if the master database's replication logger is turned off, and 0.5 seconds if it is
turned on.
To set a timeout for connection and following request attempts, use the `connectTimeout` and
`requestTimeout` values. The `maxConnectRetries` attribute configures after how many failed
connection attempts in a row the replication applier will give up and turn itself off.
You may want to set this to a high value so that temporary network outages do not lead to the
replication applier stopping itself.
The `chunkSize` attribute can be used to control the approximate maximum size of a master's
response (in bytes). Setting it to a low value may make the master respond faster (less data is
assembled before the master sends the response), but may require more request-response roundtrips.
Set it to `0` to use ArangoDB's built-in default value.
The following example will set most of the discussed properties for the current database's applier:
require("org/arangodb/replication").applier.properties({
endpoint: "tcp://master.domain.org:8529",
username: "root",
password: "secret",
adaptivePolling: true,
connectTimeout: 15,
maxConnectRetries: 100,
chunkSize: 262144,
autoStart: true
});
After the applier is now fully configured, it could theoretically be started. However, we
may first need an initial sychronisation of all collections and their data from the master before
we start the replication applier. The reason is that the replication logger of the master database
may not contain all data we need for a full synchronisation.
The only safe method for doing a full synchronisation is thus to
* make sure the replication logger of the master database is running
* perform an intiial full sync with the master database
* note the master database's replication logger `tick` value and
* start the continuous replication applier using this tick value.
The initial synchronisation for the current database is executed with the `sync` command:
require("org/arangodb/replication").sync({
endpoint: "tcp://master.domain.org:8529",
username: "root",
password: "secret
});
Warning: `sync` will do a full synchronisation of the collections in the current database with
collections present in the master database.
Any local instances of the collections and all their data are removed! Only execute this
command when you are sure you want to remove the local data!
As `sync` does a full synchronisation, it may take a while to execute.
When `sync` completes successfully, it show a list of collections it has synchronised in the
`collections` attribute. It will also return the master database's replication logger tick value
at the time the `sync` was started on the master. The tick value is contained in the `lastLogTick`
attribute of the `sync` command:
{
"lastLogTick" : "231848833079705",
"collections" : [ ... ]
}
If the master database's replication logger is turned on, you can now start the continuous
synchronisation for the current database with the command
require("org/arangodb/replication").applier.start("231848833079705");
Note that the tick values should be handled as strings. Using numeric data types for tick
values is unsafe because they might exceed the 32 bit value and the IEEE754 double accuracy
ranges.
Example Setup {#UserManualReplicationSetup}
===========================================
Setting up a working master-slave replication requires two ArangoDB instances:
- _master_: this is the instance we'll activate the replication logger on
- _slave_: on this instance, we'll start a replication applier, and this will fetch data from the
master database's events log and apply all events locally
For the following example setup, we'll use the instance *tcp://master.domain.org:8529* as the
master, and the instance *tcp://slave.domain.org:8530* as a slave.
The goal is to have all data from the database `_system` on master *tcp://master.domain.org:8529*
be replicated to the database `_system` on the slave *tcp://slave.domain.org:8530*.
On the *master*, configure the replication logger for database `_system` to start automatically
on ArangoDB startup. Additionally, set some restrictions for the event log size:
db._useDatabase("_system");
require("org/arangodb/replication").logger.properties({
autoStart: true,
maxEvents: 1048576,
maxEventsSize: 0
});
After that, start the replication logger for the `_system` database on the master:
db._useDatabase("_system");
require("org/arangodb/replication").logger.start();
On the *slave*, make sure there currently is no replication applier running in the `_system`
database:
db._useDatabase("_system");
require("org/arangodb/replication").applier.stop();
After that, do an initial sync of the slave with data from the master. Execute the following
commands on the slave:
db._useDatabase("_system");
require("org/arangodb/replication").sync({
endpoint: "tcp://master.example.org:8529",
username: "myuser",
password: "mypasswd"
});
Warning: this will replace data in the slave database with data from the master database!
Only execute the commands if you have verified you are on the correct server, in the
correct database!
Then re-configure the slave database's replication applier to point to the master database,
set the username and password for the connection, and set the `autoStart` attribute:
db._useDatabase("_system");
require("org/arangodb/replication").applier.properties({
endpoint: "tcp://master.example.org:8529",
username: "myuser",
password: "mypasswd",
autoStart: true,
adaptivePolling: true
});
After that, start the applier in the slave database:
db._useDatabase("_system");
require("org/arangodb/replication").applier.start();
After that, you should be able to monitor the state and progress of the replication
applier by executing the `state` command in the slave database:
db._useDatabase("_system");
require("org/arangodb/replication").applier.state();
You may also want to check the master and slave states via the HTTP APIs (see @ref
HttpReplication) or the web interface ("Dashboard" tab).
Replication Limitations {#UserManualReplicationLimitations}
===========================================================
The replication in ArangoDB 1.4-alpha has a few limitations still. Some of these
limitations may be removed in later versions of ArangoDB:
- the event log of the master database is currently written directly after a write operation
is carried out in the master database. In case the master crashes between having executed the
write operation and having it written into the event log, the write operation may
have been executed on the master, but may be lost for replication and not be applied
on any slaves.
- there is no feedback from the slaves to the master. If a slave cannot apply an event
it got from the master, the master will have a different state of data. In this
case, the replication applier on the slave will stop and report an error. Administrators
can then either "fix" the problem or re-sync the data from the master to the slave
and start the applier again.
- the replication is an asynchronous master-slave replication. There is currently no
way to use it as a synchronous replication, or a multi-master replication.
- at the moment it is assumed that only the replication applier executes write
operations on a slave. ArangoDB currently does not prevent users from carrying out
their own write operations on slaves, though this might lead to undefined behavior
and the replication applier stopping.
- the replication logger will log write operations for all user-defined collections and
only some system collections. Write operations for the following system collections
are excluded intentionally: `_trx`, `_replication`, `_users`, `_aal`, `_fishbowl`,
`_modules` and `_routing`. Write operations for the following system collections
will be logged: `_aqlfunctions`, `_graphs`.
- Foxx applications consist of database entries and application scripts in the file system.
The file system parts of Foxx applications are not tracked anywhere and thus not
replicated in current versions of ArangoDB. To replicate a Foxx application, it is
required to copy the application to the remote server and install it there using the
`foxx-manager` utility.
- master servers do not know which slaves are or will be connected to them. All servers
in a replication setup are currently only loosely coupled. There currently is no way
for a client to query which servers are present in a replication.
- failover must currently be handled by clients or client APIs.
- there currently is one replication logger and one replication applier per ArangoDB
database. It is thus not possible to have a slave apply the events logs from
multiple masters.
- replication is set up on a per-database level. When using ArangoDB with multiple
databases, replication must be configured individually for each database.
- the replication applier is single-threaded, but write operations on the master may
be executed in parallel if they affect different collections. Thus the replication
applier might not be able to catch up with a very powerful and loaded master.
- replication is only supported between ArangoDB 1.4 masters and 1.4 slaves. It is
currently not possible to replicate from/to other ArangoDB versions.
- a replication applier cannot apply data from itself.
- when doing the initial synchronisation from a master to a slave using the `sync`
method, collections are not prevented from being dropped in-between. That means that
when a client fetches the data for collections C1, C2, and C3 (in this order), it
will start with collection C1. Any other client may come along and drop collection
C3 in the meantime. When the client reaches collection C3, it will not find this
collection anymore, and the initial sync will fail. This will be fixed until the
1.4 stable release.
Replication Overhead {#UserManualReplicationOverhead}
=====================================================
Running the replication logger will make all data modification operations more
expensive, as the ArangoDB server will write the operations into the replication log, too.
Additionally, replication appliers that connect to an ArangoDB master will cause some
extra work for the master as it needs to process incoming HTTP requests, and respond.
Overall, turning on the replication logger may reduce throughput on an ArangoDB master
by some extent. If the replication feature is not required, the replication logger should
be turned off.
Transactions are logged to the event log on the master as an uninterrupted sequence.
While a transaction is written to the event log, the event log is blocked for other
writes. Transactions should thus be as small as possible.