Google Git
Sign in
apache / zookeeper / 7e83fd9064a2173778a65cb9fd1eada2975508ee / . / docs / zookeeperReconfig.html
blob: fabcf3a6a5d6648c4fb225f921608024fe26ee9a [file] [log] [blame]
<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN" "http://www.w3.org/TR/html4/loose.dtd">
<html>
<head>
<META http-equiv="Content-Type" content="text/html; charset=UTF-8">
<meta content="Apache Forrest" name="Generator">
<meta name="Forrest-version" content="0.9">
<meta name="Forrest-skin-name" content="pelt">
<title>ZooKeeper Dynamic Reconfiguration</title>
<link type="text/css" href="skin/basic.css" rel="stylesheet">
<link media="screen" type="text/css" href="skin/screen.css" rel="stylesheet">
<link media="print" type="text/css" href="skin/print.css" rel="stylesheet">
<link type="text/css" href="skin/profile.css" rel="stylesheet">
<script src="skin/getBlank.js" language="javascript" type="text/javascript"></script><script src="skin/getMenu.js" language="javascript" type="text/javascript"></script><script src="skin/fontsize.js" language="javascript" type="text/javascript"></script>
<link rel="shortcut icon" href="images/favicon.ico">
</head>
<body onload="init()">
<script type="text/javascript">ndeSetTextSize();</script>
<div id="top">
<!--+
|breadtrail
+-->
<div class="breadtrail">
<a href="http://www.apache.org/">Apache</a> &gt; <a href="http://zookeeper.apache.org/">ZooKeeper</a> &gt; <a href="http://zookeeper.apache.org/">ZooKeeper</a><script src="skin/breadcrumbs.js" language="JavaScript" type="text/javascript"></script>
</div>
<!--+
|header
+-->
<div class="header">
<!--+
|start group logo
+-->
<div class="grouplogo">
<a href="http://hadoop.apache.org/"><img class="logoImage" alt="Hadoop" src="images/hadoop-logo.jpg" title="Apache Hadoop"></a>
</div>
<!--+
|end group logo
+-->
<!--+
|start Project Logo
+-->
<div class="projectlogo">
<a href="http://zookeeper.apache.org/"><img class="logoImage" alt="ZooKeeper" src="images/zookeeper_small.gif" title="ZooKeeper: distributed coordination"></a>
</div>
<!--+
|end Project Logo
+-->
<!--+
|start Search
+-->
<div class="searchbox">
<form action="http://www.google.com/search" method="get" class="roundtopsmall">
<input value="zookeeper.apache.org" name="sitesearch" type="hidden"><input onFocus="getBlank (this, 'Search the site with google');" size="25" name="q" id="query" type="text" value="Search the site with google">&nbsp;
<input name="Search" value="Search" type="submit">
</form>
</div>
<!--+
|end search
+-->
<!--+
|start Tabs
+-->
<ul id="tabs">
<li>
<a class="unselected" href="http://zookeeper.apache.org/">Project</a>
</li>
<li>
<a class="unselected" href="https://cwiki.apache.org/confluence/display/ZOOKEEPER/">Wiki</a>
</li>
<li class="current">
<a class="selected" href="index.html">ZooKeeper 3.4 Documentation</a>
</li>
</ul>
<!--+
|end Tabs
+-->
</div>
</div>
<div id="main">
<div id="publishedStrip">
<!--+
|start Subtabs
+-->
<div id="level2tabs"></div>
<!--+
|end Endtabs
+-->
<script type="text/javascript"><!--
document.write("Last Published: " + document.lastModified);
// --></script>
</div>
<!--+
|breadtrail
+-->
<div class="breadtrail">
&nbsp;
</div>
<!--+
|start Menu, mainarea
+-->
<!--+
|start Menu
+-->
<div id="menu">
<div onclick="SwitchMenu('menu_1.1', 'skin/')" id="menu_1.1Title" class="menutitle">Overview</div>
<div id="menu_1.1" class="menuitemgroup">
<div class="menuitem">
<a href="index.html">Welcome</a>
</div>
<div class="menuitem">
<a href="zookeeperOver.html">Overview</a>
</div>
<div class="menuitem">
<a href="zookeeperStarted.html">Getting Started</a>
</div>
<div class="menuitem">
<a href="releasenotes.html">Release Notes</a>
</div>
</div>
<div onclick="SwitchMenu('menu_1.2', 'skin/')" id="menu_1.2Title" class="menutitle">Developer</div>
<div id="menu_1.2" class="menuitemgroup">
<div class="menuitem">
<a href="api/index.html">API Docs</a>
</div>
<div class="menuitem">
<a href="zookeeperProgrammers.html">Programmer's Guide</a>
</div>
<div class="menuitem">
<a href="javaExample.html">Java Example</a>
</div>
<div class="menuitem">
<a href="zookeeperTutorial.html">Barrier and Queue Tutorial</a>
</div>
<div class="menuitem">
<a href="recipes.html">Recipes</a>
</div>
</div>
<div onclick="SwitchMenu('menu_selected_1.3', 'skin/')" id="menu_selected_1.3Title" class="menutitle" style="background-image: url('skin/images/chapter_open.gif');">Admin &amp; Ops</div>
<div id="menu_selected_1.3" class="selectedmenuitemgroup" style="display: block;">
<div class="menuitem">
<a href="zookeeperAdmin.html">Administrator's Guide</a>
</div>
<div class="menuitem">
<a href="zookeeperQuotas.html">Quota Guide</a>
</div>
<div class="menuitem">
<a href="zookeeperJMX.html">JMX</a>
</div>
<div class="menuitem">
<a href="zookeeperObservers.html">Observers Guide</a>
</div>
<div class="menupage">
<div class="menupagetitle">Dynamic Reconfiguration</div>
</div>
</div>
<div onclick="SwitchMenu('menu_1.4', 'skin/')" id="menu_1.4Title" class="menutitle">Contributor</div>
<div id="menu_1.4" class="menuitemgroup">
<div class="menuitem">
<a href="zookeeperInternals.html">ZooKeeper Internals</a>
</div>
</div>
<div onclick="SwitchMenu('menu_1.5', 'skin/')" id="menu_1.5Title" class="menutitle">Miscellaneous</div>
<div id="menu_1.5" class="menuitemgroup">
<div class="menuitem">
<a href="https://cwiki.apache.org/confluence/display/ZOOKEEPER">Wiki</a>
</div>
<div class="menuitem">
<a href="https://cwiki.apache.org/confluence/display/ZOOKEEPER/FAQ">FAQ</a>
</div>
<div class="menuitem">
<a href="http://zookeeper.apache.org/mailing_lists.html">Mailing Lists</a>
</div>
</div>
<div id="credit"></div>
<div id="roundbottom">
<img style="display: none" class="corner" height="15" width="15" alt="" src="skin/images/rc-b-l-15-1body-2menu-3menu.png"></div>
<!--+
|alternative credits
+-->
<div id="credit2"></div>
</div>
<!--+
|end Menu
+-->
<!--+
|start content
+-->
<div id="content">
<div title="Portable Document Format" class="pdflink">
<a class="dida" href="zookeeperReconfig.pdf"><img alt="PDF -icon" src="skin/images/pdfdoc.gif" class="skin"><br>
PDF</a>
</div>
<h1>ZooKeeper Dynamic Reconfiguration</h1>
<div id="front-matter">
<div id="minitoc-area">
<ul class="minitoc">
<li>
<a href="#ch_reconfig_intro">Overview</a>
</li>
<li>
<a href="#ch_reconfig_format">Changes to Configuration Format</a>
<ul class="minitoc">
<li>
<a href="#sc_reconfig_clientport">Specifying the client port</a>
</li>
<li>
<a href="#sc_reconfig_standaloneEnabled">The standaloneEnabled flag</a>
</li>
<li>
<a href="#sc_reconfig_file">Dynamic configuration file</a>
</li>
<li>
<a href="#sc_reconfig_backward">Backward compatibility</a>
</li>
</ul>
</li>
<li>
<a href="#ch_reconfig_upgrade">Upgrading to 3.5.0</a>
</li>
<li>
<a href="#ch_reconfig_dyn">Dynamic Reconfiguration of the ZooKeeper Ensemble</a>
<ul class="minitoc">
<li>
<a href="#sc_reconfig_retrieving">Retrieving the current dynamic configuration</a>
</li>
<li>
<a href="#sc_reconfig_modifying">Modifying the current dynamic configuration</a>
<ul class="minitoc">
<li>
<a href="#sc_reconfig_general">General</a>
</li>
<li>
<a href="#sc_reconfig_incremental">Incremental mode</a>
</li>
<li>
<a href="#sc_reconfig_nonincremental">Non-incremental mode</a>
</li>
<li>
<a href="#sc_reconfig_conditional">Conditional reconfig</a>
</li>
<li>
<a href="#sc_reconfig_errors">Error conditions</a>
</li>
<li>
<a href="#sc_reconfig_additional">Additional comments</a>
</li>
</ul>
</li>
</ul>
</li>
<li>
<a href="#ch_reconfig_rebalancing">Rebalancing Client Connections</a>
</li>
</ul>
</div>
</div>
<a name="ch_reconfig_intro"></a>
<h2 class="h3">Overview</h2>
<div class="section">
<p>Prior to the 3.5.0 release, the membership and all other configuration
parameters of Zookeeper were static - loaded during boot and immutable at
runtime. Operators resorted to ''rolling restarts'' - a manually intensive
and error-prone method of changing the configuration that has caused data
loss and inconsistency in production.</p>
<p>Starting with 3.5.0, &ldquo;rolling restarts&rdquo; are no longer needed!
ZooKeeper comes with full support for automated configuration changes: the
set of Zookeeper servers, their roles (participant / observer), all ports,
and even the quorum system can be changed dynamically, without service
interruption and while maintaining data consistency. Reconfigurations are
performed immediately, just like other operations in ZooKeeper. Multiple
changes can be done using a single reconfiguration command. The dynamic
reconfiguration functionality does not limit operation concurrency, does
not require client operations to be stopped during reconfigurations, has a
very simple interface for administrators and no added complexity to other
client operations.</p>
<p>New client-side features allow clients to find out about configuration
changes and to update the connection string (list of servers and their
client ports) stored in their ZooKeeper handle. A probabilistic algorithm
is used to rebalance clients across the new configuration servers while
keeping the extent of client migrations proportional to the change in
ensemble membership.</p>
<p>This document provides the administrator manual for reconfiguration.
For a detailed description of the reconfiguration algorithms, performance
measurements, and more, please see our paper:</p>
<dl>
<dt>
<term>Shraer, A., Reed, B., Malkhi, D., Junqueira, F. Dynamic
Reconfiguration of Primary/Backup Clusters. In <em>USENIX Annual
Technical Conference (ATC) </em>(2012), 425-437</term>
</dt>
<dd>
<p>Links: <a href="https://www.usenix.org/system/files/conference/atc12/atc12-final74.pdf">paper (pdf)</a>, <a href="https://www.usenix.org/sites/default/files/conference/protected-files/shraer_atc12_slides.pdf">slides (pdf)</a>, <a href="https://www.usenix.org/conference/atc12/technical-sessions/presentation/shraer">video</a>, <a href="http://www.slideshare.net/Hadoop_Summit/dynamic-reconfiguration-of-zookeeper">hadoop summit slides</a>
</p>
</dd>
</dl>
</div>
<a name="ch_reconfig_format"></a>
<h2 class="h3">Changes to Configuration Format</h2>
<div class="section">
<a name="sc_reconfig_clientport"></a>
<h3 class="h4">Specifying the client port</h3>
<p>A client port of a server is the port on which the server accepts
client connection requests. Starting with 3.5.0 the
<em>clientPort</em> and <em>clientPortAddress
</em> configuration parameters should no longer be used. Instead,
this information is now part of the server keyword specification, which
becomes as follows:</p>
<p>
<span class="codefrag computeroutput">server.&lt;positive id&gt; = &lt;address1&gt;:&lt;port1&gt;:&lt;port2&gt;[:role];[&lt;client port address&gt;:]&lt;client port&gt;</span>
</p>
<p>The client port specification is to the right of the semicolon. The
client port address is optional, and if not specified it defaults to
"0.0.0.0". As usual, role is also optional, it can be
<em>participant</em> or <em>observer</em>
(<em>participant</em> by default).</p>
<p> Examples of legal server statements: </p>
<ul>
<li>
<p>
<span class="codefrag computeroutput">server.5 = 125.23.63.23:1234:1235;1236</span>
</p>
</li>
<li>
<p>
<span class="codefrag computeroutput">server.5 = 125.23.63.23:1234:1235:participant;1236</span>
</p>
</li>
<li>
<p>
<span class="codefrag computeroutput">server.5 = 125.23.63.23:1234:1235:observer;1236</span>
</p>
</li>
<li>
<p>
<span class="codefrag computeroutput">server.5 = 125.23.63.23:1234:1235;125.23.63.24:1236</span>
</p>
</li>
<li>
<p>
<span class="codefrag computeroutput">server.5 = 125.23.63.23:1234:1235:participant;125.23.63.23:1236</span>
</p>
</li>
</ul>
<a name="sc_reconfig_standaloneEnabled"></a>
<h3 class="h4">The standaloneEnabled flag</h3>
<p>Prior to 3.5.0, one could run ZooKeeper in Standalone mode or in a
Distributed mode. These are separate implementation stacks, and
switching between them during run time is not possible. By default (for
backward compatibility) <em>standaloneEnabled</em> is set to
<em>true</em>. The consequence of using this default is that
if started with a single server the ensemble will not be allowed to
grow, and if started with more than one server it will not be allowed to
shrink to contain fewer than two participants.</p>
<p>Setting the flag to <em>false</em> instructs the system
to run the Distributed software stack even if there is only a single
participant in the ensemble. To achieve this the (static) configuration
file should contain:</p>
<p>
<span class="codefrag computeroutput">standaloneEnabled=false</span>
</p>
<p>With this setting it is possible to start a ZooKeeper ensemble
containing a single participant and to dynamically grow it by adding
more servers. Similarly, it is possible to shrink an ensemble so that
just a single participant remains, by removing servers.</p>
<p>Since running the Distributed mode allows more flexibility, we
recommend setting the flag to <em>false</em>. We expect that
the legacy Standalone mode will be deprecated in the future.</p>
<a name="sc_reconfig_file"></a>
<h3 class="h4">Dynamic configuration file</h3>
<p>Starting with 3.5.0 we're distinguishing between dynamic
configuration parameters, which can be changed during runtime, and
static configuration parameters, which are read from a configuration
file when a server boots and don't change during its execution. For now,
the following configuration keywords are considered part of the dynamic
configuration: <em>server</em>, <em>group</em>
and <em>weight</em>.</p>
<p>Dynamic configuration parameters are stored in a separate file on
the server (which we call the dynamic configuration file). This file is
linked from the static config file using the new
<em>dynamicConfigFile</em> keyword.</p>
<p>
<strong>Example</strong>
</p>
<div class="note example">
<div class="label">zoo_replicated1.cfg</div>
<div class="content">
<title>zoo_replicated1.cfg</title>
<pre class="code">tickTime=2000
dataDir=/zookeeper/data/zookeeper1
initLimit=5
syncLimit=2
dynamicConfigFile=/zookeeper/conf/zoo_replicated1.cfg.dynamic</pre>
</div>
</div>
<div class="note example">
<div class="label">zoo_replicated1.cfg.dynamic</div>
<div class="content">
<title>zoo_replicated1.cfg.dynamic</title>
<pre class="code">server.1=125.23.63.23:2780:2783:participant;2791
server.2=125.23.63.24:2781:2784:participant;2792
server.3=125.23.63.25:2782:2785:participant;2793</pre>
</div>
</div>
<p>When the ensemble configuration changes, the static configuration
parameters remain the same. The dynamic parameters are pushed by
ZooKeeper and overwrite the dynamic configuration files on all servers.
Thus, the dynamic configuration files on the different servers are
usually identical (they can only differ momentarily when a
reconfiguration is in progress, or if a new configuration hasn't
propagated yet to some of the servers). Once created, the dynamic
configuration file should not be manually altered. Changed are only made
through the new reconfiguration commands outlined below. Note that
changing the config of an offline cluster could result in an
inconsistency with respect to configuration information stored in the
ZooKeeper log (and the special configuration znode, populated from the
log) and is therefore highly discouraged.</p>
<p>
<strong>Example 2</strong>
</p>
<p>Users may prefer to initially specify a single configuration file.
The following is thus also legal:</p>
<div class="note example">
<div class="label">zoo_replicated1.cfg</div>
<div class="content">
<title>zoo_replicated1.cfg</title>
<pre class="code">tickTime=2000
dataDir=/zookeeper/data/zookeeper1
initLimit=5
syncLimit=2
clientPort=<strong>2791</strong> // note that this line is now redundant and therefore not recommended
server.1=125.23.63.23:2780:2783:participant;<strong>2791</strong>
server.2=125.23.63.24:2781:2784:participant;2792
server.3=125.23.63.25:2782:2785:participant;2793</pre>
</div>
</div>
<p>The configuration files on each server will be automatically split
into dynamic and static files, if they are not already in this format.
So the configuration file above will be automatically transformed into
the two files in Example 1. Note that the clientPort and
clientPortAddress lines (if specified) will be automatically removed
during this process, if they are redundant (as in the example above).
The original static configuration file is backed up (in a .bak
file).</p>
<a name="sc_reconfig_backward"></a>
<h3 class="h4">Backward compatibility</h3>
<p>We still support the old configuration format. For example, the
following configuration file is acceptable (but not recommended):</p>
<div class="note example">
<div class="label">zoo_replicated1.cfg</div>
<div class="content">
<title>zoo_replicated1.cfg</title>
<pre class="code">tickTime=2000
dataDir=/zookeeper/data/zookeeper1
initLimit=5
syncLimit=2
clientPort=2791
server.1=125.23.63.23:2780:2783:participant
server.2=125.23.63.24:2781:2784:participant
server.3=125.23.63.25:2782:2785:participant</pre>
</div>
</div>
<p>During boot, a dynamic configuration file is created and contains
the dynamic part of the configuration as explained earlier. In this
case, however, the line "clientPort=2791" will remain in the static
configuration file of server 1 since it is not redundant -- it was not
specified as part of the "server.1=..." using the format explained in
the section <a href="#ch_reconfig_format">Changes to Configuration Format</a>. If a reconfiguration
is invoked that sets the client port of server 1, we remove
"clientPort=2791" from the static configuration file (the dynamic file
now contain this information as part of the specification of server
1).</p>
</div>
<a name="ch_reconfig_upgrade"></a>
<h2 class="h3">Upgrading to 3.5.0</h2>
<div class="section">
<p>Upgrading a running ZooKeeper ensemble to 3.5.0 should be done only
after upgrading your ensemble to the 3.4.6 release. Note that this is only
necessary for rolling upgrades (if you're fine with shutting down the
system completely, you don't have to go through 3.4.6). If you attempt a
rolling upgrade without going through 3.4.6 (for example from 3.4.5), you
may get the following error:</p>
<pre class="code">2013-01-30 11:32:10,663 [myid:2] - INFO [localhost/127.0.0.1:2784:QuorumCnxManager$Listener@498] - Received connection request /127.0.0.1:60876
2013-01-30 11:32:10,663 [myid:2] - WARN [localhost/127.0.0.1:2784:QuorumCnxManager@349] - Invalid server id: -65536</pre>
<p>During a rolling upgrade, each server is taken down in turn and
rebooted with the new 3.5.0 binaries. Before starting the server with
3.5.0 binaries, we highly recommend updating the configuration file so
that all server statements "server.x=..." contain client ports (see the
section <a href="#sc_reconfig_clientport">Specifying the client port</a>). As explained earlier
you may leave the configuration in a single file, as well as leave the
clientPort/clientPortAddress statements (although if you specify client
ports in the new format, these statements are now redundant).</p>
</div>
<a name="ch_reconfig_dyn"></a>
<h2 class="h3">Dynamic Reconfiguration of the ZooKeeper Ensemble</h2>
<div class="section">
<p>The ZooKeeper Java and C API were extended with getConfig and reconfig
commands that facilitate reconfiguration. Both commands have a synchronous
(blocking) variant and an asynchronous one. We demonstrate these commands
here using the Java CLI, but note that you can similarly use the C CLI or
invoke the commands directly from a program just like any other ZooKeeper
command.</p>
<a name="sc_reconfig_retrieving"></a>
<h3 class="h4">Retrieving the current dynamic configuration</h3>
<p>The dynamic configuration is stored in a special znode
ZooDefs.CONFIG_NODE = /zookeeper/config. The new
<span class="codefrag command">config</span> CLI command reads this znode (currently it is
simply a wrapper to <span class="codefrag command">get /zookeeper/config</span>). As with
normal reads, to retrieve the latest committed value you should do a
<span class="codefrag command">sync</span> first.</p>
<pre class="code">[zk: 127.0.0.1:2791(CONNECTED) 3] config
server.1=localhost:2780:2783:participant;localhost:2791
server.2=localhost:2781:2784:participant;localhost:2792
server.3=localhost:2782:2785:participant;localhost:2793
<strong>version=400000003</strong>
</pre>
<p>Notice the last line of the output. This is the configuration
version. The version equals to the zxid of the reconfiguration command
which created this configuration. The version of the first established
configuration equals to the zxid of the NEWLEADER message sent by the
first successfully established leader. When a configuration is written
to a dynamic configuration file, the version automatically becomes part
of the filename and the static configuration file is updated with the
path to the new dynamic configuration file. Configuration files
corresponding to earlier versions are retained for backup
purposes.</p>
<p>During boot time the version (if it exists) is extracted from the
filename. The version should never be altered manually by users or the
system administrator. It is used by the system to know which
configuration is most up-to-date. Manipulating it manually can result in
data loss and inconsistency.</p>
<p>Just like a <span class="codefrag command">get</span> command, the
<span class="codefrag command">config</span> CLI command accepts the <span class="codefrag option">-w</span>
flag for setting a watch on the znode, and <span class="codefrag option">-s</span> flag for
displaying the Stats of the znode. It additionally accepts a new flag
<span class="codefrag option">-c</span> which outputs only the version and the client
connection string corresponding to the current configuration. For
example, for the configuration above we would get:</p>
<pre class="code">[zk: 127.0.0.1:2791(CONNECTED) 17] config -c
400000003 localhost:2791,localhost:2793,localhost:2792</pre>
<p>Note that when using the API directly, this command is called
<span class="codefrag command">getConfig</span>.</p>
<p>As any read command it returns the configuration known to the
follower to which your client is connected, which may be slightly
out-of-date. One can use the <span class="codefrag command">sync</span> command for
stronger guarantees. For example using the Java API:</p>
<pre class="code">zk.sync(ZooDefs.CONFIG_NODE, void_callback, context);
zk.getConfig(watcher, callback, context);</pre>
<p>Note: in 3.5.0 it doesn't really matter which path is passed to the
<span class="codefrag command">sync() </span> command as all the server's state is brought
up to date with the leader (so one could use a different path instead of
ZooDefs.CONFIG_NODE). However, this may change in the future.</p>
<a name="sc_reconfig_modifying"></a>
<h3 class="h4">Modifying the current dynamic configuration</h3>
<p>Modifying the configuration is done through the
<span class="codefrag command">reconfig</span> command. There are two modes of
reconfiguration: incremental and non-incremental (bulk). The
non-incremental simply specifies the new dynamic configuration of the
system. The incremental specifies changes to the current configuration.
The <span class="codefrag command">reconfig</span> command returns the new
configuration.</p>
<p>A few examples are in: <span class="codefrag filename">ReconfigTest.java</span>,
<span class="codefrag filename">ReconfigRecoveryTest.java</span> and
<span class="codefrag filename">TestReconfigServer.cc</span>.</p>
<a name="sc_reconfig_general"></a>
<h4>General</h4>
<p>
<strong>Removing servers:</strong> Any server can
be removed, including the leader (although removing the leader will
result in a short unavailability, see Figures 6 and 8 in the <a href="https://www.usenix.org/conference/usenixfederatedconferencesweek/dynamic-recon%EF%AC%81guration-primarybackup-clusters">paper</a>). The server will not be shut-down automatically.
Instead, it becomes a "non-voting follower". This is somewhat similar
to an observer in that its votes don't count towards the Quorum of
votes necessary to commit operations. However, unlike a non-voting
follower, an observer doesn't actually see any operation proposals and
does not ACK them. Thus a non-voting follower has a more significant
negative effect on system throughput compared to an observer.
Non-voting follower mode should only be used as a temporary mode,
before shutting the server down, or adding it as a follower or as an
observer to the ensemble. We do not shut the server down automatically
for two main reasons. The first reason is that we do not want all the
clients connected to this server to be immediately disconnected,
causing a flood of connection requests to other servers. Instead, it
is better if each client decides when to migrate independently. The
second reason is that removing a server may sometimes (rarely) be
necessary in order to change it from "observer" to "participant" (this
is explained in the section <a href="#sc_reconfig_additional">Additional comments</a>).</p>
<p>Note that the new configuration should have some minimal number of
participants in order to be considered legal. If the proposed change
would leave the cluster with less than 2 participants and standalone
mode is enabled (standaloneEnabled=true, see the section <a href="#sc_reconfig_standaloneEnabled">The standaloneEnabled flag</a>), the reconfig will not be
processed (BadArgumentsException). If standalone mode is disabled
(standaloneEnabled=false) then its legal to remain with 1 or more
participants.</p>
<p>
<strong>Adding servers:</strong> Before a
reconfiguration is invoked, the administrator must make sure that a
quorum (majority) of participants from the new configuration are
already connected and synced with the current leader. To achieve this
we need to connect a new joining server to the leader before it is
officially part of the ensemble. This is done by starting the joining
server using an initial list of servers which is technically not a
legal configuration of the system but (a) contains the joiner, and (b)
gives sufficient information to the joiner in order for it to find and
connect to the current leader. We list a few different options of
doing this safely.</p>
<ol>
<li>
<p>Initial configuration of joiners is comprised of servers in
the last committed configuration and one or more joiners, where
<strong>joiners are listed as observers.</strong>
For example, if servers D and E are added at the same time to (A,
B, C) and server C is being removed, the initial configuration of
D could be (A, B, C, D) or (A, B, C, D, E), where D and E are
listed as observers. Similarly, the configuration of E could be
(A, B, C, E) or (A, B, C, D, E), where D and E are listed as
observers. <strong>Note that listing the joiners as
observers will not actually make them observers - it will only
prevent them from accidentally forming a quorum with other
joiners.</strong> Instead, they will contact the servers in the
current configuration and adopt the last committed configuration
(A, B, C), where the joiners are absent. Configuration files of
joiners are backed up and replaced automatically as this happens.
After connecting to the current leader, joiners become non-voting
followers until the system is reconfigured and they are added to
the ensemble (as participant or observer, as appropriate).</p>
</li>
<li>
<p>Initial configuration of each joiner is comprised of servers
in the last committed configuration + <strong>the
joiner itself, listed as a participant.</strong> For example, to
add a new server D to a configuration consisting of servers (A, B,
C), the administrator can start D using an initial configuration
file consisting of servers (A, B, C, D). If both D and E are added
at the same time to (A, B, C), the initial configuration of D
could be (A, B, C, D) and the configuration of E could be (A, B,
C, E). Similarly, if D is added and C is removed at the same time,
the initial configuration of D could be (A, B, C, D). Never list
more than one joiner as participant in the initial configuration
(see warning below).</p>
</li>
<li>
<p>Whether listing the joiner as an observer or as participant,
it is also fine not to list all the current configuration servers,
as long as the current leader is in the list. For example, when
adding D we could start D with a configuration file consisting of
just (A, D) if A is the current leader. however this is more
fragile since if A fails before D officially joins the ensemble, D
doesn&rsquo;t know anyone else and therefore the administrator will have
to intervene and restart D with another server list.</p>
</li>
</ol>
<div class="note">
<div class="label">Warning</div>
<div class="content">
<title>Warning</title>
<p>Never specify more than one joining server in the same initial
configuration as participants. Currently, the joining servers don&rsquo;t
know that they are joining an existing ensemble; if multiple joiners
are listed as participants they may form an independent quorum
creating a split-brain situation such as processing operations
independently from your main ensemble. It is OK to list multiple
joiners as observers in an initial config.</p>
</div>
</div>
<p>Finally, note that once connected to the leader, a joiner adopts
the last committed configuration, in which it is absent (the initial
config of the joiner is backed up before being rewritten). If the
joiner restarts in this state, it will not be able to boot since it is
absent from its configuration file. In order to start it you&rsquo;ll once
again have to specify an initial configuration.</p>
<p>
<strong>Modifying server parameters:</strong> One
can modify any of the ports of a server, or its role
(participant/observer) by adding it to the ensemble with different
parameters. This works in both the incremental and the bulk
reconfiguration modes. It is not necessary to remove the server and
then add it back; just specify the new parameters as if the server is
not yet in the system. The server will detect the configuration change
and perform the necessary adjustments. See an example in the section
<a href="#sc_reconfig_incremental">Incremental mode</a> and an exception to this
rule in the section <a href="#sc_reconfig_additional">Additional comments</a>.</p>
<p>It is also possible to change the Quorum System used by the
ensemble (for example, change the Majority Quorum System to a
Hierarchical Quorum System on the fly). This, however, is only allowed
using the bulk (non-incremental) reconfiguration mode. In general,
incremental reconfiguration only works with the Majority Quorum
System. Bulk reconfiguration works with both Hierarchical and Majority
Quorum Systems.</p>
<p>
<strong>Performance Impact:</strong> There is
practically no performance impact when removing a follower, since it
is not being automatically shut down (the effect of removal is that
the server's votes are no longer being counted). When adding a server,
there is no leader change and no noticeable performance disruption.
For details and graphs please see Figures 6, 7 and 8 in the <a href="https://www.usenix.org/conference/usenixfederatedconferencesweek/dynamic-recon%EF%AC%81guration-primarybackup-clusters">paper</a>.</p>
<p>The most significant disruption will happen when a leader change
is caused, in one of the following cases:</p>
<ol>
<li>
<p>Leader is removed from the ensemble.</p>
</li>
<li>
<p>Leader's role is changed from participant to observer.</p>
</li>
<li>
<p>The port used by the leader to send transactions to others
(quorum port) is modified.</p>
</li>
</ol>
<p>In these cases we perform a leader hand-off where the old leader
nominates a new leader. The resulting unavailability is usually
shorter than when a leader crashes since detecting leader failure is
unnecessary and electing a new leader can usually be avoided during a
hand-off (see Figures 6 and 8 in the <a href="https://www.usenix.org/conference/usenixfederatedconferencesweek/dynamic-recon%EF%AC%81guration-primarybackup-clusters">paper</a>).</p>
<p>When the client port of a server is modified, it does not drop
existing client connections. New connections to the server will have
to use the new client port.</p>
<p>
<strong>Progress guarantees:</strong> Up to the
invocation of the reconfig operation, a quorum of the old
configuration is required to be available and connected for ZooKeeper
to be able to make progress. Once reconfig is invoked, a quorum of
both the old and of the new configurations must be available. The
final transition happens once (a) the new configuration is activated,
and (b) all operations scheduled before the new configuration is
activated by the leader are committed. Once (a) and (b) happen, only a
quorum of the new configuration is required. Note, however, that
neither (a) nor (b) are visible to a client. Specifically, when a
reconfiguration operation commits, it only means that an activation
message was sent out by the leader. It does not necessarily mean that
a quorum of the new configuration got this message (which is required
in order to activate it) or that (b) has happened. If one wants to
make sure that both (a) and (b) has already occurred (for example, in
order to know that it is safe to shut down old servers that were
removed), one can simply invoke an update
(<span class="codefrag command">set-data</span>, or some other quorum operation, but not
a <span class="codefrag command">sync</span>) and wait for it to commit. An alternative
way to achieve this was to introduce another round to the
reconfiguration protocol (which, for simplicity and compatibility with
Zab, we decided to avoid).</p>
<a name="sc_reconfig_incremental"></a>
<h4>Incremental mode</h4>
<p>The incremental mode allows adding and removing servers to the
current configuration. Multiple changes are allowed. For
example:</p>
<p>
<span class="codefrag userinput">&gt; reconfig -remove 3 -add
server.5=125.23.63.23:1234:1235;1236</span>
</p>
<p>Both the add and the remove options get a list of comma separated
arguments (no spaces):</p>
<p>
<span class="codefrag userinput">&gt; reconfig -remove 3,4 -add
server.5=localhost:2111:2112;2113,6=localhost:2114:2115:observer;2116</span>
</p>
<p>The format of the server statement is exactly the same as
described in the section <a href="#sc_reconfig_clientport">Specifying the client port</a> and
includes the client port. Notice that here instead of "server.5=" you
can just say "5=". In the example above, if server 5 is already in the
system, but has different ports or is not an observer, it is updated
and once the configuration commits becomes an observer and starts
using these new ports. This is an easy way to turn participants into
observers and vise versa or change any of their ports, without
rebooting the server.</p>
<p>ZooKeeper supports two types of Quorum Systems &ndash; the simple
Majority system (where the leader commits operations after receiving
ACKs from a majority of voters) and a more complex Hierarchical
system, where votes of different servers have different weights and
servers are divided into voting groups. Currently, incremental
reconfiguration is allowed only if the last proposed configuration
known to the leader uses a Majority Quorum System
(BadArgumentsException is thrown otherwise).</p>
<p>Incremental mode - examples using the Java API:</p>
<pre class="code">List&lt;String&gt; leavingServers = new ArrayList&lt;String&gt;();
leavingServers.add("1");
leavingServers.add("2");
byte[] config = zk.reconfig(null, leavingServers, null, -1, new Stat());</pre>
<pre class="code">List&lt;String&gt; leavingServers = new ArrayList&lt;String&gt;();
List&lt;String&gt; joiningServers = new ArrayList&lt;String&gt;();
leavingServers.add("1");
joiningServers.add("server.4=localhost:1234:1235;1236");
byte[] config = zk.reconfig(joiningServers, leavingServers, null, -1, new Stat());
String configStr = new String(config);
System.out.println(configStr);</pre>
<p>There is also an asynchronous API, and an API accepting comma
separated Strings instead of List&lt;String&gt;. See
src/java/main/org/apache/zookeeper/ZooKeeper.java.</p>
<a name="sc_reconfig_nonincremental"></a>
<h4>Non-incremental mode</h4>
<p>The second mode of reconfiguration is non-incremental, whereby a
client gives a complete specification of the new dynamic system
configuration. The new configuration can either be given in place or
read from a file:</p>
<p>
<span class="codefrag userinput">&gt; reconfig -file newconfig.cfg
</span>//newconfig.cfg is a dynamic config file, see <a href="#sc_reconfig_file">Dynamic configuration file</a>
</p>
<p>
<span class="codefrag userinput">&gt; reconfig -members
server.1=125.23.63.23:2780:2783:participant;2791,server.2=125.23.63.24:2781:2784:participant;2792,server.3=125.23.63.25:2782:2785:participant;2793</span>
</p>
<p>The new configuration may use a different Quorum System. For
example, you may specify a Hierarchical Quorum System even if the
current ensemble uses a Majority Quorum System.</p>
<p>Bulk mode - example using the Java API:</p>
<pre class="code">ArrayList&lt;String&gt; newMembers = new ArrayList&lt;String&gt;();
newMembers.add("server.1=1111:1234:1235;1236");
newMembers.add("server.2=1112:1237:1238;1239");
newMembers.add("server.3=1114:1240:1241:observer;1242");
byte[] config = zk.reconfig(null, null, newMembers, -1, new Stat());
String configStr = new String(config);
System.out.println(configStr);</pre>
<p>There is also an asynchronous API, and an API accepting comma
separated String containing the new members instead of
List&lt;String&gt;. See
src/java/main/org/apache/zookeeper/ZooKeeper.java.</p>
<a name="sc_reconfig_conditional"></a>
<h4>Conditional reconfig</h4>
<p>Sometimes (especially in non-incremental mode) a new proposed
configuration depends on what the client "believes" to be the current
configuration, and should be applied only to that configuration.
Specifically, the <span class="codefrag command">reconfig</span> succeeds only if the
last configuration at the leader has the specified version.</p>
<p>
<span class="codefrag userinput">&gt; reconfig -file &lt;filename&gt; -v &lt;version&gt;</span>
</p>
<p>In the previously listed Java examples, instead of -1 one could
specify a configuration version to condition the
reconfiguration.</p>
<a name="sc_reconfig_errors"></a>
<h4>Error conditions</h4>
<p>In addition to normal ZooKeeper error conditions, a
reconfiguration may fail for the following reasons:</p>
<ol>
<li>
<p>another reconfig is currently in progress
(ReconfigInProgress)</p>
</li>
<li>
<p>the proposed change would leave the cluster with less than 2
participants, in case standalone mode is enabled, or, if
standalone mode is disabled then its legal to remain with 1 or
more participants (BadArgumentsException)</p>
</li>
<li>
<p>no quorum of the new configuration was connected and
up-to-date with the leader when the reconfiguration processing
began (NewConfigNoQuorum)</p>
</li>
<li>
<p>
<span class="codefrag userinput">-v x</span> was specified, but the version
<span class="codefrag userinput">y</span> of the latest configuration is not
<span class="codefrag userinput">x</span> (BadVersionException)</p>
</li>
<li>
<p>an incremental reconfiguration was requested but the last
configuration at the leader uses a Quorum System which is
different from the Majority system (BadArgumentsException)</p>
</li>
<li>
<p>syntax error (BadArgumentsException)</p>
</li>
<li>
<p>I/O exception when reading the configuration from a file
(BadArgumentsException)</p>
</li>
</ol>
<p>Most of these are illustrated by test-cases in
<span class="codefrag filename">ReconfigFailureCases.java</span>.</p>
<a name="sc_reconfig_additional"></a>
<h4>Additional comments</h4>
<p>
<strong>Liveness:</strong> To better understand
the difference between incremental and non-incremental
reconfiguration, suppose that client C1 adds server D to the system
while a different client C2 adds server E. With the non-incremental
mode, each client would first invoke <span class="codefrag command">config</span> to find
out the current configuration, and then locally create a new list of
servers by adding its own suggested server. The new configuration can
then be submitted using the non-incremental
<span class="codefrag command">reconfig</span> command. After both reconfigurations
complete, only one of E or D will be added (not both), depending on
which client's request arrives second to the leader, overwriting the
previous configuration. The other client can repeat the process until
its change takes effect. This method guarantees system-wide progress
(i.e., for one of the clients), but does not ensure that every client
succeeds. To have more control C2 may request to only execute the
reconfiguration in case the version of the current configuration
hasn't changed, as explained in the section <a href="#sc_reconfig_conditional">Conditional reconfig</a>. In this way it may avoid blindly
overwriting the configuration of C1 if C1's configuration reached the
leader first.</p>
<p>With incremental reconfiguration, both changes will take effect as
they are simply applied by the leader one after the other to the
current configuration, whatever that is (assuming that the second
reconfig request reaches the leader after it sends a commit message
for the first reconfig request -- currently the leader will refuse to
propose a reconfiguration if another one is already pending). Since
both clients are guaranteed to make progress, this method guarantees
stronger liveness. In practice, multiple concurrent reconfigurations
are probably rare. Non-incremental reconfiguration is currently the
only way to dynamically change the Quorum System. Incremental
configuration is currently only allowed with the Majority Quorum
System.</p>
<p>
<strong>Changing an observer into a
follower:</strong> Clearly, changing a server that participates in
voting into an observer may fail if error (2) occurs, i.e., if fewer
than the minimal allowed number of participants would remain. However,
converting an observer into a participant may sometimes fail for a
more subtle reason: Suppose, for example, that the current
configuration is (A, B, C, D), where A is the leader, B and C are
followers and D is an observer. In addition, suppose that B has
crashed. If a reconfiguration is submitted where D is said to become a
follower, it will fail with error (3) since in this configuration, a
majority of voters in the new configuration (any 3 voters), must be
connected and up-to-date with the leader. An observer cannot
acknowledge the history prefix sent during reconfiguration, and
therefore it does not count towards these 3 required servers and the
reconfiguration will be aborted. In case this happens, a client can
achieve the same task by two reconfig commands: first invoke a
reconfig to remove D from the configuration and then invoke a second
command to add it back as a participant (follower). During the
intermediate state D is a non-voting follower and can ACK the state
transfer performed during the second reconfig comand.</p>
</div>
<a name="ch_reconfig_rebalancing"></a>
<h2 class="h3">Rebalancing Client Connections</h2>
<div class="section">
<p>When a ZooKeeper cluster is started, if each client is given the same
connection string (list of servers), the client will randomly choose a
server in the list to connect to, which makes the expected number of
client connections per server the same for each of the servers. We
implemented a method that preserves this property when the set of servers
changes through reconfiguration. See Sections 4 and 5.1 in the <a href="https://www.usenix.org/conference/usenixfederatedconferencesweek/dynamic-recon%EF%AC%81guration-primarybackup-clusters">paper</a>.</p>
<p>In order for the method to work, all clients must subscribe to
configuration changes (by setting a watch on /zookeeper/config either
directly or through the <span class="codefrag command">getConfig</span> API command). When
the watch is triggered, the client should read the new configuration by
invoking <span class="codefrag command">sync</span> and <span class="codefrag command">getConfig</span> and if
the configuration is indeed new invoke the
<span class="codefrag command">updateServerList</span> API command. To avoid mass client
migration at the same time, it is better to have each client sleep a
random short period of time before invoking
<span class="codefrag command">updateServerList</span>.</p>
<p>A few examples can be found in:
<span class="codefrag filename">StaticHostProviderTest.java</span> and
<span class="codefrag filename">TestReconfig.cc</span>
</p>
<p>Example (this is not a recipe, but a simplified example just to
explain the general idea):</p>
<pre class="code">
public void process(WatchedEvent event) {
synchronized (this) {
if (event.getType() == EventType.None) {
connected = (event.getState() == KeeperState.SyncConnected);
notifyAll();
} else if (event.getPath()!=null &amp;&amp; event.getPath().equals(ZooDefs.CONFIG_NODE)) {
// in prod code never block the event thread!
zk.sync(ZooDefs.CONFIG_NODE, this, null);
zk.getConfig(this, this, null);
}
}
}
public void processResult(int rc, String path, Object ctx, byte[] data, Stat stat) {
if (path!=null &amp;&amp; path.equals(ZooDefs.CONFIG_NODE)) {
String config[] = ConfigUtils.getClientConfigStr(new String(data)).split(" "); // similar to config -c
long version = Long.parseLong(config[0], 16);
if (this.configVersion == null){
this.configVersion = version;
} else if (version &gt; this.configVersion) {
hostList = config[1];
try {
// the following command is not blocking but may cause the client to close the socket and
// migrate to a different server. In practice its better to wait a short period of time, chosen
// randomly, so that different clients migrate at different times
zk.updateServerList(hostList);
} catch (IOException e) {
System.err.println("Error updating server list");
e.printStackTrace();
}
this.configVersion = version;
} } }</pre>
</div>
<p align="right">
<font size="-2"></font>
</p>
</div>
<!--+
|end content
+-->
<div class="clearboth">&nbsp;</div>
</div>
<div id="footer">
<!--+
|start bottomstrip
+-->
<div class="lastmodified">
<script type="text/javascript"><!--
document.write("Last Published: " + document.lastModified);
// --></script>
</div>
<div class="copyright">
Copyright &copy;
2008-2013 <a href="http://www.apache.org/licenses/">The Apache Software Foundation.</a>
</div>
<!--+
|end bottomstrip
+-->
</div>
</body>
</html>