A healthchecker for Anycasted services.
Contents
anycast-healthchecker monitors a service by doing periodic health checks and based on the result instructs Bird daemon to either advertise or withdraw the route to reach the monitored service. As a result Bird will only advertise routes for healthy services.
Bird must be configured in a certain way to interface properly with anycast-healthchecker. The configuration is detailed later in this document.
anycast-healthchecker is a Python program, which uses the daemon library to implement a well-behaved Unix daemon process and threading to run multiple service checks in parallel.
Anycast is a network address scheme where traffic from a sender has more than one potential receivers, but only one of them receives it. Routing protocols, decide which one of the potential receivers will actually receive traffic based on the topology of the network. The main attribute which contributes to the decision is the cost of the network path between a sender and a receiver.
Cost is a protocol specific (usually integer) value that only has meaning within a protocol that is used as a metric of distance. Routing protocols provide default values for common topologies (BGP associates the cost of a path with the number of autonomous systems between the sender and the receiver, OSPF calculates the default cost based on the bandwidth of links), but its main use is to allow administrative control over traffic flow by specifying cost according to business needs.
The nearest receiver to a sender always receives the traffic and this only changes if something changes on the network, another receiver with best path to the sender appears or current receiver disappears. If multiple receivers have the same distance from the sender more than one of the receivers might receive traffic, based on further details of the network configuration.
The three drawings below exhibit how traffic is routed between a sender and multiple potential receivers when something changes on network. In this example BGP routing protocol is used:
These potential receivers use BGP or OSPF by running an Internet Routing daemon, Bird or Quagga, to simultaneously announce the same destination IP address from different places on the network. Due to the nature of Anycast receivers can be located on any network across a global network infrastructure.
Anycast doesn't balance traffic as only one receiver attracts traffic from senders. For instance, if there are two receivers which announce the same destination IP address in certain location, traffic will be distributed across those two receivers unevenly as senders can be spread across the network in an uneven way.
Anycast is being used as a mechanism to switch traffic between and within data-centers for the following main reasons:
- the switch of traffic occurs without the need to enforce a change to clients
In case of loss of a service in one location, traffic to that location will be switched to another data-center without any manual intervention and most importantly without pushing a change to clients which you don't always control.
- the switch happens within few milliseconds
The same technology can be used for balancing traffic using Equal-Cost Multi-Path.
ECMP routing is a network technology where traffic can be routed over multiple paths. In the context of routing protocols, path is the route a packet has to take in order to be delivered to a destination. Because these multiple paths have the same cost, traffic is balanced across them.
This provides the possibility to perform load-balancing of traffic across multiple servers. Routers are the devices which perform load-balancing of traffic and most of them use a deterministic way to select the server based on the following four properties of IP packets:
- source IP
- source PORT
- destination IP
- destination PORT
Each unique combination of values for those four properties is called network flow. For each different network flow a different destination server is selected so traffic is evenly balanced across all servers. These servers run an Internet Routing daemon in the same way as with Anycast case but with the major difference that all servers receive traffic.
The main characteristic of this type of load-balancing is that is stateless. Router balances traffic to a destination IP address based on the quadruple network flow without the need to understand and inspect protocols above Layer 3. As a result it is very cheap in terms of resources and very fast at the same time. This is commonly advertised as traffic balancing at wire-speed.
anycast-healthchecker can be utilized in Anycast and ECMP environments.
The current release of anycast-healthchecker supports only the Bird daemon which you have to configure in a specific way. Thus, it is mandatory to explain very briefly how Bird handles advertisements for routes.
Bird maintains a routing information base (RIB) and various protocols import/export routes to/from it. The diagram below illustrates how Bird advertises routes for IPs assigned to the loopback interface to the rest of the network using BGP protocol. Bird can also import routes learned via BGP/OSPF protocols, but this part of the routing process is irrelevant to the functionality of anycast-healthchecker.
A route is always associated with a service which runs locally on the box. The Anycasted service is a daemon (HAProxy, Nginx, Bind etc) which processes incoming traffic and listens to an IP (Anycast Service Address) for which a route exists in the RIB and advertised by Bird.
As it is exhibited in the above diagram a route is advertised only when:
- The IP is assigned to the loopback interface.
- direct protocol from Bird imports a route for that IP in the RIB.
- BGP/OSPF protocols export that route from the RIB to a network peer.
The route associated with the Anycasted service must be either advertised or withdrawn based on the health of the service, otherwise traffic will always be routed to the local node regardless of the status of the service.
Bird provides filtering capabilities with the help of a simple programming language. A filter can be used to either accept or reject routes before they are exported from the RIB to the network.
A list of IP prefixes (<IP>/<prefix length>) is stored in a text file. IP prefixes that are not included in the list are filtered-out and they do not get exported from the RIB to the network. The white-list text file is sourced by Bird upon startup, reload and reconfiguration. The following diagram illustrates how this technique works:
This configuration logic allows a separate process to update the list by adding or removing IP prefixes and trigger a reconfiguration of Bird in order to advertise or withdraw routes. anycast-healthchecker is that separate process. It monitors Anycasted services and based on the status of the health checks updates the list of IP prefixes.
Bird does not allow the definition of a list with no elements and when that happens
Bird will emit an error and refuses to start. Because of this anycast-healthchecker
makes sure that there is always an IP prefix in the list, see dummy_ip_prefix
configuration option in Daemon section.
Because anycast-healthchecker is very much tied in with Bird daemon, the configuration of Bird is explained first. Next, the configuration of anycast-healthchecker (including the configuration for the health checks) is covered and finally the options for invoking the program from the command line will be described.
The logic described in How anycast-healthchecker works can be accomplished by configuring:
- an
includestatement to source other configuration files inbird.conf - a function,
match_route, as an export filter for the routing protocol (BGP or OSPF) - a list of IP prefixes for routes which allowed to be exported by Bird
anycast-healthchecker does not install any of the files mentioned in this section.
The most important parts are the lines include "/etc/bird.d/*.conf"; and
export where match_route();. The former statement causes inclusion of files
and the latter forces all routes to pass from the match_route function before
they are exported. BGP protocol is used in the below bird configuration example
but OSPF protocol can be used as well:
include "/etc/bird.d/*.conf";
protocol device {
scan time 10;
}
protocol direct direct1 {
interface "lo";
export none;
import all;
}
template bgp bgp_peers {
import none;
export where match_route();
local as 64815;
}
protocol bgp BGP1 from bgp_peers {
disabled no;
neighbor 10.248.7.254 as 64814;
}
match-route.conf file configures the match_route function which performs
the white and black listing of IP prefixes by looking up the IP prefix of the
route in a list and accepts the export if it finds a matching entry:
function match_route()
{
return net ~ ACAST_PS_ADVERTISE;
}
anycast-prefixes.conf file defines a list of IP prefixes which is stored in
a variable with the name ACAST_PS_ADVERTISE. The name of the variable can
be anything meaningful but bird_variable setting must be changed
accordingly in order for anycast-healthchecker to modify it.
anycast-prefixes.conf file is not part of the anycast-healthchecker
package, thus administrators need install an initial version with the following
content and after the launch of anycast-healthchecker it should not
modified by any other process(es) as its content is managed by
anycast-healthchecker.
define ACAST_PS_ADVERTISE =
[
10.189.200.255/32
];
anycast-healthchecker daemon removes IP prefixes from the list for which a
service check is not configured. But, the IP prefix set in dummy_ip_prefix
does not need a service check configuration.
anycast-healthchecker uses the popular INI format for its configuration files. This is an example configuration file for the daemon (/etc/anycast-healthchecker.conf):
[DEFAULT] interface = lo [daemon] pidfile = /var/run/anycast-healthchecker/anycast-healthchecker.pid bird_conf = /etc/bird.d/anycast-prefixes.conf bird_variable = ACAST_PS_ADVERTISE bird_reconfigure_cmd = sudo /usr/sbin/birdc configure loglevel = debug log_maxbytes = 104857600 log_backups = 8 log_file = /var/log/anycast-healthchecker/anycast-healthchecker.log stderr_file = /var/log/anycast-healthchecker/stderr.log stdout_file = /var/log/anycast-healthchecker/stdout.log dummy_ip_prefix = 10.189.200.255/32 purge_ip_prefixes = false keep_bird_changes = false changes_counter = 128
Above settings are used as defaults when daemon is launched without a
configuration file. The daemon does not need to run as root as long as it
has sufficient privileges to modify the Bird configuration (anycast-prefixes.conf)
and trigger a reconfiguration of Bird by running birdc configure.
In the above example sudo is used for that purpose (sudoers file has
been modified for that purpose).
Below are the default settings for all service checks, see Configuring checks for services for an explanation of the parameters. Settings in this section can be overwritten in other sections.
| interface: | lo |
|---|---|
| check_interval: | 10 |
| check_timeout: | 2 |
| check_rise: | 2 |
| check_fail: | 2 |
| check_disabled: | true |
| on_disable: | withdraw |
| ip_check_disabled: | false |
Settings for anycast-healthchecker daemon
- pidfile Defaults to /var/run/anycast-healthchecker/anycast-healthchecker.pid
File to store the process id of the daemon
- bird_conf Defaults to /etc/bird.d/anycast-prefixes.conf
File with the list of IP prefixes allowed to be exported. If this file is a symbolic link then the destination and the link itself must be on the same mounted filesystem.
- bird_variable Defaults to ACAST_PS_ADVERTISE
The name of the list defined in bird_conf
- bird_reconfigure_cmd Defaults to sudo /usr/sbin/birdc configure
Command to trigger a reconfiguration of Bird daemon
- loglevel Defaults to debug
Log level to use, possible values are: debug, info, warning, error, critical
- log_file Defaults to /var/log/anycast-healthchecker/anycast-healthchecker.log
File to log messages to
- log_maxbytes Defaults to 104857600 (bytes)
Maximum size in bytes for log files
- log_backups Defaults to 8
Number of old log files to maintain
- stderr_file Defaults to /var/log/anycast-healthchecker/stderr.log
File to redirect standard error to
- stdout_file Defaults to /var/log/anycast-healthchecker/stdout.log
File to redirect standard output to
- dummy_ip_prefix Defaults to 10.189.200.255/32
An IP prefix in the form <IP>/<prefix length> which will be always available in
the list defined by bird_variable to avoid having an empty list.
The dummy_ip_prefix must not be used by any service or assigned to the
interface set with interface or configured anywhere on the network as
anycast-healthchecker does not perform any checks for it.
- purge_ip_prefixes Defaults to false
Purge IP-Prefixes from bird_conf on start-up which don't have a service check associated with them.
NOTE: These IP-Prefixes are always removed from bird_conf when
bird_conf is updated during the life time of the daemon.
purge_ip_prefixes is considered only during start-up and was introduced in
order to be compatible with previous behavior, which didn't remove those
IP-Prefixes on start-up.
- keep_bird_changes Defaults to false
Keep a history of changes for bird_conf file by copying it to a directory.
During the startup of the daemon a directory with the name history is
created under the directory where bird_conf file resides. The daemon has to
have sufficient privileges to create that directory.
- changes_counter Defaults to 128
How many bird_conf files to keep in the history directory.
anycast-healthchecker daemon can be configured to send logging messages over HTTP to a central place in addition to write them to log files. It builds a JSON blob with a specific data structure, which is not configurable at the moment.
The following settings can be added to the [daemon] section for enabling JSON logging.
- json_logging Defaults to false
true enables JSON logging false disables it
- http_server Unset by default
Server name to send JSON logging over HTTP protocol
- http_server_port Unset by default
Port to connect
- http_server_protocol Unset by default
HTTP protocol to use, either http or https
- http_server_timeout Unset by default
How long to wait for the server to send data before giving up, as a float number.
JSON messages are send using http POST requests which are executed in blocking
mode which means that possible long delays will make the health checks to be
delayed as well.
http_server_timeout accepts floating point numbers as values which are
passed to underlying request module as a single timeout which will be applied
to both the connect and the read timeouts.
The configuration for a single service check is defined in one section. Here is an example:
[foo.bar.com] check_cmd = /usr/bin/curl --fail --silent http://10.52.12.1/ check_interval = 10 check_timeout = 2 check_fail = 2 check_rise = 2 check_disabled = false on_disabled = withdraw ip_prefix = 10.52.12.1/32
The name of the section becomes the name of the service check and appears in the log files for easier searching of error/warning messages.
- check_cmd Unset by default
The command to run to determine the status of the service based on the return code. Complex health checking should be wrapped in a script. When check command fails, the stdout and stderr appears in the log file.
- check_interval Defaults to 2 (seconds)
How often to run the check
- check_timeout Defaults to 2 (seconds)
Maximum time in seconds for the check command to complete. anycast-healthchecker will try kill the check if it doesn't return after check_timeout seconds. If check_cmd runs under another user account (root) via sudo then it wouldn't be killed. anycast-healthchecker could run as root to overcome this problem, but it is highly recommended to run it as normal user.
- check_fail Defaults to 2
A service is considered DOWN after this many consecutive unsuccessful health checks
- check_rise Defaults to 2
A service is considered HEALTHY after this many consecutive successful health checks
- check_disabled Defaults to false
true disables the check, false enables it
- on_disable Defaults to withdraw
What to do when check is disabled, either withdraw or advertise
- ip_prefix Unset by default
IP prefix associated with the service. It must be assigned to the
interface set in interface parameter unless ip_check_disabled is set to
true
- ip_check_disabled Defaults to false
true disables the assignment check of ip_prefix to the interface set in
interface, false enables it.
If the check_cmd checks the availability of the service by sending a
request to the Anycasted IP address then that request may be served by another
node which advertises the same IP address on the network. This usually happens
when the Anycasted IP address is not assigned to loopback or any other
interface on the local node.
Therefore, it should be only enabled in environments where the network or the
network configuration of the local node prevents the request from check_cmd
to be forwarded to another node.
- interface Defaults to lo
The name of the interface that ip_prefix is assigned to
Multiple sections may be combined in one file or provide one file per section.
File must be stored under one directory and their name should use .conf
as suffix (foo.bar.com.conf).
Daemon CLI usage:
anycast-healthchecker --help
A simple healthchecker for Anycasted services.
Usage:
anycast-healthchecker [-f <file> -d <directory> -c ] [-p | -P]
Options:
-f, --file <file> configuration file with settings for the daemon
[default: /etc/anycast-healthchecker.conf]
-d, --dir <dir> directory with configuration files for service checks
[default: /etc/anycast-healthchecker.d]
-c, --check perform a sanity check on configuration
-p, --print show default settings for daemon and service checks
-P, --print-conf show configuration
-v, --version show version
-h, --help show this screen
The daemon can be launched by supplying a configuration file and a directory with configuration files for service checks:
anycast-healthchecker -f ./anycast-healthchecker.conf -d ./anycast-healthchecker.d
At the root of the project there is System V init and a Systemd unit file for proper integration with OS startup tools.
Under contrib/systemd and contrib/SysVinit directories there are the necessary Unit service and startup files which can be used to start the daemon on boot.
Under contrib/nagios directory there is a nagios plugin to check if daemon is up and if all threads are running.
Use pip:
pip install anycast-healthchecker
From Source:
sudo python setup.py install
Build (source) RPMs:
python setup.py clean --all; python setup.py bdist_rpm
Build a source archive for manual installation:
python setup.py sdist
Bump version in anycast_healthchecker/__init__.py
Commit above change with:
git commit -av -m'RELEASE 0.1.3 version'
Create a signed tag, pbr will use this for the version number:
git tag -s 0.1.3 -m 'bump release'
Create the source distribution archive (the archive will be placed in the dist directory):
python setup.py sdist
pbr will update ChangeLog file and we want to squeeze them to the previous commit thus we run:
git commit -av --amend
Move current tag to the last commit:
git tag -fs 0.1.3 -m 'bump release'
Push changes:
git push;git push --tags
I would love to hear what other people think about anycast_healthchecker and provide feedback. Please post your comments, bug reports and wishes on my issues page.
At the root of the project there is a local_run.sh script which you can use for testing purposes. It does the following:
- Creates the necessary directory structure under $PWD/var to store configuration and log files
- Generates configuration for the daemon and for 2 service checks
- Generates bird configuration(anycast-prefixes.conf)
- Installs anycast-healthchecker with
python3.4 setup.py install, requires python virtualenvironment, use the excellent tool virtualenvwrapper - Assigns 4 IP addresses (10.52.12.[1-4]) to loopback interface
- Checks if bird daemon runs but it does not try to start if it's running
- Starts the daemon as normal user and not as root
Requirements for running local_run.sh and having a workable setup
- python3.4 installation available
- Bird installed and configured as it is mentioned in Bird configuration
- sudo access to run sudo birdc configure
- sudo access to assign IPs on the loopback interface
The following people have contributed to project with feedback, commits and code reviews
- Károly Nagy (@charlesnagy)
- Nick Demou (@ndemou)
- Ralf Ertzinger (@alufu)
Apache 2.0
This program was originally developed for Booking.com. With approval from Booking.com, the code was generalised and published as Open Source on github, for which the author would like to express his gratitude.
Project website: https://github.com/unixsurfer/anycast_healthchecker
Author: Pavlos Parissis <pavlos.parissis@gmail.com>