har: a High-Availability Resolver

Overview

har is a "high availability resolver." As you probably know, a DNS resolver's duty is to convert host names (e.g., api.example.com) into IPv4 or IPv6 addresses (e.g., 1.2.3.4).

Unlike bind, djbdns, PowerDNS, or other DNS resolvers, har is not a full-service resolver authoritative for top-level or second-level domains (e.g. mycompany.com). har is not designed to accept connections from other hosts, nor does har provide features such as caching or recursion. Instead, har is designed to resolve hostnames that you configure into IP addresses of hosts that are actually available, based on criteria you specify. Queries for non-configured hostnames are simply forwarded to a full-service resolver.

Why use har?

har is intended to be a building block for large-scale, highly-available services. For example, if your service architecture involves a front-end web server that retrieves data from an API layer, har can help ensure that the front-end code connects to the closest available API server.

har provides the following benefits over HTTP proxies:

Most HTTP proxies don't maintain connection and response latency statistics. har helps direct clients to both an available and proximate server. This makes it easy to implement a design in which a client falls back to a server in another region (whose response latency necessarily will be longer) only if all of the servers in the same region are unavailable.
Proxied connections usually disguise the connection's origin address (from the server's point of view) and the destination address (from the client's point of view). This can present a challenge for debugging service issues. With har, you don't need to worry about that: clients can directly connect to the server's address.
har helps isolate service issues: since every client runs its own copy, the failure of a har instance only affects the client.
har improves speed: no extra network hop is needed that could introduce additional response latency.

Design

hard (the har daemon) is a DNS proxy server. It answers DNS queries on the IPv4 and IPv6 loopback interfaces by looking at the "question" section of the request packet. If the request is for an A or AAAA record (i.e., an IP address), and the subject of the request (i.e., the hostname) is configured in har's configuration file, har will consider itself the authoritative server and respond to the request Otherwise, har will simply forward the request to an authoritative DNS server.

While har waits for queries, it periodically polls a list of one or more candidates associated with each configured host. A candidate is a specific host whose IP address could be associated with the configured hostname. For example, candidates for the host api.example.com could be api1.east.example.com, api2.east.example.com, api1.west.example.com, and so forth. Candidates are polled and ranked either via HTTP or TCP health checks. Check state is stored in a local sqlite3 database.

+-----------+
|  program  |
+-----------+
     ||
     || <- DNS query: "api.example.com" 
     ||    (localhost, port 53)
     \/
+----------+
|          |       ---------------          ----------
|          | ---> (har config file)   |--> (sqlite3 db)
|          |       ---------------    |     ----------
|   hard   | --------------------------
|          |
|          | == HTTP health check ==> http://api1.west.example.com
|          | == HTTP health check ==> http://api1.east.example.com
|          | == HTTP health check ==> ...
+----------+
     ||
     ||  <- outbound DNS queries ("api1.west.example.com")
     \/
+-------------------------------------+
| upstream resolver (ns1.example.com) | 
+-------------------------------------+

Requirements

har requires the following shared libraries to be installed on the client:

At this time, har has been tested only on CentOS 5 Linux (x86_64 platform).
Support for additional operating systems and distributions is forthcoming -- any assistance in this area would be most appreciated.

Building har

gcc -std=c99 -D_POSIX_SOURCE -Wall -pedantic -g har.c -o har -lev -lldns -lconfuse -lcurl -luriparser -lsqlite3

Configuration

TBD

Authors

Michael S. Fischer michael+har@dynamine.net is the primary author and maintainer.

Support

Contact the author for support questions. Since har is a volunteer project, support is available only as the author's time permits.

Bug reports and feature requests can be filed at https://github.com/otterley/har/issues.

Copyright and License

See the LICENSE file included with this distribution for the terms under which har is licensed.

Source and contributions

Contributions to har are welcome and encouraged. The official GitHub repository is located at https://github.com/otterley/har.

vshankar / har