DNS Parallel Quering library.
The objective of DNSpq is to query multiple DNS servers in parallel when trying to resolve a query. The pq in DNSpq hence stands for Parallel Query. Querying multiple servers in parallel is non-standard and aggressive towards nameservers, so not suited for use other than for a specific use-case, whis is most likely constrained to a local area network and dedicated DNS servers.
DNSpq queries (in parallel) multiple DBS servers and takes the first answer back to the client. But DNSpq is not just doing that, it tries to be smart in what it does, when dealing with multiple servers to make the setup as resilient and robust as possible. The characteristics of DNSpq are:
- Each query is sent to all servers at the same time
- The total waiting time for such a request is 500ms (half a second)
- Each query is retried once, in case of no response within 300ms
- Failure responses are considered no responses, hence additional responses are waited for
- As soon as a successful response is received, that response is returned to the caller
As a consequence of above, broken or unavailable servers responding that way don't cause failure responses per definition, in most cases other responses are positive, such as when a server gets restarted. The client becomes greedy as well, since it asks for more than it is willing to retrieve, but this is all to improve the overal response time in case of server failure or downtime.
DNSpq doesn't have a cache. It only supports A-type queries, and simple
responses to those. The library, which is wrapped in a nss module
(libnss_dnspq.so.2) aborts on any attempt to do something which is not
a simple A-type query, and a simple response to that. This makes it
easy to have the library fallback queries to the normal glibc resolver.
The configuration of DNSpq nss module goes in /etc/resolv-dnspq.conf. This file supports pool-based syntax to allow multiple pools to be setup, and also multiple providers for the same pool. An example config might look like this:
# Syntax of this file is: .<domain> <server>[:<port>] [<server>[:<port>] ...]
# The client picks a random server pair for domain if multiple matches exist.
.my-pool 10.197.182.25:53001 10.197.182.26:53002
.my-pool 10.197.182.25:53002 10.197.182.26:53003
.my-pool 10.197.182.25:53003 10.197.182.26:53001
In this file, a pool called my-pool is defined with three providers,
all using two servers. The pool can be triggered by resolving anything
from the .my-pool domain, e.g. myhost.my-pool. Upon each query,
DNSpq will pick a random provider for the .my-pool pool. Once it has
chosen one pool, it will send the DNS query to all of the servers listed
for that pool to their designated IP address and port numbers. One can
play with this file in many ways to achieve balancing, sharding and
more.
Fabian Groffen
This library 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.