NAME

gdnsd-plugin-weighted - gdnsd plugin implementing "weighted" records

SYNOPSIS

Example plugin config:

  plugins => {
    weighted => {
      multi = false # default
      service_types = up
      up_thresh => 0.5 # default
      corpwww => {
        lb01 = [ lb01.example.org., 99 ]
        lb02 = [ lb02.example.org., 15 ]
        lb03 = [ lb03.example.org., 1 ]
      }
      frontwww6 => {
        service_types = up
        multi = true
        wwwhost01 = [ 2001:db8::123, 4 ]
        wwwhost02 = [ 2001:db8::456, 1 ]
        wwwhost03 = [ 2001:db8::789, 2 ]
      }
      pubwww => {
        service_types = [ web_check, foo ]
        up_thresh => 0.01,
        pubhost01 = [ 192.0.2.1, 44 ]
        pubhost02 = [ 192.0.2.2, 11 ]
        pubhost03 = [ 192.0.2.3, 11 ]
        pubhost04 = [ 192.0.2.4, 11 ]
      }
      cdnwww => {
        service_types = web_check
        datacenter1 => {
          d1-lb1 = [ 127.0.0.1, 2 ]
          d1-lb2 = [ 127.0.0.2, 2 ]
        }
        datacenter2 => {
          d2-lb1 = [ 127.0.0.3, 2 ]
          d2-lb2 = [ 127.0.0.4, 2 ]
          d2-lb3 = [ 127.0.0.5, 1 ]
        }
      }
      mixed => {
        multi => false,
        addrs_v4 => {
          lb1 = [ 127.0.0.3, 2 ]
          lb2 = [ 127.0.0.4, 2 ]
        }
        addrs_v6 => {
          multi => true
          www6set1 = {
            lb01 => [ 2001:db8::123, 4 ]
            lb02 => [ 2001:db8::456, 1 ]
          }
          www6set2 = {
            lb01 => [ 2001:db8::789, 4 ]
            lb02 => [ 2001:db8::ABC, 1 ]
          }
        }
      }
      cn => {
        service_types = my_cn_check
        foo = [ lb01.example.org., 99 ]
        bar = [ lb02.example.org., 15 ]
      }
    }
  }

Zonefile RRs referencing the above, in example.com:

  www.corp   300 DYNC weighted!corpwww
  www6.front 300 DYNA weighted!frontwww6
  www        300 DYNC weighted!pubwww
  cdn        300 DYNA weighted!cdnwww
  mixed-a    300 DYNA weighted!mixed
  cnames     300 DYNC weighted!cn

DESCRIPTION

gdnsd-plugin-weighted can be used to return one (or a subset) of several address records, or one of several CNAME records based on dynamic-weighted probabilities.

CONFIGURATION - TOP LEVEL

At the top level, there are three special parameter keys: "service_types", "up_thresh", and "multi". "multi" is ignored for CNAME-based resources. All of these keys are inherited and override-able at the per-resource and per-address-family levels.

"service_types" sets how the applicable addresses or CNAMEs are monitored. The top-level default "service_types" is "up", which is a built-in service type provided by gdnsd. For more information about configuring non-default service type's, see the main gdnsd.config(5) documentation.

"multi" is a boolean that can be "true" or "false", and defaults to "false". "multi" controls the behavior of the algorithm for selecting result addresses, discussed in detail later.

"up_thresh" defines a floating point fraction of summed address weights in the range "(0.0 - 1.0]", defaulting to 0.5, and is used to influence failure/failover behavior.

Other than those three, the rest of the top level keys are the names of your resources, and their values are the configuration of each resource.

CONFIGURATION - PER-RESOURCE

Inside a given resource's configuration hash, again the three address-related parameters "services_types", "multi", and "up_thresh" may be specified to override their settings per-resource.

There are two basic configuration modes within a resource:

1) Explicit per-family address sub-stanzas. In this mode, the resource contains one or more of the keys "addrs_v4" and "addrs_v6". Usually one would use both together, as it's simpler to use the second option when configuring a single address family.

The contents of each stanza configure response RRs of the given address type for this resource, and the 3 behavioral parameters "service_types", "multi", and "up_thresh" can be overridden per-address-family as well.

2) Automatic top-level detection of just one address family or CNAMEs. In this mode, you can configure the top-level of a resource with direct entries, so long as they are matching set of a single type: all IPv4 addresses, all IPv6 address, or all CNAMEs, and the type will be auto-detected.

Resources which contain weighted lists of CNAMEs rather than addresses can only be used with "DYNC" RRs in zonefiles, whereas those that contain only addresses can be used in either "DYNC" or "DYNA" RRs.

CONFIGURATION - CNAMES

When configuring cnames, the value of each item should be "[ CNAME, WEIGHT ]", and the resource will be useful for "DYNC" zonefile records, resolving to a weighted CNAME record in responses. The selection algorithm based on weights and monitoring results is as documented below for addresses in the THE UNGROUPED SINGLE CASE, since groups of CNAMEs cannot be configured, and the "multi" option is not valid for them.

CNAME data must be fully qualified (ends in "."), and a "DYNC" cannot be used to emit a CNAME that points into the same zone that contained the "DYNC".

CONFIGURATION - ADDRESSES

With the exception that "addrs_v4" and "addrs_v6" must contain only addresses of the correct family (or in the top-level auto-detect case, the top level entries must all be of the same family), the two stanzas behave identically. When both are present, they are both used in every "DYNA" response (as gdnsd always includes opposite-family records in the Additional section of A/AAAA queries).

Within either address family type, there are two different binary dimensions (multi -> true/false, and grouped-vs-ungrouped) upon which the configuration and behavior hinge, leading to four different possible cases: ungrouped-single, ungrouped-multi, grouped-single, and grouped-multi. Each will be discussed in detail below:

THE UNGROUPED SINGLE CASE

This is the simplest case. The code detects this case when it sees that "multi" is false (the default), and that the values of the keys are arrays rather than sub-hashes. Each hash key is an address label, and each value is an array of "[ IPADDR, WEIGHT ]".

When answering a query in this case, first the weights are converted to dynamic weights. The dynamic weight of an address is its configured weight if the monitored state is "UP", or zero if the monitored state is "DOWN". The dynamic weights are summed to produce a dynamic weight total, and then a single address to respond with is chosen from the set, with each address having the odds "addr_dynamic_weight / total_dynamic_weight".

However, if the "total_dynamic_weight" is less than "ceil(up_thresh * total_configured_weight)", then the dynamic weights are all reset to their configured full values so that the response odds are the same as if all were "UP", and resource-level failure is signalled to any upper-layer meta-plugin (e.g. metafo or geoip) when applicable.

Example (X could be a whole resource, or an addrs_v4 stanza):

  X => {
    multi => false # default
    # odds below assume no addresses are down:
    lb01 => [ 192.0.2.1, 45 ] # 25% chance (45/180)
    lb02 => [ 192.0.2.1, 60 ] # 33% chance (60/180)
    lb03 => [ 192.0.2.1, 75 ] # 42% chance (75/180)
  }

THE UNGROUPED MULTI CASE

This case is detected when, (as above) the values of the keys are arrays of "[ IPADDR, WEIGHT]", but the parameter "multi" is true. The change from the above behavior is primarily that multiple addresses from the weighted set can be returned in each response. The "maximum", rather than the sum, of the dynamic weights (again, zero for down addresses, configured-weight otherwise), is found, and the odds of each address's inclusion in the response set is "addr_dyanmic_weight / max_dynamic_weight".

This means all non-"DOWN" addresses which share the group's maximum dynamic weight value will always be included, whereas others will be optionally included depending on the odds. At least one address is always returned (because logically, at least one address has the maximum weight, giving it a 100% chance), and sometimes the full non-"DOWN" set will be returned.

"up_thresh" behaves as in the previous case: If the sum of the dynamic weight values is less than "ceil(up_thresh * total_configured_weight)", then the dynamic weights are all set to their configured values and the result set is calculated as if all were "UP", while signalling resource-level failure to upstream meta-plugins (geoip or metafo).

Example:

  X => {
    multi => true
    # odds below assume no addresses are down:
    lb01 => [ 192.0.2.1, 45 ] # 75% chance (45/60)
    lb02 => [ 192.0.2.1, 60 ] # 100% chance (60/60)
    lb03 => [ 192.0.2.1, 60 ] # 100% chance (60/60)
    # overall possible result-sets:
    # lb01,lb02,lb03 -> 75%
    # lb02,lb03 -> 25%
  }

THE GROUPED SINGLE CASE

The grouped cases are detected when the keys' values are sub-hashes at the outer level rather than arrays of "[ IPADDR, WEIGHT]". In the grouped case, first the set is divided into named groups, and then within each group individual addresses are configured as "addrlabel => [ IPADDR, WEIGHT ]".

Example:

   X => {
     group1 => {
       lb01 => [ 192.0.2.1, 10 ]
       lb02 => [ 192.0.2.1, 20 ]
       lb03 => [ 192.0.2.1, 30 ]
     }
     group2 => {
       lb01 => [ 192.0.2.7, 10 ]
       lb02 => [ 192.0.2.8, 20 ]
       lb03 => [ 192.0.2.9, 30 ]
     }
   }

The grouped single case, of course, occurs when the configuration layout is as shown above, and the "multi" parameter is "false" (the default).

In grouped-single mode, essentially the groups are weighted against each other similarly to the single case for ungrouped addresses, resulting in the choice of a single group from the set of groups. Then the addresses within the chosen group are weighted against each other in multi-style, returning potentially more than one address from the chosen group.

Specifically, each group's odds of being the single group chosen is "group_dyn_weight / total_dyn_weight", where the group's dynamic weight is the sum of the dynamic weights within it ("DOWN" addresses are zero), and the total dynamic weight is the dynamic sum of all groups. Then within each group, the odds of each address being included in the multi-response set is "addr_dyn_weight / group_max_dyn_weight".

"up_thresh" operates on all groups as a whole, and if the non-"DOWN" sum of all weights in all groups fails to meet the standard of "ceil(up_thresh * total_sum_configured_weight)", then all addresses will be treated as if they are "UP" for selection purposes, and resource-level failure will be signalled upstream.

THE GROUPED MULTI CASE

You can probably infer this one's behavior from reading about the previous three cases. The difference from the previous grouped-single case is that the multi-vs-single behaviors are reversed. Multiple groups are chosen based on the dynamic maximum weight between the groups, and a single weighted address is returned from the subset within each chosen group. All of the details above logically apply in the way you would expect, as all of these four cases internally share the same code and logic, they just apply different bits of it to different subsets of the problem.

GENERAL NOTES ON ADDRESS MODE CASES ABOVE

Note that any time multi-selection is in effect at a layer (the top layer when multi is true, or within a group when when multi is false), the minimum count of chosen items will be the count of items that share the maximum weight within the set. e.g. a set of items with weights "30, 30, 30, 20, 20" will always choose at least 3/5 items (because the first three have 100% odds of inclusion), and the total response set will range as high as all 5 items with some probability.

A practical use-case example for grouped-single:

Splitting groups on subnet boundaries in grouped-single mode gives the result that a single response packet never mixes subnets. This would enable your DNS-based balancing to defeat certain forms of client-level Destination Address Selection interference, while still returning multiple addresses per response (all from one subnet).

A practical use-case for grouped-multi:

Suppose you have a large set of addresses which can be logically grouped into subsets that have some shared failure risk (e.g. subpartitions of a datacenter which share infrastructure). With grouped-multi behavior, clients will get up to N (count of groups) addresses in a round-robin response, but a given response set will never contain two addresses from the same group/subset. This maximizes the chance that the client can successfully fail over to another address in the list when its primary selection fails, since the total set in each response does not share any per-subset failure mode.

LIMITS

All weights must be positive integer values greater than zero and less than 2^20 (1048576).

There is a limit of 64 addresses, address-groups, or cnames at the top level of a resource (or per address family in the addrs_v4/addrs_v6 cases), and a limit of 64 addresses within each address group in the grouped modes.

COPYRIGHT AND LICENSE

This file is part of gdnsd.

gdnsd-plugin-weighted is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

gdnsd-plugin-weighted is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with gdnsd-plugin-weighted. If not, see <http://www.gnu.org/licenses/>.