When AAAA multi-target / dual stack support was
added via #2461 it broke ownership of domains across
different clusters with different ingress records types.
For example if 2 clusters manage the same zone,
1 cluster uses A records and the other uses CNAME
records, when each record type is treated as a separate
planning record, it will cause ownership to bounce back
and forth and records to be constantly created and
deleted.
This change updates the planner to keep track of multiple
current records for a domain. This allows for A and AAAA
records to exist for a domain while allowing record type
changes.
The planner will ignore desired records for a domain that
represent conflicting record types allowed by RFC 1034 3.6.2.
For example if the desired records for a domain contains
a CNAME record plus any other record type no changes for
that domain will be planned.
The planner now contains an owned record filter provided
by the registry. This allows the planner to accurately plan
create updates when there are record type changes between
the current and desired endpoints. Without this filter the
planner could add create changes for domains not owned
by the controller.
Log whenever a target is not able to be parsed as an IP address. This is expected
to occur fairly often (for example, with CNAME targets), but allows more visibility
into how targets are being compared.
NOTE: depending on the number and type of targets, this could be quite noisy.
Previously there was no distinction between an IP address and any other string
when doing a comparison to determine which is "less" when determining which endpoint to actually create.
This explicitly handles IP addresses and will always prefer
them over non-IP addresses when determining which of two targets is less.
This adds support for all AWS Route53 routing policies, namely:
* simple (default)
* weighted
* latency
* failover
* geolocation
* multi value answer
These routing policies allow to create multiple records with the same
name, but different "SetIdentifiers". Therefor, as a prerequisite for
implementing support for above routing policies, there is a new
"abstraction layer" added that handles this SetIdentifier by adding a
new attribute in the Endpoint struct and adding another level in the
plan table.
* First stab at NodePort support. Testing incomplete
* Fix up the unit tests
* Remove some deadcode in the unittests
* gather node ips once and add support for srv records
* Make sure we match gofmt simple
* Move the nodes to the testcase and add a test for clusters that only have internal ip addresses
* Somehow forgot about the weight field in the records
* Add SRV as a supported record type
* Endpoint.Target is now Endpoint.Targets. This is its own type representing mutliple targets for a single DNS name while adding some convenience for sorting and comparing
* Made everything compile and tests run through with the new Endpoint.Targets
* The ingress source can now properly handle multiple target ips per host
* Added custom conflict resolver, to better understand how conflict resolution has to work for me
* My custom conflict resolver behaves a bit different than the PerResource resolver, therefore I needed to modify the expected test result
Removed unnecessary FIXME
* The ingress source now creates CNAME endpoints with multiple targets to let the DNS provider decide how to handle multiple CNAME targets. This could be interesting for weighted targets etc.
* Adopted the expected results to the new way we create endpoints for CNAMEs
* Removed Add method from Targets since manipulating the slice through here is unnecessary complicated and doesn't deliver enough convenience
* Reverted ConflictResolver to the original one. There is some discussing to do what the best way is to handle conflicts
* Added missing documenting comment to IsLess of Targets
* Added documenting comments to Targets,Targets.Same and NewTargets to clarify their intention and usage
* Service source now also generates endpoints with multiple targets
* Service and Ingress source now sort all Targets for every Endpoint to make order of Targets predictable
* Endpoints generated by the Google Cloud DNS provider now also have sorted Targets to make order of Targets predictable
* Modified provider dyn to be able to compile with multi target changes
* Fixed small nitpicks, so my code is acceptable
* Fixed merge method after updating to new Targets. Replacing '!=' with .Same of course needs a boolean negation
* Tests for dyn provider now also use the new Targets instead of Target
* Simplified extractServiceIps as implied by linki to make it more readable
* ref: change service ClusterIP retrieval again
* Added entry to CHANGELOG.md describing the new features contained in this PR
* Make suitableType() be Endpoint method
With this change it becomes possible to work with endpoint
of empty type in packages other than "provider". Also
it seems logical for a smart property getter without side effects
to be a method rather than a function in different package
* Make plan computation work correctly with multi-target domains
* fix drawing
* drop comments
* fix boilerplate header
* fix comment
* fix the bug with empty map
* rework registry to support random lables
* serialize->serializeLabel function rename
* golint for err variable naming
* add additional test
* add tests for current case where one resource can generate multiple endpoints
* make labels have its own type, add serialization as a method
* add comment for exported error
* use greater rather than not equal zero
* update changelog
