openconfig / kne / 29960987883
42%

DEFAULT BRANCH: main
Ran
Jobs 1
Files 62
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 42.133% (-0.008%) from 42.141%
29960987883

push

github

web-flow 
Remove limitation preventing high link counts (#717)

* Add testing specifically for the 1500 link case

* Use pod affinity / antiaffinity for determinism

Rather than leaving it to chance, explicitly make a test
for local veth vs remote vxlan

* Create a helper to replace koko

This has these properties for better performance
* Gathers all changes for a given pod (i.e. netns) into a batch
* Sets up veths, moves them, etc within a single netns context change
  for the pod batch. (Well, actually two.)
* Uses deterministic names so that idempotent recovery from a mid-batch
  crash is doable.

* Shift the work of creating interfaces to meshnetd

This simplifies the CNI plugin so that it can reliably return within
K8s' time limit.

It has meshnetd do the work of creating interfaces and moving them into
the pod's netns. This has important benefits
 * It eliminates the skip / don't skip multiprocess coordination between
   paired pods, which otherwise causes many network round trips both
   between the plugin and meshnetd, and also meshnetd and kubeapi.
 * Meshnetd can much more easily batch operations within a pod
   namespace. This reduces the number of expensive netns system calls
   with large numbers of links

It has a side effect that the CNI plugin returns before all interfaces
are present. (Really, before ANY are present.) That lets the pod go
RUNNING before the interfaces are ready.

However, this was always true for interfaces that are skipped (that is,
the other end of the link is responsible for setting it up.) The
init-wait container already exists to address this already-present
problem.

* Use init-wait to make sure all interfaces are present

Before the pod is running, thus before pings happen

* Have the CNI plugin wait for link up

This is bounded, it waits for up to 5 seconds.

For existing common small topologies, this will make
sure the interfaces are all ready before the pod is
considered running (before the main container starts.)

Th... (continued)

5401 of 12819 relevant lines covered (42.13%)

0.48 hits per line

Coverage Regressions

Lines Coverage ∆ File
1
67.26
 -0.3% topo/node/arista/arista.go
Subprojects
ID Flag name Job ID Ran Files Coverage
1 /jobs/185031500 29960987883.1 62
42.13
 GitHub Action Run
Source Files on build 29960987883