Make throttling nmagent fetches for nodesubnet more dynamic #3023
Conversation
santhoshmprabhu
changed the title
…de-subnet-intelligent-refresh
|
/azp run Azure Container Networking PR |
|
Azure Pipelines successfully started running 1 pipeline(s). |
…de-subnet-intelligent-refresh
|
/azp run Azure Container Networking PR |
|
Azure Pipelines successfully started running 1 pipeline(s). |
timraymond
left a comment
timraymond
left a comment
There was a problem hiding this comment.
I think I'm going to recant some of what I said in the last PR now that I understand why we're doing this debouncing. As I mentioned during the design review, rate limits and wireserver are a fact of life, so it now feels weird to have arbitrary logic treating the nmagent client carefully because it might blow our rate limit. The design intention of the NMAgent client was always that users can do nmagent.DoSomethingForMe() and it will block while it deals with all of the weirdness that happens with NMAgent and Wireserver. This includes async response handling... and I'm thinking it should also include respecting rate limits. If the request is going to blow the rate limit, the WireserverTransport in the NMAgent client needs to block that request until the request will succeed. Otherwise, we're just setting the user up for failure.
Doing this with tickers is definitely the right way (in particular the whole ticker resetting thing is how I've done respecting rate limits in the past). However, it should be done in https://github.com/Azure/azure-container-networking/blob/master/nmagent/internal/wireserver.go so that it's shared across the client.
…de-subnet-intelligent-refresh
|
/azp run Azure Container Networking PR |
|
/azp run Azure Container Networking PR |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run Azure Container Networking PR |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run Azure Container Networking PR |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run Azure Container Networking PR |
|
Azure Pipelines successfully started running 1 pipeline(s). |
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
…thub.com:Azure/azure-container-networking into sanprabhu/cilium-node-subnet-intelligent-refresh
|
/azp run Azure Container Networking PR |
|
Azure Pipelines successfully started running 1 pipeline(s). |
santhoshmprabhu
deleted the
sanprabhu/cilium-node-subnet-intelligent-refresh
branch
* feat(CNS): Early work on better throttling in NMAgent fetch for nodesubnet * feat(CNS): Update NMAgent fetches to be async with binary exponential backoff * chore: check for empty nmagent response * test: update test for empty response * style: make linter happy * chore: fix some comments * fix: Fix bug in refresh * refactor: Address comments * refactor: ignore primary ip * refactor: move refresh out of ipfetcher * test: add ip fetcher tests * fix: remove broken import * fix: fix import * fix: fix linting * fix: fix some failing tests * chore: Remove unused function * test: test updates * fix: address comments * chore: add missed file * chore: add comment about static interval * feat: address Evan's comment to require Equal method on cached results * chore: add missed file * feat: more efficient equality * refactor: address Evan's comment * refactor: address Tim's comment * fix: undo accidental commit * fix: make linter happy * fix: make linter happy
Reason for Change:
We previously added fetch of secondary IPs to the NMAgent client in CNS. This PR makes the fetch asynchronous, with the ability to adjust the frequency dynamically. Specifically,
This is implemented using a time Ticker, but we wrap it with a wrapper so as to mock it for tests. This PR also includes some additional checks for empty responses from the nmagent (this was raised in this comment). Also have UTs.
Issue Fixed:
NA
Requirements:
Notes: