Linear cache: Add new UpdateResources method to update multiple resources at once without doing a stow update and limit unneeded allocations when processing delta watches

[valerian-roche]

In use cases where linear cache is used with eds and delta watches, the user can currently only update the cache in two ways:

• full stow set (SetResources): this requires the user to build the entire list to then pass it to the cache, which then hashes everything to figure out what changed. This is very resource intensive (in our case ~1s of CPU). After that it will only check once for delta watches and return all changed endpoints at once
• provide detailed updates one by one (through UpdateResource and DeleteResource): this does not require the user to keep a full view of the data, and does not consume much cpu in hashing (only hashing the resources actually updated), but it will check all delta watches for each call, and will forward endpoints changed a key at a time (more traffic with clients)

In our case the main issue switching to the second part is the cpu usage (and time spent) on updates when updating multiple resources (though with still num_updates << total_keys, e.g. 50 vs 5000) while multiple delta watches exist (~20)
We then end up spending more than 120ms of CPU (for ~50 updates/deletes) with profiles like (this is aggregated for updates between 0 and 50):

In order to improve this, two changes are in this PR:

• Do not pre-allocate the slice of filtered resources in most cases
  • When the delta watch is non-wildcard, this is irrelevant to pre-allocate for the total list of resources, and most checks will return no change at all
  • When the delta watch is wildcard, pre-allocate when there is no version passed, but not in other cases when most resources have not been updated
• Provide a method to update/delete multiple resources at once. This allows only one iteration across watches, improving resource usage and limiting traffic between the control-plane and the clients

After the fix, runtime of the update (only includes iterating through calls to UpdateResource/DeleteResource before and on call to UpdateResources after) is greatly improved and stable (blue line below)

[alecholmez]

Thanks for this PR! I'll review this first thing tomorrow. I just approved the run, can you go ahead and fix up the lint checks?

[valerian-roche]

Thanks for this PR! I'll review this first thing tomorrow. I just approved the run, can you go ahead and fix up the lint checks?

Just pushed the fix for the lint (at least working locally)

[alecholmez]

Just a few minor nits/questions. Overall looks great thanks for the enhancements!

[alecholmez]

Looks great, thanks for the PR! Once Github fixes actions we'll let the build run then get this merged if all is well.

[valerian-roche]

Would an owner be able to trigger the checks to confirm it is all good?

[alecholmez]

Would an owner be able to trigger the checks to confirm it is all good?

I am an owner :) unfortunately the APIs are actually down: https://www.githubstatus.com/ Github seems to be having an outage. I'm watching that page to see when they get resolved. Sorry for the delay!

[valerian-roche]

My bad, didn't see that it was down, sorry about that!

[alecholmez]

@valerian-roche can you push a dummy commit? I think CI needs a trigger. Some instructions here to do so: https://github.com/envoyproxy/go-control-plane/blob/main/CONTRIBUTING.md#triggering-ci-re-run-without-making-changes

[alecholmez]

@valerian-roche just merged! Thanks again nice work

[valerian-roche]

Thanks for the fast review and merge, greatly appreciated!