resolver: Make EndpointMap's Get, Set and Delete operations O(1)

[arjan-bal]

Fixes: #8173

This PR changes only the internal implementation of EndpointMap to use a map keyed by an unordered list of addresses. This avoids the need to iterate over the entire map to check the existence of a single element. This brings down the average case time complexity for Get, Set and Delete from O(n) to O(1) per operation, where n is the number of endpoints in the map.

Benchmarks

Code:

func BenchmarkEndpointMap(b *testing.B) {
	em := NewEndpointMap()
	for i := range b.N {
		em.Set(Endpoint{
			Addresses: []Address{{Addr: fmt.Sprintf("%d.%d.%d.%d", i, i, i, i)}},
		}, i)
	}
	for i := range b.N {
		em.Get(Endpoint{
			Addresses: []Address{{Addr: fmt.Sprintf("%d.%d.%d.%d", i, i, i, i)}},
		})
	}
	for i := range b.N {
		em.Delete(Endpoint{
			Addresses: []Address{{Addr: fmt.Sprintf("%d.%d.%d.%d", i, i, i, i)}},
		})
	}
}

Without any changes

goos: linux
goarch: amd64
pkg: google.golang.org/grpc/resolver
cpu: Intel(R) Xeon(R) CPU @ 2.60GHz
BenchmarkEndpointMap-48            10000            839907 ns/op
PASS

With base64 encoding the address list

goos: linux
goarch: amd64
pkg: google.golang.org/grpc/resolver
cpu: Intel(R) Xeon(R) CPU @ 2.60GHz
BenchmarkEndpointMap-48           373237              3770 ns/op
PASS

With JSON encoding the address list

goos: linux
goarch: amd64
pkg: google.golang.org/grpc/resolver
cpu: Intel(R) Xeon(R) CPU @ 2.60GHz
BenchmarkEndpointMap-48           289248              4275 ns/op
PASS

RELEASE NOTES:

• resolver: EndpointMap's Get, Set and Delete operations are now O(1) on average.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.00%. Comparing base (1f6b0cf) to head (1887f0a).
Report is 5 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #8179      +/-   ##
==========================================
- Coverage   82.42%   82.00%   -0.43%     
==========================================
  Files         393      410      +17     
  Lines       39162    40233    +1071     
==========================================
+ Hits        32280    32993     +713     
- Misses       5571     5878     +307     
- Partials     1311     1362      +51     

Files with missing lines	Coverage Δ
resolver/map.go	96.19% <100.00%> (-0.56%)	⬇️

... and 39 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[atollena]

This is still O(n) in the number of addresses, no?

[arjan-bal]

encodeEndpoint is O(n) where n is the number of addresses within a single endpoint. Practically a single endpoint should contain around 2 addresses or so.

em.endpoints[x] is O(n) where n is also the number of addresses within a single endpoint. This is independent of the number of endpoints in the map, which can be much larger than n.

[easwars]

Do you think it makes sense to checkin some/all of the benchmark code that you have for posterity?

[easwars]

Nit: return nil, false here to be explicit about the second return value.

[arjan-bal]

Done.

[arjan-bal]

Do you think it makes sense to checkin some/all of the benchmark code that you have for posterity?

Added the micro benchmark mentioned in the PR description.