Add Go benchmarks for hot paths: Check, Lookup, GetRoute, statsd emission

[dolph]

Summary

The codebase has no *_test.go benchmarks. Every performance discussion (this issue, #11, and the proposed fixes in #21) ends up arguing about ballpark numbers. A small benchmark suite gives a real baseline before any rework and prevents regressions after.

Scope

Add Benchmark* functions covering the per-check hot paths and the metric pipeline. Suggested set:

1. BenchmarkCheck (destinations_bench_test.go) — full Check() against localhost with a stub statsd drain. Measures end-to-end per-check overhead including DNS, routing, dial, HTTPS.

2. BenchmarkLookup (resolver_bench_test.go) — pure DNS path. Pairs with #-DNS-caching once filed.

3. BenchmarkGetRoute (router_bench_test.go) — quantifies routing.New() per-call cost. Pairs with the routing-table-cache issue.

4. BenchmarkDestinationIncrement + BenchmarkTagsRebuild (destinations_bench_test.go) — allocations in the metric tag path. Pairs with the dest.tags() caching issue.

5. BenchmarkStatsdSender (statsd_bench_test.go) — connection-per-metric cost; needs a UDP listener on 127.0.0.1:0 as a sink. Pairs with Statsd emitter can stall checks when statsd is down; opens a new connection per metric #11.

6. BenchmarkGetLocalIPs (source_bench_test.go) — quantifies the per-log-line cost flagged in Logs are unstructured plain text without levels; GetLocalIPs runs per log line #21.

7. BenchmarkMonitorThroughput — simulates 100/500/1000 destinations to measure goroutine + GC + statsd-queue behavior with b.RunParallel and a stub server.

Helpful conventions

• All benchmarks call b.ReportAllocs() so allocs/op is in the output.
• Use testing.B.Run(name, ...) to parameterize by destination count for the monitor-throughput benchmark.
• Add a make bench target (or ./build.sh bench) running go test -bench=. -benchmem ./....
• Optionally wire into CI as a smoke run (-benchtime=1x) and a separate manual go test -bench=. -benchtime=10s workflow for serious measurement.

Stub statsd sink

Most of the above benchmarks need a way to keep the queue <- s send from blocking. Two options:

// Option A: drain into /dev/null
func init() { go func() { for range queue {} }() }

// Option B: replace the sender for the test

Either works; Option B is cleaner once #11 lands and the sender becomes an interface.

Why bother

• Establishes a baseline so post-fix improvements for Statsd emitter can stall checks when statsd is down; opens a new connection per metric #11, the routing-cache issue, the DNS-cache issue, and the tags-cache issue are quantifiable instead of speculative.
• Catches regressions when refactors land (e.g. switching to log/slog per Logs are unstructured plain text without levels; GetLocalIPs runs per log line #21).
• Cheap to maintain — no external dependencies.

Related

Cross-references: #11 (statsd sender), #21 (GetLocalIPs per log line), plus the three perf issues filed alongside this one.