go-race-detector-workflow

Go's concurrency model (goroutines + channels + the memory model) differs structurally from OS-thread models. The race detector is compiled into the binary at build time via ThreadSanitizer instrumentation; goroutine leaks are a separate failure class that the race detector does not cover. This skill walks both checks, from first run to CI gate.

race-condition-test-author covers multi-language deterministic interleaving (barriers, jcstress, TSan for C/C++). This skill focuses exclusively on Go: the -race flag, GORACE tuning, stress amplification, and goleak.

Step 1 - Enable the race detector

Per go.dev/doc/articles/race_detector, add -race to any go command:

go test -race ./...
go run  -race main.go
go build -race ./cmd/server

The flag compiles ThreadSanitizer instrumentation into the binary. It requires cgo and a C compiler (on Linux/FreeBSD/Windows; Darwin ships its own). Supported platforms as of Go 1.22: linux/amd64, linux/arm64, linux/ppc64le, linux/s390x, linux/loong64, freebsd/amd64, netbsd/amd64, darwin/amd64, darwin/arm64, windows/amd64 (mingw-w64 runtime v8+ required).

Per go.dev/doc/articles/race_detector, expected overhead:

• Execution time: 2-20x slower.
• Memory: 5-10x increase.
• Additional 8 bytes per defer/recover (unbounded in long-running goroutines; budget accordingly in CI timeouts).

Step 2 - Read a race report

A detected race prints two goroutine stacks to stderr:

WARNING: DATA RACE
Write at 0x00c0000b4010 by goroutine 7:
  main.(*Cache).set+0x6c
    /home/user/app/cache.go:38

Previous read at 0x00c0000b4010 by goroutine 6:
  main.(*Cache).get+0x44
    /home/user/app/cache.go:22

Goroutine 7 (running) created at:
  main.runWorker+0x34
    /home/user/app/main.go:71

The report names the conflicting accesses (read vs. write), the memory address, and the goroutine creation sites. Fix by protecting all accesses to the address with the same synchronization primitive (mutex, atomic, or channel hand-off).

Step 3 - Tune GORACE options

Per go.dev/doc/articles/race_detector, set GORACE before the command:

GORACE="log_path=/tmp/race/report halt_on_error=1 history_size=2" \
  go test -race ./...

Useful options:

Step 4 - Stress with -count and -cpu

The race detector only fires on races that actually execute. A single go test -race run on a lightly-contended path may produce zero output and still miss a real race. Amplify coverage:

# Run each test 10 times per package
go test -race -count=10 ./...

# Exercise multiple GOMAXPROCS values
go test -race -cpu=1,2,4,8 ./...

# Combine: 5 runs at each GOMAXPROCS
go test -race -count=5 -cpu=1,2,4 ./...

-cpu sets GOMAXPROCS for each comma-separated value, then re-runs. Running at GOMAXPROCS=1 surfaces sequencing bugs; higher values surface true parallel races. Combining both increases scheduler interleaving diversity without extra code.

Step 5 - Check loop-variable capture with go vet

Per the [Go vet documentation at go.dev/cmd/vet], go vet flags the classic loop-variable-capture anti-pattern that frequently causes races when goroutines close over a range variable:

// Before Go 1.22 - race: all goroutines capture the same &v
for _, v := range items {
    go func() { process(v) }()  // vet warns here
}

// Fix: copy the variable
for _, v := range items {
    v := v
    go func() { process(v) }()
}

Run before -race to filter out this class early:

go vet ./...
go test -race ./...

In Go 1.22+, range variables are per-iteration by default; the capture pattern is still worth auditing in code that may be compiled with older toolchains.

Step 6 - Detect goroutine leaks with goleak

A goroutine that starts but never stops is a leak: the race detector ignores it (no concurrent access violation), but the goroutine holds resources and inflates memory over time.

Install per github.com/uber-go/goleak:

go get -u go.uber.org/goleak

Per-test: VerifyNone

import "go.uber.org/goleak"

func TestWorkerPool(t *testing.T) {
    defer goleak.VerifyNone(t)

    pool := NewWorkerPool(4)
    pool.Submit(func() { /* work */ })
    pool.Shutdown()
    // VerifyNone fires after Shutdown() returns;
    // any still-running worker goroutine fails the test.
}

Per github.com/uber-go/goleak, VerifyNone is incompatible with t.Parallel(): goleak cannot associate a specific goroutine with a specific parallel sub-test.

Package-level: VerifyTestMain

For packages that use t.Parallel(), wrap the test runner instead:

func TestMain(m *testing.M) {
    goleak.VerifyTestMain(m)
}

VerifyTestMain runs the full test binary, then checks for leaked goroutines once all tests have completed.

Filtering expected goroutines

Third-party libraries sometimes leave background goroutines that are intentional. Silence them with options per pkg.go.dev/go.uber.org/goleak:

// Ignore a goroutine whose top-of-stack is this function
goleak.VerifyNone(t,
    goleak.IgnoreTopFunction("database/sql.(*DB).connectionOpener"),
)

// Ignore a function anywhere in the stack (v1.3.0+)
goleak.VerifyNone(t,
    goleak.IgnoreAnyFunction("google.golang.org/grpc.(*ccBalancerWrapper).watcher"),
)

// Snapshot existing goroutines at test start; ignore them at end
opt := goleak.IgnoreCurrent()
// ... test logic ...
goleak.VerifyNone(t, opt)

Prefer IgnoreTopFunction over IgnoreCurrent when the library goroutine is identifiable by name: IgnoreCurrent silences goroutines that were already running at snapshot time, which can mask leaks introduced before the snapshot.

Step 7 - CI matrix

Gate both checks in CI. Run -race in at least one matrix dimension (per go.dev/doc/articles/race_detector: "It is recommended to always run race-enabled tests"):

jobs:
  test:
    strategy:
      matrix:
        go-version: ["1.22", "1.23"]
        race: ["", "-race"]
    steps:
      - uses: actions/setup-go@v5
        with:
          go-version: ${{ matrix.go-version }}

      - name: Run tests
        env:
          GORACE: "log_path=/tmp/race/report halt_on_error=0"
        run: |
          go vet ./...
          go test ${{ matrix.race }} -count=3 -cpu=1,4 -timeout=10m ./...

      - name: Upload race reports
        if: failure()
        uses: actions/upload-artifact@v4
        with:
          name: race-reports-${{ matrix.go-version }}-${{ matrix.race }}
          path: /tmp/race/report*

-race adds 2-20x overhead; set -timeout to at least 5-10x your non-race run time. Upload log_path files on failure so the report survives the run.

Anti-patterns

Limitations

• The race detector only fires on races that execute in the instrumented run. Low-probability interleavings require stress (-count, -cpu) or barrier-based deterministic tests (see race-condition-test-author).
• Per go.dev/doc/articles/race_detector, cgo is required; cross-compiled binaries (e.g., GOOS=linux GOARCH=arm on a Mac) will not run with -race unless the target toolchain supports TSan.
• Per github.com/uber-go/goleak, goleak requires one of the two most recent minor versions of Go; older toolchains are not supported.
• The race detector adds 8 bytes per defer/recover that accumulates until the goroutine exits, not until the deferred function returns. Long- running service binaries built with -race can leak memory faster than typical tests reveal.
• goleak does not distinguish between a goroutine that will stop shortly and one that is genuinely leaked. VerifyNone has a brief internal retry loop, but tests that start background goroutines with long startup delays can produce false positives; use IgnoreTopFunction to suppress known cases.

References

• go.dev/doc/articles/race_detector - canonical reference for -race, GORACE options, report format, overhead, and platform support
• github.com/uber-go/goleak - goleak source, VerifyNone, VerifyTestMain, and filter option docs
• pkg.go.dev/go.uber.org/goleak - full API reference for IgnoreTopFunction, IgnoreAnyFunction, IgnoreCurrent, Cleanup
• race-condition-test-author - multi-language deterministic interleaving (barriers, jcstress, TSan for C/C++); this skill is the Go-specific complement
• deadlock-detection-harness - deadlock patterns; separate from data races