worker

go-worker

go-worker provides a simple way to manage and execute prioritized tasks concurrently, backed by a TaskManager with a worker pool and a priority queue.

Breaking changes (January 2026)

• Stop() removed. Use StopGraceful(ctx) or StopNow().
• Local result streaming uses SubscribeResults(buffer); GetResults() is now a compatibility shim and the legacy local StreamResults() is removed (gRPC StreamResults remains).
• RegisterTasks now returns an error.
• Task.Execute replaces Fn in examples.
• NewGRPCServer requires a handler map.
• Rate limiting is deterministic: burst is min(maxWorkers, maxTasks) and ExecuteTask uses the shared limiter.
• gRPC durable tasks use RegisterDurableTasks and the new DurableTask message.
• When a durable backend is configured, use RegisterDurableTask(s) instead of RegisterTask(s).
• DurableBackend now requires Extend (lease renewal support for custom backends).

Features

• Task prioritization: tasks are scheduled by priority.
• Concurrent execution: tasks run in a worker pool with strict rate limiting.
• Middleware: wrap the TaskManager for logging/metrics, etc.
• Results: fan-out subscriptions via SubscribeResults.
• Cancellation: cancel tasks before or during execution.
• Retries: exponential backoff with capped delays.
• Durability: optional Redis-backed durable task queue (at-least-once, lease-based).

Architecture

flowchart LR
    Client[Client code] -->|register tasks| TaskManager
    TaskManager --> Queue[Priority Queue]
    Queue -->|dispatch| Worker1[Worker]
    Queue -->|dispatch| WorkerN[Worker]
    Worker1 --> Results[Result Broadcaster]
    WorkerN --> Results

gRPC Service

go-worker exposes its functionality over gRPC through the WorkerService. The service allows clients to register tasks, stream results, cancel running tasks and query their status.

Handlers and Payloads

The server registers handlers keyed by name. Each handler consists of a Make function that constructs the expected payload type, and a Fn function that executes the task logic using the unpacked payload.

Clients send a Task message containing a name and a serialized payload using google.protobuf.Any. The server automatically unpacks the Any payload into the correct type based on the registered handler and passes it to the corresponding function. For durable tasks, use RegisterDurableTasks with the DurableTask message (the payload is still an Any).

handlers := map[string]worker.HandlerSpec{
    "create_user": {
        Make: func() protoreflect.ProtoMessage { return &workerpb.CreateUserPayload{} },
        Fn: func(ctx context.Context, payload protoreflect.ProtoMessage) (any, error) {
            p := payload.(*workerpb.CreateUserPayload)
            return &workerpb.CreateUserResponse{UserId: "1234"}, nil
        },
    },
}

srv := worker.NewGRPCServer(tm, handlers)

For production, configure TLS credentials and interceptors (logging/auth) on the gRPC server; see __examples/grpc for a complete setup. For a Redis-backed durable gRPC example, see __examples/grpc_durable.

Security defaults to follow in production:

• Use TLS (prefer mTLS) for gRPC; the durable gRPC example uses insecure credentials for local demos only.
• Scrub payloads and auth metadata from logs; log task IDs or correlation IDs instead of PII.
• Implement auth via WithGRPCAuth and redact/validate tokens inside interceptors.

Durable gRPC client (example)

Use RegisterDurableTasks for persisted tasks (payload is still Any). Results stream is shared with non-durable tasks.

payload, _ := anypb.New(&workerpb.SendEmailPayload{
    To:      "ops@example.com",
    Subject: "Hello durable gRPC",
    Body:    "Persisted task",
})

resp, err := client.RegisterDurableTasks(ctx, &workerpb.RegisterDurableTasksRequest{
    Tasks: []*workerpb.DurableTask{
        {
            Name:           "send_email",
            Payload:        payload,
            IdempotencyKey: "durable:send_email:ops@example.com",
        },
    },
})
if err != nil {
    log.Fatal(err)
}

Authorization hook

You can enforce authentication/authorization at the gRPC boundary with WithGRPCAuth. Return a gRPC status error to control the response code (e.g., Unauthenticated or PermissionDenied).

auth := func(ctx context.Context, method string, _ any) error {
 md, _ := metadata.FromIncomingContext(ctx)
 values := md.Get("authorization")
 if len(values) == 0 {
  return status.Error(codes.Unauthenticated, "missing token")
 }

 token := strings.TrimSpace(strings.TrimPrefix(values[0], "Bearer "))
 if token != "expected-token" {
  return status.Error(codes.Unauthenticated, "missing or invalid token")
 }

 return nil
}

srv := worker.NewGRPCServer(tm, handlers, worker.WithGRPCAuth(auth))

Note on deadlines: When the client uses a stream context with a deadline, exceeding the deadline will terminate the stream but does not cancel the tasks running on the server. To properly handle cancellation, use separate contexts for task execution or cancel tasks explicitly.

API Example (gRPC)

tm := worker.NewTaskManagerWithDefaults(context.Background())
handlers := map[string]worker.HandlerSpec{
    "create_user": {
        Make: func() protoreflect.ProtoMessage { return &workerpb.CreateUserPayload{} },
        Fn: func(ctx context.Context, payload protoreflect.ProtoMessage) (any, error) {
            p := payload.(*workerpb.CreateUserPayload)
            return &workerpb.CreateUserResponse{UserId: "1234"}, nil
        },
    },
}

srv := worker.NewGRPCServer(tm, handlers)

gs := grpc.NewServer()
workerpb.RegisterWorkerServiceServer(gs, srv)
// listen and serve ...

client := workerpb.NewWorkerServiceClient(conn)

// register a task with payload
payload, err := anypb.New(&workerpb.CreateUserPayload{
    Username: "newuser",
    Email:    "newuser@example.com",
})
if err != nil {
    log.Fatal(err)
}

_, _ = client.RegisterTasks(ctx, &workerpb.RegisterTasksRequest{
    Tasks: []*workerpb.Task{
        {
            Name:           "create_user",
            Payload:        payload,
            CorrelationId:  uuid.NewString(),
            IdempotencyKey: "create_user:newuser@example.com",
            Metadata:       map[string]string{"source": "api_example", "role": "admin"},
        },
    },
})

// cancel by id
_, _ = client.CancelTask(ctx, &workerpb.CancelTaskRequest{Id: "<task-id>"})

// get task information
res, _ := client.GetTask(ctx, &workerpb.GetTaskRequest{Id: "<task-id>"})
fmt.Println(res.Status)

API Usage Examples

Quick Start

tm := worker.NewTaskManager(context.Background(), 2, 10, 5, 30*time.Second, time.Second, 3)

task := &worker.Task{
    ID:       uuid.New(),
    Priority: 1,
    Ctx:      context.Background(),
    Execute:  func(ctx context.Context, _ ...any) (any, error) { return "hello", nil },
}

if err := tm.RegisterTask(context.Background(), task); err != nil {
    log.Fatal(err)
}

results, cancel := tm.SubscribeResults(1)
res := <-results
cancel()

fmt.Println(res.Result)

Result backpressure

By default, full subscriber buffers drop new results. You can change the policy:

tm.SetResultsDropPolicy(worker.DropOldest)

GetResults() remains as a compatibility shim and returns a channel with a default buffer. Prefer SubscribeResults(buffer) so you can control buffering and explicitly unsubscribe.

Initialization

Create a new TaskManager by calling the NewTaskManager() function with the following parameters:

• ctx is the base context for the task manager (used for shutdown and derived task contexts)
• maxWorkers is the number of workers to start. If <= 0, it will default to the number of available CPUs
• maxTasks is the maximum number of queued tasks, defaults to 10
• tasksPerSecond is the rate limit of tasks that can be executed per second. If <= 0, rate limiting is disabled (the limiter uses a burst size of min(maxWorkers, maxTasks) for deterministic throttling)
• timeout is the default timeout for tasks, defaults to 5 minutes
• retryDelay is the default delay between retries, defaults to 1 second
• maxRetries is the default maximum number of retries, defaults to 3 (0 disables retries)

tm := worker.NewTaskManager(context.Background(), 4, 10, 5, 30*time.Second, 1*time.Second, 3)

Durable backend (Redis)

Durable tasks use a separate DurableTask type and a handler registry keyed by name. The default encoding is protobuf via ProtoDurableCodec. When a durable backend is enabled, RegisterTask/RegisterTasks are disabled in favor of RegisterDurableTask(s). See __examples/durable_redis for a runnable example.

client, err := rueidis.NewClient(rueidis.ClientOption{
    InitAddress: []string{"127.0.0.1:6379"},
})
if err != nil {
    log.Fatal(err)
}
defer client.Close()

backend, err := worker.NewRedisDurableBackend(client)
if err != nil {
    log.Fatal(err)
}

handlers := map[string]worker.DurableHandlerSpec{
    "send_email": {
        Make: func() proto.Message { return &workerpb.SendEmailRequest{} },
        Fn: func(ctx context.Context, payload proto.Message) (any, error) {
            req := payload.(*workerpb.SendEmailRequest)
            // process request
            return &workerpb.SendEmailResponse{MessageId: "msg-1"}, nil
        },
    },
}

tm := worker.NewTaskManagerWithOptions(
    context.Background(),
    worker.WithDurableBackend(backend),
    worker.WithDurableHandlers(handlers),
)

err = tm.RegisterDurableTask(context.Background(), worker.DurableTask{
    Handler: "send_email",
    Message: &workerpb.SendEmailRequest{To: "ops@example.com"},
    Retries: 5,
})
if err != nil {
    log.Fatal(err)
}

Defaults: lease is 30s, poll interval is 200ms, Redis dequeue batch is 50, and lease renewal is disabled (configurable via options).

Operational notes (durable Redis):

• Key hashing: Redis Lua scripts touch multiple keys; for clustered Redis, all keys must share the same hash slot. The backend auto-wraps the prefix in {} to enforce this (e.g., {go-worker}:ready).
• DLQ: Failed tasks are pushed to a dead-letter list ({prefix}:dead). There is no built-in replay yet; you should build a replay tool or manually re-enqueue as needed.
• DLQ replay: See __examples/durable_dlq_replay for a small Lua-based replay utility (dry-run by default; use -apply to replay).
• Multi-node workers: Multiple workers can safely dequeue from the same backend. Lease timeouts handle worker crashes, but tune WithDurableLease for your workload.
• Lease renewal: enable WithDurableLeaseRenewalInterval for long-running tasks to extend leases while a task executes.
• Visibility: Ready and processing queues live in sorted sets; you can inspect sizes via ZCARD on {prefix}:ready and {prefix}:processing.
• Inspect utility: __examples/durable_queue_inspect prints ready/processing/dead counts; use -show-ids to display IDs (hidden by default to avoid leaking sensitive identifiers).

Multi-node coordination (durable Redis)

Durable processing is at-least-once. When multiple nodes consume from the same Redis backend:

• Lease sizing: set WithDurableLease longer than your worst-case task duration (plus buffer). If a task exceeds its lease, it can be requeued and run again on another node.
• Lease renewal (optional): set WithDurableLeaseRenewalInterval (less than the lease duration) to extend leases while tasks run.
• Idempotency: enforce idempotency at the task level (idempotency key + handler-side dedupe) because duplicates are possible on retries and lease expiry.
• Throughput control: worker count and polling interval are per node. If you need a global rate limit across nodes, enforce it externally or in the handler.
• Clock skew: Redis uses server time for scores; keep node clocks in sync to avoid uneven dequeue/lease timing.
• Isolation: use distinct prefixes per environment/region/tenant to avoid cross-talk.

Checklist:

• Set WithDurableLease above p99 task duration (plus buffer).
• Enable WithDurableLeaseRenewalInterval for tasks that can exceed the lease duration.
• Keep task handlers idempotent; always use idempotency keys for external side effects.
• Tune WithDurablePollInterval based on desired responsiveness vs. Redis load.
• Scale WithMaxWorkers per node based on CPU and downstream throughput.

Example:

go run __examples/durable_queue_inspect/main.go -redis-addr=localhost:6380 -redis-password=supersecret -redis-prefix=go-worker
go run __examples/durable_queue_inspect/main.go -redis-addr=localhost:6380 -redis-password=supersecret -redis-prefix=go-worker -show-ids -peek=5

Sample output:

ready=3 processing=1 dead=0
ready IDs: 8c0f8b2d-0a4d-4a3b-9ad7-2d2a5b7f5d12, 9b18d5f2-3b7f-4d7a-9dd1-1bb1a3a56c55

DLQ replay example (dry-run by default):

go run __examples/durable_dlq_replay/main.go -redis-addr=localhost:6380 -redis-password=supersecret -redis-prefix=go-worker -batch=100
go run __examples/durable_dlq_replay/main.go -redis-addr=localhost:6380 -redis-password=supersecret -redis-prefix=go-worker -batch=100 -apply

Optional retention can be configured to prevent unbounded task registry growth:

tm.SetRetentionPolicy(worker.RetentionPolicy{
    TTL:        24 * time.Hour,
    MaxEntries: 100000,
})

Retention applies only to terminal tasks (completed/failed/cancelled/etc). Running or queued tasks are never evicted. Cleanup is best-effort: it runs on task completion and periodically when TTL > 0. If CleanupInterval is unset, the default interval is clamp(TTL/2, 1s, 1m). If MaxEntries is lower than the number of active tasks, the registry may exceed the limit until tasks finish.

Task lifecycle hooks can be configured for structured logging or tracing:

tm.SetHooks(worker.TaskHooks{
    OnQueued: func(task *worker.Task) {
        // log enqueue
    },
    OnStart: func(task *worker.Task) {
        // log start
    },
    OnFinish: func(task *worker.Task, status worker.TaskStatus, _ any, err error) {
        // log completion
        _ = err
        _ = status
    },
})

Tracing hooks can be configured with a tracer implementation:

tm.SetTracer(myTracer)

See __examples/tracing for a minimal logger-based tracer. See __examples/otel_tracing for OpenTelemetry tracing with a stdout exporter.

OpenTelemetry metrics

To export metrics with OpenTelemetry, configure a meter provider and pass it to the task manager:

exporter, err := stdoutmetric.New(stdoutmetric.WithPrettyPrint())
if err != nil {
    log.Fatal(err)
}

reader := sdkmetric.NewPeriodicReader(exporter)
mp := sdkmetric.NewMeterProvider(sdkmetric.WithReader(reader))
defer func() {
    _ = mp.Shutdown(context.Background())
}()

if err := tm.SetMeterProvider(mp); err != nil {
    log.Fatal(err)
}

See __examples/otel_metrics for a complete runnable example. See __examples/otel_metrics_otlp for an OTLP/HTTP exporter example.

Emitted metrics:

• tasks_scheduled_total
• tasks_running
• tasks_completed_total
• tasks_failed_total
• tasks_cancelled_total
• tasks_retried_total
• results_dropped_total
• queue_depth
• task_latency_seconds

Registering Tasks

Register new tasks by calling the RegisterTasks() method of the TaskManager struct and passing in a variadic number of tasks.

id := uuid.New()

task := &worker.Task{
    ID:          id,
    Name:        "Some task",
    Description: "Here goes the description of the task",
    Priority:    10,
    Ctx:         context.Background(),
    Execute: func(ctx context.Context, _ ...any) (any, error) {
        time.Sleep(time.Second)
        return fmt.Sprintf("task %s executed", id), nil
    },
    Retries:    3,
    RetryDelay: 2 * time.Second,
}

task2 := &worker.Task{
    ID:       uuid.New(),
    Priority: 10,
    Ctx:      context.Background(),
    Execute:  func(ctx context.Context, _ ...any) (any, error) { return "Hello, World!", nil },
}

if err := tm.RegisterTasks(context.Background(), task, task2); err != nil {
    log.Fatal(err)
}

Stopping the Task Manager

Use StopGraceful to stop accepting new tasks and wait for completion, or StopNow to cancel tasks immediately.

ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
defer cancel()

_ = tm.StopGraceful(ctx)
// or
// tm.StopNow()

Results

Subscribe to results with a dedicated channel per subscriber.

results, cancel := tm.SubscribeResults(10)

ctx, cancelWait := context.WithTimeout(context.Background(), 5*time.Second)
defer cancelWait()

_ = tm.Wait(ctx)
cancel()

for res := range results {
    fmt.Println(res)
}

Cancellation

You can cancel a Task by calling the CancelTask() method of the TaskManager struct and passing in the task ID as a parameter.

_ = tm.CancelTask(task.ID)

You can cancel all tasks by calling the CancelAll() method of the TaskManager struct.

tm.CancelAll()

Middleware

You can apply middleware to the TaskManager by calling the RegisterMiddleware() function and passing in the TaskManager and the middleware functions.

srv := worker.RegisterMiddleware[worker.Service](tm,
    func(next worker.Service) worker.Service {
        return middleware.NewLoggerMiddleware(next, logger)
    },
)

Example

package main

import (
    "context"
    "fmt"
    "time"

    "github.com/google/uuid"
    worker "github.com/hyp3rd/go-worker"
    "github.com/hyp3rd/go-worker/middleware"
)

func main() {
    tm := worker.NewTaskManager(context.Background(), 4, 10, 5, 3*time.Second, 30*time.Second, 3)

    var srv worker.Service = worker.RegisterMiddleware[worker.Service](tm,
        func(next worker.Service) worker.Service {
            return middleware.NewLoggerMiddleware(next, middleware.DefaultLogger())
        },
    )

    task := &worker.Task{
        ID:       uuid.New(),
        Priority: 1,
        Ctx:      context.Background(),
        Execute: func(ctx context.Context, _ ...any) (any, error) {
            return 2 + 5, nil
        },
    }

    _ = srv.RegisterTasks(context.Background(), task)

    results, cancel := srv.SubscribeResults(10)
    defer cancel()

    ctx, cancelWait := context.WithTimeout(context.Background(), 5*time.Second)
    defer cancelWait()
    _ = srv.Wait(ctx)

    for res := range results {
        fmt.Println(res)
    }
}

Versioning

This project follows Semantic Versioning.

Contribution Guidelines

We welcome contributions! Fork the repository, create a feature branch, run the linters and tests, then open a pull request.

Feature Requests

To propose new ideas, open an issue using the Feature request template.

Newcomer-Friendly Issues

Issues labeled good first issue or help wanted are ideal starting points for new contributors.

Release Notes

See CHANGELOG for the history of released versions.

License

This project is licensed under the MIT License - see the LICENSE file for details.