dureq

dureq

A distributed, Redis-native job scheduling and workflow orchestration system built in Go.

dureq is short for durable redis execution queue.

99.9% vibe-coded.

Inspired by Asynq, gocron, and river.

Why v1? — Rollback from the Actor Model

This repository is the v1 (current) implementation of dureq.

A v2 rewrite was attempted using the Hollywood actor framework, moving from v1's pull-based Redis Streams model to a push-based actor model with cluster-wide singleton actors (Scheduler, Dispatcher, Orchestrator) and per-node worker actors (WorkerSupervisor, WorkerActor, EventBridge, Notifier, Heartbeat, ClusterGuard).

The v2 actor model was rolled back to v1 for the following reasons:

1. Debugging was a nightmare

Actor boundaries destroy stack traces. When a handler panics or returns an error, the error is caught in recover(), converted to a string, and passed through 4-5 actor hops (WorkerActor -> WorkerSupervisor -> EventBridge -> Orchestrator) — each hop flattening the error to a plain string. By the time you see the error in Redis or logs, all type information, wrapped error context, and stack traces are gone. You're left with "render failed for user user-006: GPU out of memory" and no idea where in the actor chain things went wrong.

In v1, errors propagate through normal Go function calls. You get real stack traces, errors.Is() / errors.As() works, and you can set breakpoints anywhere in the call chain.

2. Delayed and lagging job dispatch

v2's dispatch path requires 7 actor hops for a simple job:

NotifierActor -> DispatcherActor -> WorkerSupervisorActor -> WorkerActor (spawn)
-> handler goroutine -> WorkerActor -> WorkerSupervisor -> EventBridgeActor

Each hop involves actor mailbox queuing, message serialization (protobuf for cross-node), and scheduling overhead. The Dispatcher singleton becomes a bottleneck — every job in the entire cluster must route through a single actor for node selection. Under load, the Dispatcher's mailbox backs up and dispatch latency spikes unpredictably.

v1's path is direct: client writes to Redis Stream -> worker polls stream -> executes. No intermediate actors, no singleton bottleneck, no message serialization between components on the same node.

3. Memory allocation explosion

v2 copies the job payload at minimum 2x on the dispatch path:

• Job -> DispatchJobMsg (protobuf allocation + payload byte copy)
• DispatchJobMsg -> ExecuteJobMsg (another protobuf allocation + payload copy)

Every job completion also allocates: ExecutionDoneMsg -> WorkerDoneMsg -> DomainEventMsg — each carrying the full JobRun struct and error strings through the actor chain. For high-throughput workloads, this creates significant GC pressure.

v1 passes job data by pointer within the same process. The only serialization boundary is Redis itself.

4. Operational complexity without proportional benefit

v2 introduced Hollywood cluster membership, protobuf code generation, singleton election via ClusterGuard, capacity reporting via WorkerStatusMsg (2s interval), and a complex actor wiring phase during startup. This added ~15 source files, protobuf definitions, and a make genproto build step — all to solve a problem (push-based dispatch) that v1 solves adequately with a simple Redis Pub/Sub notification that wakes the fetch loop.

The theoretical latency advantage of push-based dispatch (4-12ms vs polling) did not justify the complexity cost in practice, especially since v1's Pub/Sub wake mechanism already eliminates most polling latency.

Architecture

Client (Go)
  |
  |  enqueue via Redis
  v
+-----------------------------------------------------------+
|                         Redis                              |
|  Jobs . Schedules . Streams . Locks . Events . Results     |
+-----------------------------------------------------------+
  |                         ^
  v  Pub/Sub notify         |  persist results
+-----------------------------------------------------------+
|  Node                                                      |
|                                                            |
|  +-------------+  +-------------+  +---------+             |
|  |  Scheduler  |  | Dispatcher  |  | Worker  |             |
|  |  (leader)   |  | (direct)    |  | Pool    |             |
|  +------+------+  +------+------+  +----+----+             |
|         |                |              |                   |
|         v                v              v                   |
|  Sorted Set scan   XADD to Stream   XREADGROUP              |
|  -> dispatch due   + Pub/Sub wake   -> execute handler      |
|                                     -> ack / retry          |
|                                                            |
|  +---------------+  +----------------+                     |
|  | Orchestrator  |  | Monitor API    |                     |
|  | (leader)      |  | HTTP/gRPC/WS   |                     |
|  +---------------+  +----------------+                     |
+-----------------------------------------------------------+
         ^
         |  HTTP / WebSocket
+--------+--------+      +------------------+
|    durequi      |      |    dureqctl       |
| (Web Dashboard) |      | (Terminal UI)     |
+-----------------+      +------------------+

• Leader-elected coordinator: only the leader runs Scheduler + Orchestrator
• All nodes are workers: every node polls Redis Streams and executes handlers
• Direct Redis data flow: no actors, no RPC between components on the same node
• Single dependency: Redis only (Standalone, Sentinel, or Cluster)

Features

Core

• Flexible Scheduling — Immediate, one-time, interval, cron, daily, weekly, monthly with timezone support
• Workflow Orchestration — DAG-based task dependencies with automatic state advancement and child workflows
• Batch Processing — Optional onetime preprocessing, chunked parallel item execution, failure policies
• Group Aggregation — Collect individual tasks into groups, automatically flush and aggregate into a single job based on grace period, max delay, or max size
• Mux Handler — Pattern-based task type routing (order.*), global and per-handler middleware
• Priority Queues — User-defined tiers with weighted fair queuing
• Task Queue Rate Limiting — Distributed token bucket rate limiter per queue (Redis Lua-backed)

Reliability

• Retry Policies — Exponential backoff with jitter, error classification (retryable, non-retryable, rate-limited)
• Overlap Policies — Control concurrent runs for recurring jobs (allow all, skip, buffer one, buffer all, replace)
• Catchup / Backfill — Recover missed executions within a configurable window
• Schedule Jitter — Random offset on scheduled execution times to prevent thundering herd
• Exactly-Once Execution — Per-run distributed locks prevent duplicate handler invocation
• Unique Keys — Deduplication via unique keys with lookup and manual deletion
• Worker Versioning — BuildID-style safe deployments; version-mismatched work is re-enqueued for matching workers
• ScheduleToStart Timeout — Fail jobs that wait too long in the queue before starting
• Workflow Signals — Send asynchronous external data to running workflows

Operations

• Leader Election — Automatic failover with configurable TTL and epoch-based fencing tokens
• Dynamic Configuration — Runtime-updateable settings (concurrency, timeouts, retry) via Redis, per-handler overrides
• Archival / Retention — Automatic cleanup of completed/dead/cancelled jobs after configurable retention period
• Node Drain — Graceful worker drain: stop accepting new work while finishing in-flight tasks
• Audit Trail — Per-job state transition history stored as capped Redis Streams
• Progress Reporting — In-flight progress updates from handlers via types.ReportProgress(ctx, data)
• Distributed Locking — Redis-based locks with automatic extension

Observability

• Web Dashboard — React-based UI (durequi) with real-time WebSocket updates
• Terminal UI — Interactive TUI (dureqctl) for cluster monitoring
• Monitoring APIs — HTTP REST, gRPC, and WebSocket for full cluster observability
• Payload Search — JSONPath-based search across jobs, workflows, and batches
• Multi-Tenancy — Key prefix isolation per tenant

Getting Started

Prerequisites

• Go 1.25+
• Redis 7+

Quick Example

Server:

package main

import (
    "context"
    "encoding/json"
    "fmt"
    "time"

    "github.com/FDK0901/dureq/pkg/dureq"
    "github.com/FDK0901/dureq/pkg/types"
)

func main() {
    srv, _ := dureq.NewServer(
        dureq.WithRedisURL("redis://localhost:6379"),
        dureq.WithNodeID("node-1"),
        dureq.WithMaxConcurrency(10),
    )

    srv.RegisterHandler(types.HandlerDefinition{
        TaskType:    "email.send",
        Concurrency: 5,
        Timeout:     30 * time.Second,
        Handler: func(ctx context.Context, payload json.RawMessage) error {
            fmt.Printf("Sending email: %s\n", payload)
            return nil
        },
    })

    ctx := context.Background()
    srv.Start(ctx)
    // ... signal handling, srv.Stop()
}

Client:

package main

import (
    "context"
    "time"

    "github.com/FDK0901/dureq/pkg/dureq"
    "github.com/FDK0901/dureq/pkg/types"
)

func main() {
    cli, _ := dureq.NewClient(dureq.WithClientRedisURL("redis://localhost:6379"))
    defer cli.Close()

    ctx := context.Background()

    // Immediate job
    cli.Enqueue(ctx, "email.send", map[string]string{
        "to":      "user@example.com",
        "subject": "Hello",
    })

    // Recurring job (every 30 seconds, bounded)
    start := time.Now().Add(time.Minute)
    end := start.Add(24 * time.Hour)
    interval := types.Duration(30 * time.Second)
    cli.EnqueueScheduled(ctx, &dureq.EnqueueRequest{
        TaskType: "report.generate",
        Payload:  []byte(`{}`),
        Schedule: types.Schedule{
            Type:     types.ScheduleDuration,
            Interval: &interval,
            StartsAt: &start,
            EndsAt:   &end,
        },
    })
}

Scheduling

Overlap Policies

For recurring jobs, control what happens when a new execution is due while a previous one is still running:

Workflows

Define DAG-based workflows where tasks execute in dependency order:

cli.EnqueueWorkflow(ctx, types.WorkflowDefinition{
    Name: "order-processing",
    Tasks: []types.WorkflowTask{
        {Name: "validate", TaskType: "order.validate", Payload: payload},
        {Name: "charge",   TaskType: "order.charge",   DependsOn: []string{"validate"}},
        {Name: "reserve",  TaskType: "order.reserve",  DependsOn: []string{"validate"}},
        {Name: "ship",     TaskType: "order.ship",     DependsOn: []string{"charge", "reserve"}},
        {Name: "notify",   TaskType: "order.notify",   DependsOn: []string{"ship"}},
    },
})

The orchestrator validates the DAG (no cycles, all dependencies exist), dispatches root tasks first, and advances downstream tasks as dependencies complete. On retry, only failed tasks and their dependents are re-executed — completed tasks are preserved.

Batches

Process collections of items with optional shared preprocessing:

onetimeType := types.TaskType("image.download_template")

cli.EnqueueBatch(ctx, types.BatchDefinition{
    Name:            "birthday-cards",
    OnetimeTaskType: &onetimeType,
    OnetimePayload:  templatePayload,
    ItemTaskType:    "image.overlay_text",
    Items:           items,
    ChunkSize:       100,
    FailurePolicy:   types.BatchContinueOnError,
})

• Onetime preprocessing: shared setup task runs once before items begin
• Chunked dispatch: configurable chunk size provides backpressure
• Failure policies: continue_on_error (collect all results) or fail_fast (stop on first failure)
• Per-item results: individual success/failure stored and queryable

Group Aggregation

Collect individual events into groups and automatically aggregate them into a single job:

// Server: define aggregator + register handler
aggregator := types.GroupAggregatorFunc(
    func(group string, payloads []json.RawMessage) (json.RawMessage, error) {
        // Merge individual payloads into one
        return json.Marshal(merged)
    },
)

srv, _ := dureq.NewServer(
    dureq.WithRedisURL("redis://localhost:6379"),
    dureq.WithGroupAggregation(types.GroupConfig{
        Aggregator:  aggregator,
        GracePeriod: 3 * time.Second,  // flush 3s after last item added
        MaxDelay:    15 * time.Second, // force flush after 15s regardless
        MaxSize:     10,               // flush immediately when 10 items
    }),
)

// Client: enqueue individual items into a group
cl.EnqueueGroup(ctx, types.EnqueueGroupOption{
    Group:    "user-alice",
    TaskType: "analytics.process",
    Payload:  eventPayload,
})

Groups are flushed when any trigger fires: grace period elapsed (no new items), max delay reached, or max size hit. The aggregator merges all buffered payloads into a single job payload dispatched to the registered handler.

The flush uses an atomic Lua script, so multiple nodes can safely run the aggregation processor concurrently — only one node will claim each group's messages.

Error Handling & Retries

// Default: 3 attempts, 5s initial delay, 5m max, 2x multiplier, 10% jitter
types.DefaultRetryPolicy()

// Custom
&types.RetryPolicy{
    MaxAttempts:  5,
    InitialDelay: 10 * time.Second,
    MaxDelay:     10 * time.Minute,
    Multiplier:   3.0,
    Jitter:       0.2,
}

Error classification controls retry behavior:

return fmt.Errorf("transient failure: %w", err)              // retryable (default)
return &types.NonRetryableError{Err: err}                    // permanent, skip retries
return &types.RateLimitedError{Err: err, RetryAfter: 30*time.Second} // retry after delay

Handler Context

Inside a handler, access job metadata via context:

func handler(ctx context.Context, payload json.RawMessage) error {
    jobID    := types.GetJobID(ctx)
    attempt  := types.GetAttempt(ctx)
    taskType := types.GetTaskType(ctx)
    headers  := types.GetHeaders(ctx)

    types.ReportProgress(ctx, json.RawMessage(`{"percent": 50}`))

    return nil
}

Middleware

// Global — applies to ALL handlers
srv.Use(loggingMiddleware, metricsMiddleware)

// Per-handler — applied after global middleware
srv.RegisterHandler(types.HandlerDefinition{
    TaskType: "order.*",
    Handler:  orderHandler,
    Middlewares: []types.MiddlewareFunc{timingMiddleware},
})

Pattern-based routing: register "order.*" to match order.validate, order.charge, order.ship, etc. Use types.GetTaskType(ctx) inside the handler to dispatch.

Monitoring API

Endpoints

Jobs:

Workflows:

Batches:

Bulk Operations:

Audit Trail:

Node Drain:

Other:

WebSocket:

Event Types

Configuration

Server Options

Client Options

Dynamic Configuration

Settings can be updated at runtime without restarting nodes. Overrides are stored in Redis and polled every 10 seconds.

Global overrides:

# Via Redis directly
redis-cli HSET dureq:config:global MaxConcurrency 200
redis-cli HSET dureq:config:global SchedulerTickInterval 5s

Per-handler overrides:

redis-cli HSET dureq:config:handler:email.send Concurrency 20
redis-cli HSET dureq:config:handler:email.send Timeout 60s
redis-cli HSET dureq:config:handler:email.send MaxAttempts 5

These overrides take precedence over values set at server startup. Delete a field to revert to the startup default.

Project Structure

dureq/
+-- cmd/
|   +-- dureqd/             # Server daemon (HTTP/gRPC APIs)
|   +-- dureqctl/            # Interactive terminal UI (Bubble Tea)
+-- internal/
|   +-- server/              # Server setup, handler registry, options
|   +-- worker/              # Redis Streams consumer, goroutine pool, retry
|   +-- dispatcher/          # Job -> Stream dispatch, priority tier mapping
|   +-- scheduler/           # Leader-only schedule scanner, orphan detection
|   +-- workflow/            # DAG orchestrator, batch state machine
|   +-- aggregation/         # Group aggregation processor (leader-independent)
|   +-- store/               # Redis persistence (Lua scripts, streams, pagination)
|   +-- client/              # Go client SDK
|   +-- monitor/             # HTTP REST, gRPC, WebSocket APIs
|   +-- election/            # Redis-based leader election with epoch fencing
|   +-- lock/                # Distributed locking with auto-extension
|   +-- dynconfig/           # Dynamic configuration (runtime-updateable via Redis)
|   +-- ratelimit/           # Distributed token bucket rate limiter
|   +-- archival/            # Retention-based cleanup of old jobs
|   +-- cache/               # In-memory schedule and job cache
|   +-- lifecycle/           # Graceful shutdown coordination
+-- pkg/
|   +-- types/               # Public domain types (Job, Schedule, Workflow, Batch)
+-- examples/                # Runnable demos
+-- tests/
|   +-- integration/         # Integration tests (requires Redis)
+-- proto/
|   +-- dureq/monitor/v1/   # gRPC service definitions
+-- gen/                     # Generated protobuf code

Examples

cd examples/workflow_with_mux
go run main.go

Dependencies