1) Problem Clarification / Làm rõ bài toán

EN

A distributed task queue executes background jobs asynchronously:

send email
resize images
process payments
run ML workloads
schedule jobs
retry failed tasks

VI

Distributed task queue xử lý job nền không đồng bộ:

gửi email
resize ảnh
xử lý thanh toán
chạy ML
job theo lịch
retry job bị lỗi

2) General Architecture / Kiến trúc tổng quan

Producer → Queue/Broker → Workers → Result Store / Callback

VI

Producer → Queue → Worker → Lưu kết quả hoặc callback.

3) Push vs Pull Model

EN

Push-based (SQS, Redis queues)

Broker delivers messages to workers.

Pull-based (Kafka)

Workers poll for new messages.

VI

Push (SQS/Redis)

Queue đẩy job sang worker.

Pull (Kafka)

Worker chủ động kéo job.

4) Core Requirements

EN

✔ scalable workers
✔ retry logic
✔ idempotency
✔ dead-letter handling
✔ visibility timeout
✔ task timeouts
✔ ordering rules

VI

✔ scale worker
✔ retry
✔ idempotent
✔ dead-letter
✔ visibility timeout
✔ timeout job
✔ thứ tự xử lý

PART A — REDIS/ CELERY / SIDEKIQ STYLE QUEUES

5) Redis Queue Model

EN

Redis lists/zsets store tasks:

LPUSH queue task
BRPOP queue

Workers pop tasks and process them.

VI

Redis list/zset lưu job:

LPUSH → push
BRPOP → worker lấy job

6) Celery Architecture

EN

Components:

Broker: Redis / RabbitMQ
Worker
Beat: scheduler
Result backend: Redis / DB

Supports:

ETA, countdown
retry, exponential backoff
task chaining

VI

Celery gồm:

Broker
Worker
Beat (scheduler)
Result backend

Hỗ trợ:

ETA
retry/backoff
chain, group, chord

7) Sidekiq Model

EN

Ruby workers, Redis queue.

Features:

concurrency
retry set
scheduling via sorted sets

VI

Sidekiq dùng Redis:

đa luồng
retry thông minh
scheduled job dùng sorted set

PART B — AWS SQS STYLE QUEUES

8) SQS Key Concepts

EN

At-least-once delivery
Visibility timeout
Dead-letter queue
Long polling

VI

At-least-once delivery
Visibility timeout
Dead-letter queue (DLQ)
Long polling

9) Visibility Timeout

EN

When worker receives a message:

message becomes invisible for N seconds
if worker fails or crashes → message becomes visible again (retry)

VI

Worker nhận job → job bị ẩn tạm thời.
Nếu worker chết → job xuất hiện lại → retry.

10) Dead Letter Queue (DLQ)

EN

After X retries → move message to DLQ.

Allows investigation of failing tasks.

VI

Sau X lần retry → đẩy vào DLQ để điều tra thủ công.

PART C — KAFKA WORKER MODEL

11) Kafka Worker Characteristics

EN

✔ partition-level ordering
✔ consumer groups for parallelism
✔ at-least-once or exactly-once
✔ offset management
✔ no automatic retry → must implement logic

VI

✔ ordering theo partition
✔ consumer group để scale
✔ at-least-once hoặc exactly-once
✔ quản lý offset
✔ không retry tự động → phải tự code

12) Exactly-Once with Kafka

EN

Combine:

idempotent producer
transactional consumer
atomic offset commit

VI

Dùng:

producer idempotent
consumer transactional
commit offset nguyên tử

13) Handling Failures in Kafka Workers

EN

retry topic
DLQ topic
backoff logic
poison message protection

VI

Kafka xử lý lỗi bằng:

retry topic
DLQ topic
backoff
tránh poison message

PART D — TASK SCHEDULING

14) Scheduling Jobs

EN

Techniques:

cron scheduler
delayed queue (Redis sorted-set)
SQS delay seconds
Kafka delay topic

VI

Cách lên lịch job:

cron
Redis sorted-set
SQS delay
Kafka delay topic

15) Idempotency — CRITICAL for Safety

EN

Workers must avoid double-processing:

Techniques:

idempotent operation
job deduplication key
database unique constraints
Redis SETNX

VI

Phải đảm bảo job không xử lý 2 lần:

logic idempotent
dedupe key
DB unique
Redis SETNX

PART E — WORKER SCALING

16) Worker Auto-Scaling

EN

Base scaling on:

queue depth
message age
processing time
CPU usage

VI

Scale worker dựa trên:

số job trong queue
job bị treo lâu
thời gian xử lý
CPU

17) Worker Concurrency Control

EN

Use:

per-worker concurrency
per-task concurrency limits
rate limit

VI