Custom Kafka Producers and Consumers Development

Kafka clients in production are not just "send a message" and "receive a message". Often, incorrect configuration compromises business data consistency. We recently encountered a project where auto-commit and lack of idempotency caused up to 10% of orders to be lost during consumer restarts. After implementing manual offset management and an idempotent producer, losses were reduced to zero. The cost of such losses for an average online store can reach 2 million rubles per month. We specialize in developing reliable producers and consumers for high-load web applications in Java and Python. This article covers key problems, typical configurations, and practical cases—from idempotency setup to parallel processing with order preservation.

Problems We Solve

Message Loss During Broker or Consumer Failures

Without proper configuration of acks and idempotency, you risk losing critical business data. For example, with acks=1, a message can be lost if the leader fails before replication. We use acks=all with min.insync.replicas=2 and enable.idempotence=true. This eliminates loss even if one broker fails. Infrastructure savings by avoiding expensive solutions: up to 40%.

Order Violation and Duplication

Consumers without offset management may process the same message twice, and parallel processing destroys order. The solution: manual commit after processing and routing by key within a partition queue. On one project, this reduced validation errors by 30%.

Long Rebalance and Re-processing

When a consumer fails and a rebalance occurs, the group may re-read large amounts of data. We configure session.timeout.ms and max.poll.interval.ms so rebalance takes no more than 5 seconds, and we use DLQ to avoid blocking on "bad" messages.

How We Guarantee Delivery

According to Apache Kafka documentation, delivery guarantees are determined by the combination of acks, idempotency, and transactions. Let's look at key patterns.

Idempotent Producer with acks=all

For financial transactions, order events, and any critical stream, we set enable.idempotence=true and acks=all. This eliminates duplicates even with retries. Throughput impact is minor (5-10%), but reliability is significantly improved.

// Java — idempotent producer
Properties props = new Properties();
props.put(ProducerConfig.BOOTSTRAP_SERVERS_CONFIG, "kafka-1:9092,kafka-2:9092,kafka-3:9092");
props.put(ProducerConfig.KEY_SERIALIZER_CLASS_CONFIG, StringSerializer.class);
props.put(ProducerConfig.VALUE_SERIALIZER_CLASS_CONFIG, KafkaAvroSerializer.class);
props.put("schema.registry.url", "http://schema-registry:8081");

// Reliability
props.put(ProducerConfig.ACKS_CONFIG, "all");
props.put(ProducerConfig.ENABLE_IDEMPOTENCE_CONFIG, true);
props.put(ProducerConfig.MAX_IN_FLIGHT_REQUESTS_PER_CONNECTION, 5);
props.put(ProducerConfig.RETRIES_CONFIG, Integer.MAX_VALUE);
props.put(ProducerConfig.DELIVERY_TIMEOUT_MS_CONFIG, 120_000);

// Performance
props.put(ProducerConfig.BATCH_SIZE_CONFIG, 65536);          // 64KB batch
props.put(ProducerConfig.LINGER_MS_CONFIG, 5);               // Wait up to 5ms to fill batch
props.put(ProducerConfig.BUFFER_MEMORY_CONFIG, 67_108_864);  // 64MB buffer
props.put(ProducerConfig.COMPRESSION_TYPE_CONFIG, "lz4");

KafkaProducer<String, OrderEvent> producer = new KafkaProducer<>(props);

Sending with error handling:

public CompletableFuture<RecordMetadata> sendOrderEvent(OrderEvent event) {
    ProducerRecord<String, OrderEvent> record = new ProducerRecord<>(
        "order-events",
        event.getOrderId(),
        event
    );

    CompletableFuture<RecordMetadata> future = new CompletableFuture<>();

    producer.send(record, (metadata, exception) -> {
        if (exception != null) {
            if (exception instanceof RetriableException) {
                log.error("Retriable error, Kafka will retry: {}", exception.getMessage());
            } else {
                log.error("Fatal producer error for order {}: {}", event.getOrderId(), exception.getMessage());
                future.completeExceptionally(exception);
            }
        } else {
            log.debug("Sent to partition {} offset {}", metadata.partition(), metadata.offset());
            future.complete(metadata);
        }
    });

    return future;
}

Properties consumerProps = new Properties();
consumerProps.put(ConsumerConfig.BOOTSTRAP_SERVERS_CONFIG, "kafka-1:9092,kafka-2:9092,kafka-3:9092");
consumerProps.put(ConsumerConfig.GROUP_ID_CONFIG, "order-processor");
consumerProps.put(ConsumerConfig.KEY_DESERIALIZER_CLASS_CONFIG, StringDeserializer.class);
consumerProps.put(ConsumerConfig.VALUE_DESERIALIZER_CLASS_CONFIG, KafkaAvroDeserializer.class);
consumerProps.put("schema.registry.url", "http://schema-registry:8081");
consumerProps.put("specific.avro.reader", true);

// Disable auto-commit
consumerProps.put(ConsumerConfig.ENABLE_AUTO_COMMIT_CONFIG, false);
consumerProps.put(ConsumerConfig.AUTO_OFFSET_RESET_CONFIG, "earliest");

// Session timeout
consumerProps.put(ConsumerConfig.SESSION_TIMEOUT_MS_CONFIG, 30_000);
consumerProps.put(ConsumerConfig.HEARTBEAT_INTERVAL_MS_CONFIG, 10_000);

consumerProps.put(ConsumerConfig.MAX_POLL_RECORDS_CONFIG, 500);
consumerProps.put(ConsumerConfig.MAX_POLL_INTERVAL_MS_CONFIG, 300_000);

KafkaConsumer<String, OrderEvent> consumer = new KafkaConsumer<>(consumerProps);
consumer.subscribe(List.of("order-events"), new ConsumerRebalanceListener() {
    @Override
    public void onPartitionsRevoked(Collection<TopicPartition> partitions) {
        commitCurrentOffsets();
    }

    @Override
    public void onPartitionsAssigned(Collection<TopicPartition> partitions) {
        log.info("Assigned partitions: {}", partitions);
    }
});

Map<TopicPartition, OffsetAndMetadata> pendingOffsets = new HashMap<>();

try {
    while (!shutdown.get()) {
        ConsumerRecords<String, OrderEvent> records = consumer.poll(Duration.ofMillis(100));

        for (ConsumerRecord<String, OrderEvent> record : records) {
            try {
                processOrder(record.value());

                pendingOffsets.put(
                    new TopicPartition(record.topic(), record.partition()),
                    new OffsetAndMetadata(record.offset() + 1)
                );
            } catch (NonRetriableException e) {
                sendToDlq(record, e);
                pendingOffsets.put(
                    new TopicPartition(record.topic(), record.partition()),
                    new OffsetAndMetadata(record.offset() + 1)
                );
            }
        }

        if (!pendingOffsets.isEmpty()) {
            consumer.commitSync(pendingOffsets);
            pendingOffsets.clear();
        }
    }
} finally {
    consumer.close();
}

Why Manual Offset Management is Better Than Auto-Commit

auto-commit creates an illusion of simplicity but hides a risk: the offset is committed before processing. If the consumer crashes between commit and processing, the message is lost irretrievably. We always disable enable.auto.commit and commit manually after successful processing of the record. This is a standard practice for production systems.

Comparison of acks Modes

Comparison of Java and Python Clients

How to Organize Parallel Processing Without Losing Order

One poll thread + a pool of workers keyed by partition key is the classic pattern. We preserve order within a single order's events while parallelizing across orders.

Map<Integer, BlockingQueue<ConsumerRecord<String, OrderEvent>>> partitionQueues = new HashMap<>();
ExecutorService workers = Executors.newFixedThreadPool(12);

for (ConsumerRecord<String, OrderEvent> record : records) {
    int partitionIndex = record.partition() % NUM_WORKERS;
    workerQueues.get(partitionIndex).offer(record);
}

Turnkey Development Process

1. Analytics — define topics, message keys, format (Avro/JSON/Protobuf), consumer groups.
2. Implementation — write producers with idempotency, consumers with manual commit, DLQ.
3. Testing — unit tests with EmbeddedKafka, integration tests with Testcontainers, load testing.
4. Deployment — configure monitoring (consumer lag, errors) and alerts.

What's Included in the Work

• Repository with code (Java/Python) and configuration documentation.
• Docker images configured for staging/production.
• Deployment and rollback instructions.
• Training for the customer's team (1-2 hours).
• Code warranty: 3 months of free support.

Timeline

From 6 to 10 working days depending on complexity. Request a project evaluation — we'll respond within one day.

Common Mistakes (Checklist)

• [ ] enable.idempotence not enabled — duplicates on retries.
• [ ] auto-commit used — message loss on consumer failure.
• [ ] max.poll.records too large — long rebalance.
• [ ] No handling of NonRetriableException — message lost without DLQ.
• [ ] Mismatched serializer versions between producer and consumer — deserialization errors.

If you recognize your project in these issues, contact us. We have developed producers and consumers for 15+ projects handling up to 100,000 messages per second.

Guarantee: We fix deadlines in the contract and provide 3 months of post-delivery support.

Backend Development Services: Laravel, Node.js, Go, Django, PostgreSQL

On a production server at 3:14 AM, the Laravel Jobs queue stopped processing. 40,000 unprocessed jobs in Redis. Cause: worker crashed due to a memory leak in one of the Jobs (leak via a static variable in an Eloquent observer), supervisor didn't restart it because of misconfigured stopwaitsecs. This is not a hypothetical scenario — it's Tuesday. We analyzed such an incident on a project with 500 RPS load: diagnosis took 4 hours, fix — 20 minutes. So you don't lose money on downtime, we offer backend development services with a focus on production-grade reliability. We'll assess your project in 2 days.

Backend is what works when no one is watching. Or doesn't work. We guarantee you'll have the first option.

How do we ensure production-grade reliability from day one?

What we do correctly from day one

Service Layer over Fat Controllers. Controller receives HTTP request, validates it via Form Request, passes data to Service, returns response. Business logic in Service, not Controller. This sounds trivial, but most legacy projects have controllers with 500 lines and SQL queries inside.

Repository Pattern we use cautiously. If you just wrap Model::where(...) in a repository method — that's boilerplate without benefit. Repository is justified when: you need to abstract from the data source (DB + cache + external API) or when query logic is complex enough to isolate.

Jobs, Events, Listeners. Everything that can be async — make async. Sending email, PDF generation, external API sync, aggregate recalculation — into Queue. Laravel Horizon for queue monitoring in Redis: see throughput, failed jobs, processing time per queue.

How Octane handles high load

Laravel Octane with RoadRunner or Swoole keeps the app in memory between requests — removes bootstrap overhead (config loading, class autoloading) on each HTTP request. Gain: 3–8x on synthetic benchmarks, 2–4x on real applications. Important: no state between requests in static variables — that leads to exactly the incidents from the beginning. We use this in projects with >1000 RPS.

What to do about N+1 queries

N+1 is the most common cause of slow pages in Laravel apps. Standard story: page worked fine on dev with 10 records, on production with 10,000 — 8-second load.

Laravel Debugbar in dev environment shows the number of queries per page. More than 20 queries per page — signal for audit.

Model::preventLazyLoading(! app()->isProduction());

Telescope for profiling in staging: logs all queries, jobs, mail, notifications with time detail. Numbers: after implementing eager loading, page load time drops from 8s to 0.3s — 27 times faster.

PostgreSQL: indexes that are actually needed

PostgreSQL 14+ is the primary DB on all projects. We use PgBouncer + PostgreSQL combination. 10+ years experience, more than 50 backend projects, 5 years on the market.

How PostgreSQL helps avoid slow queries

Composite indexes for frequent WHERE + ORDER BY. If you have WHERE user_id = ? AND status = ? ORDER BY created_at DESC — you need (user_id, status, created_at DESC). A separate index on (user_id) doesn't help much with sorting.

Partial indexes. If 95% of queries go with WHERE status = 'active':

CREATE INDEX idx_orders_active ON orders (created_at DESC)
WHERE status = 'active';

The index is small, fast, covers the main load.

GIN indexes for JSONB and arrays. @> operator without GIN index — seq scan. With index — fast even on millions of rows.

GIN for full-text search. to_tsvector + GIN instead of LIKE '%query%'. LIKE without index is always seq scan. With pg_trgm extension and gin_trgm_ops — supports LIKE with index, useful for CRM search by partial match.

Connection pooling: why it's more important than it seems

Rails, Laravel, Django open a new connection to PostgreSQL for each PHP/Python process. With 100 workers — 100 connections. PostgreSQL starts degrading from 200–300 active connections — overhead on connection management becomes significant.

PgBouncer — connection pooler in front of PostgreSQL. Transaction pooling mode: connection to PostgreSQL is occupied only during a transaction, returned to pool between requests. 1000 application workers → 20–50 actual connections to PostgreSQL. This reduces latency by 40% and hosting costs by 30%.

Node.js with Fastify: when it's better than Laravel

Node.js is justified for:

• Realtime: WebSocket servers, Server-Sent Events, chat, live updates
• Streaming: large files, video, streaming data
• High I/O concurrency: many parallel requests to external APIs without heavy business logic
• Serverless: Lambda/Cloud Functions — Node.js starts faster than PHP

Fastify over Express: 2–3 times faster on benchmarks, built-in JSON Schema validation, better TypeScript support, plugin architecture.

Typical realtime architecture: Laravel — core business logic and REST API. Node.js + Socket.io or ws — WebSocket server. Laravel publishes events to Redis Pub/Sub, Node.js subscribes and broadcasts to clients. This separation allows scaling the WebSocket server independently of the main app.

Go: microservices and high load

Go we use for:

• High-load microservices (>10,000 RPS)
• Background workers with strict latency requirements
• DevOps tools and CLI
• gRPC services in microservice architecture

Goroutines — thousands of times cheaper than OS threads. 10,000 concurrent connections on Go is normal on one server.

But Go is not a silver bullet. Development is slower than Laravel: more boilerplate, no ORM at Eloquent level, error handling with if err != nil everywhere. Justified only when performance is a real requirement, not an assumption.

Django and Python backend

Django with DRF (Django REST Framework) — for tasks where Python is needed: ML pipelines, data processing, integrations with AI tools.

Celery for background tasks — similar to Laravel Queue but more complex to configure. Celery Beat for cron tasks.

Django ORM vs raw SQL: ORM is convenient for CRUD. For analytical queries with multiple JOINs, window functions, and CTEs — connection.execute() with raw SQL is more readable and predictable.

Redis: not just cache

Redis in our projects plays multiple roles:

Redis Cluster for horizontal scaling. Sentinel for automatic failover on standalone setups.

Deployment and infrastructure

Docker + docker-compose — standard for local development and production. Each service in a container: PHP-FPM/Octane, Nginx, PostgreSQL, Redis, Queue Worker, Scheduler.

CI/CD via GitHub Actions:

1. Run tests (PHPUnit / Pest, Vitest, Playwright)
2. Build Docker image
3. Push to Container Registry
4. Deploy: docker pull → docker-compose up -d on server, or Kubernetes rolling update

Zero-downtime deploy for Laravel: php artisan down --secret=TOKEN is not needed with proper configuration. Strategy: new container starts next to the old one, Nginx switches traffic after health check, old container stops.

Monitoring: Sentry for exception tracking with alerting in Slack/Telegram. Grafana + Prometheus (or Grafana Cloud) for metrics: CPU, memory, request rate, queue depth, database connection count. Alerts on: error rate > 1%, p99 latency > 2s, queue depth > 1000 jobs.

What's included in turnkey work

• Architecture design (API documentation, DB schema, service diagram)
• Implementation according to agreed specification with code review
• CI/CD, monitoring, alerting setup
• Load testing (k6, wrk) with report
• Handover of source code, access, deployment instructions
• Training of customer's team (2-3 sessions)
• Warranty support for 1 month after delivery

Timeline benchmarks

Pricing is calculated individually after analyzing load, integrations, and business logic. Contact us for a free audit of your current backend — get an optimization plan in 2 days. Request a consultation.