A deep comparison of the 3 leading streaming messaging platforms — Google Pub/Sub / Apache Kafka / AWS Kinesis. We break down pricing, performance, and features along the serverless / self-managed / managed axis.
| Aspect | Pub/Sub | Pub/Sub Lite | Kafka (MSK) | Kinesis Data Streams |
|---|---|---|---|---|
| Managed model | ◎ Serverless | ○ Zonal | △ Node mgmt | ◎ Shard mgmt |
| Ordering | Ordering Key | Partition | Partition | Shard + Partition Key |
| Push delivery | ◎ | — | — | — |
| Pull delivery | ◎ | ◎ | ◎ | ◎ |
| Exactly-once | ○ (Pull only) | ○ | ○ (Trans API) | — |
| Throughput ceiling | Unlimited | Reserved capacity | Unlimited | Per-shard (1000 msg/s) |
| Billing unit | $40/TB | Reserved capacity | Nodes + storage | Shard-hours + data |
| SLA | 99.95% | 99.5% | 99.9% | 99.9% |
| Service | Approx. monthly cost |
|---|---|
| Pub/Sub | ~$40 (data only) |
| Pub/Sub Lite | ~$5-10 (depending on reserved capacity) |
| Kafka (MSK m5.large × 3) | ~$500 |
| Kafka (Confluent Standard) | ~$1000+ (cluster cost) |
| Kinesis Data Streams (10 Shard) | ~$110 (shards) + $40 (1 TB data) |
| Feature | Pub/Sub | Kafka | Kinesis |
|---|---|---|---|
| Message retention | 7 days (default) | Unlimited (configurable) | 1 day (default) / up to 365 days |
| Replay | Snapshot / Seek | Offset rewind | Sequence number rewind |
| Schema Registry | Pub/Sub Schemas | Confluent Schema Registry | Glue Schema Registry |
| Stream Processing | Dataflow / Beam | Kafka Streams / Flink | Kinesis Data Analytics (Flink) |
| Connectors | Limited | Hundreds via Kafka Connect | Firehose / Lambda |
| Dead Letter | ◎ Topic | Custom | Custom |
| Aspect | Push | Pull |
|---|---|---|
| Support | Pub/Sub only | All 3 services |
| Subscriber implementation | Just an HTTPS endpoint | SDK + polling loop |
| Scaling | Auto, serverless | By subscriber count |
| Main use case | Cloud Run / Functions | Custom apps, Dataflow |
| Use case | Recommended | Why |
|---|---|---|
| Serverless + GCP | Pub/Sub | Push delivery + Eventarc |
| Large-scale streaming ETL | Pub/Sub + Dataflow | Serverless + unified Beam model |
| Need the Kafka ecosystem | Pub/Sub Lite or Confluent | Kafka-compatible |
| Multi-cloud + OSS | Kafka (Confluent) | Runs on AWS / Azure / GCP |
| AWS integration | Kinesis | Direct Firehose / Lambda / S3 hookups |
| IoT with hundreds of millions of devices | Pub/Sub | Unlimited scale |
| Long-term message retention | Kafka | Unlimited retention |
Pub/Sub vs Kafka vs Kinesis: which one should I choose?
Want serverless → Pub/Sub. OSS compatibility at scale → Kafka (Confluent / MSK). Deep AWS integration → Kinesis. Pub/Sub Lite is rapidly gaining traction as a Kafka alternative.
How does Pub/Sub Lite compare to Kafka?
Pub/Sub Lite is the cost-optimized variant designed as a Kafka alternative (~10x cheaper). It supports Kafka-compatible clients (Kafka Connect, etc.) but requires zone-level capacity reservation.
Which has the highest throughput?
Kafka is theoretically unlimited (just add nodes). Pub/Sub handles millions of msg/sec. Kinesis Data Streams is capped at 1000 msg/sec per shard.
How is ordering guaranteed?
Kafka guarantees order within a partition (same for Pub/Sub Lite). Pub/Sub uses an Ordering Key. Kinesis guarantees order via shard + partition key.
What about exactly-once delivery?
Pub/Sub supports it on Pull subscriptions with the exactly-once setting. Kafka uses the Transactional API. Kinesis requires dedup logic on the consumer side.
How does pricing work?
Pub/Sub charges $40/TB of ingested data. Pub/Sub Lite uses reserved capacity. Kafka (MSK) bills node-hours + storage. Kinesis charges shard-hours + ingest. At high volume, Kafka tends to be cheaper.
Do they all support a Schema Registry?
Yes, all three. Pub/Sub Schemas (Avro/Proto), Kafka Schema Registry (Confluent), and Kinesis uses Glue Schema Registry.
Is Push delivery unique to Pub/Sub?
Yes, it is a Pub/Sub differentiator. Kafka and Kinesis are pull-only. Push delivery is especially powerful as a front-end for Cloud Run / Cloud Functions.
Related articles - streaming comparisons
Pub/Sub 完全ガイド｜料金・Push vs Pull・Ordering Key・Kafka 比較 (GCP)
Google Cloud Pub/Sub の全機能解説。Push vs Pull、Ordering Key、Exactly-once、Dead Letter、Schema Registry、Pub/Sub Lite、AWS SNS/SQS / Kafka 比較、料金体系を網羅。
GCP vs AWS ストレージ・DB 徹底比較｜GCS/S3・BigQuery/Redshift・Spanner/DynamoDB (2026)
GCP と AWS のストレージ・データベースを徹底比較。Cloud Storage vs S3、BigQuery vs Redshift、Spanner vs DynamoDB / Aurora DSQL、Cloud SQL vs RDS、AlloyDB vs Aurora、Firestore vs DynamoDB、Bigtable vs DynamoDB を 2026 年最新版で網羅。
GCP vs AWS コンピュート徹底比較｜EC2/GCE・GKE/EKS・Lambda/Cloud Run・料金 (2026)
GCP と AWS のコンピュートサービスを徹底比較。Compute Engine vs EC2、GKE vs EKS、Cloud Run vs Lambda、App Engine vs Elastic Beanstalk、GPU/TPU、Arm 系 (Axion vs Graviton)、料金体系・Sustained Use Discount を 2026 年最新版で網羅。
Gemini vs GPT-4 vs Claude vs Llama 徹底比較｜LLM API 選び方・料金 (2026)
Google Gemini 2.0 / OpenAI GPT-4o / Anthropic Claude Opus 4 / Meta Llama 3.3 の徹底比較。性能 / コード / 推論 / マルチモーダル / 料金 / コンテキスト長 / GCP・AWS・Azure 経由利用を 2026 年最新版で網羅。
* All products are trademarks of their respective owners. Always check the official vendor pages for the latest pricing.
Practice with certification-focused question sets
Go to GCP Exam PrepNicheeLab Editorial Team
NicheeLab editorial team focused on data engineering and cloud certification learning. Content is structured around practical study needs and official exam domains.
Google Cloud Certification Roadmap (2026)
Choose your GCP certification path — Foundational, Associate...
CDL Cloud Digital Leader: Complete Exam Guide (2026)
Pass the Cloud Digital Leader exam — cloud business value, G...
GAIL Generative AI Leader: Complete Exam Guide (2026)
Pass the Generative AI Leader exam — Gemini, Vertex AI, Work...
Vertex AI Fundamentals for GCP Certs (2026)
Vertex AI basics every cert candidate needs — Workbench, Pip...
Associate Cloud Engineer (ACE): Complete Guide (2026)
Pass the Associate Cloud Engineer exam — Console, gcloud, pr...