Skip to content

SYSTEM Cited by 1 source

Amazon Kinesis Data Streams

Amazon Kinesis Data Streams is AWS's managed durable streaming substrate — a distributed, sharded, append-only log for real-time data ingestion. Comparable in architectural role to Apache Kafka (see systems/kafka): producers append records, consumers read by shard + sequence number, records are retained for a configurable TTL (24 h default, up to 365 days). Scales by shards; each shard supports a fixed ingest rate.

Stub page — minimal viable for the 2025-12-11 conversational- observability ingest. Expand as future Kinesis-specific sources land.

Role in the conversational-observability blueprint (2025-12-11)

Kinesis Data Streams is the buffer tier in AWS's telemetry-to-RAG pipeline on EKS. It sits between Fluent Bit's in- cluster telemetry collection and Lambda's embedding-generation compute. Responsibilities:

  • Absorb spikes — ingest bursts from the cluster are smoothed before hitting Lambda concurrency.
  • Durability during downstream outages — if Lambda / Bedrock / OpenSearch hiccup, records remain in Kinesis for their retention window and are re-processed on recovery.
  • Enable batching — Lambda's Kinesis event source supports batch sizes, which AWS's "Pro tip" explicitly recommends for cost-efficient embedding generation.

Kinesis vs SQS vs Kafka (positioning)

  • SQS (systems/aws-sqs) — message queue with at-least-once delivery and no replay. Good for task fanout, not for ordered replay or multi-consumer broadcast.
  • Kafka (systems/kafka) — self-managed / MSK; richer ecosystem (Kafka Streams, Connect); more operational work; higher theoretical throughput ceiling.
  • Kinesis Data Streams — AWS-managed Kafka-shaped primitive; tight integration with AWS-native consumers (Lambda, Firehose); fewer operational levers; sufficient for most AWS-resident streaming workloads.

Caveats

  • Stub page based on one source. Production design considerations (shard count, resharding, consumer fanout limits, enhanced fan-out consumers, on-demand vs provisioned capacity, KCL client library) not covered.
  • Encryption at rest / in transit via KMS is mentioned in the source but not expanded here.

Seen in

Last updated · 200 distilled / 1,178 read