Analytics services overlap at the edges, so classify the job before the product.
| Job | Start with |
|---|---|
| Serverless ad hoc SQL over S3 | Athena |
| Managed analytical warehouse | Redshift |
| ETL and shared metadata catalog | Glue |
| Govern a multi-source data lake | Lake Formation |
| Managed Hadoop/Spark cluster | EMR |
| Search and log analytics | OpenSearch |
| BI dashboards | QuickSight |
| Stateful stream processing | Managed Service for Apache Flink |
| Managed Kafka compatibility | MSK |
Athena cost and speed depend on bytes scanned: columnar Parquet/ORC, compression, partitioning, and selecting only needed columns matter. Redshift is optimized for repeated analytical workloads and can query S3 through Spectrum. Glue crawlers/catalog describe data; Glue jobs transform it. Lake Formation centralizes data-lake permissions across services.
For ingestion, Kinesis Data Streams retains records for custom consumers and replay; Data Firehose buffers and delivers to supported destinations with minimal operations. MSK fits Kafka APIs and ecosystem requirements. OpenSearch is for search, not a system-of-record replacement for DynamoDB; streams can feed an index.
ML service questions are mostly capability mapping: Rekognition images/video, Textract documents, Transcribe speech-to-text, Polly text-to-speech, Comprehend language, Translate translation, Lex conversational interfaces, and SageMaker custom ML workflows.
Exam lens: “serverless and query S3 occasionally” points to Athena; “warehouse with repeated BI joins” points to Redshift. Both can read S3, but the operating model differs.
Review card
- Optimize Athena by reducing bytes scanned.
- Glue catalogs/transforms; Lake Formation governs.
- Redshift is a warehouse; Spectrum extends queries to S3.
- Data Streams retains/replays; Firehose delivers.
- Match managed AI services by input and output modality.
Analytics services 在邊界上會 overlap，所以先分類 job，再選 product。
| Job | 優先想 |
|---|---|
| 對 S3 做 serverless ad hoc SQL | Athena |
| Managed analytical warehouse | Redshift |
| ETL 與 shared metadata catalog | Glue |
| 治理 multi-source data lake | Lake Formation |
| Managed Hadoop/Spark cluster | EMR |
| Search/log analytics | OpenSearch |
| BI dashboard | QuickSight |
| Stateful stream processing | Managed Service for Apache Flink |
| Kafka compatibility | MSK |
Athena 的成本/速度取決於 bytes scanned：要用 columnar Parquet/ORC、compression、partitioning，並只選需要的 columns。Redshift 適合重複 analytical workload，也可透過 Spectrum query S3。Glue crawler/catalog 描述 data；Glue job transform data；Lake Formation 中央管 data-lake permissions。
Ingestion 中，Kinesis Data Streams 保留 record、支援 custom consumer/replay；Data Firehose 以低維運 buffer/deliver 到 destination。MSK 對應 Kafka API/ecosystem。OpenSearch 是 search，不是 DynamoDB 的 system-of-record 替代品；可用 stream 喂 index。
ML service 題多為 capability mapping：Rekognition 圖像/影片、Textract 文件、Transcribe 語音轉字、Polly 字轉語音、Comprehend 語言、Translate 翻譯、Lex 對話介面、SageMaker custom ML workflow。
考試視角： 「Serverless、偶爾 query S3」是 Athena；「重複 BI join 的 warehouse」是 Redshift。兩者都能讀 S3，operating model 不同。
複習卡
- Athena 優化就是減少 bytes scanned。
- Glue catalog/transform；Lake Formation govern。
- Redshift 是 warehouse；Spectrum 把 query 延伸到 S3。
- Data Streams retain/replay；Firehose deliver。
- Managed AI service 按 input/output modality 選。