cert-bank

Guides

이 글을 한국어로 읽기

Amazon RDS vs Aurora vs DynamoDB: Three Questions That Pick the Database in an AWS Exam Scenario

Amazon RDS, Amazon Aurora and Amazon DynamoDB appear in almost every AWS exam, from Cloud Practitioner to Solutions Architect – Professional. The questions rarely ask "which database is best"; they describe an application and hide two or three clues in the wording. This comparison shows how to read those clues, which exams go deepest into which service, and the features that settle the close calls.

Study guide · Comparison · Amazon RDS, Amazon Aurora and Amazon DynamoDB

Amazon RDS vs Aurora vs DynamoDB: Three Questions That Pick the Database in an AWS Exam Scenario

Three questions, in this order

  1. How will the data be read? If every access is by a known key (get this customer's cart, this device's last readings), at very large or unpredictable scale, with millisecond latency and no servers to manage, the answer is DynamoDB. If the scenario mentions joins across tables, ad hoc reporting queries or an existing SQL schema, it is relational.
  2. Is there an engine to keep? A lift of an existing Oracle, SQL Server, Db2, MariaDB, MySQL or PostgreSQL database with minimal change points to RDS for that engine. Changing engines is a migration question (AWS DMS plus schema conversion), not a database choice.
  3. If it is MySQL or PostgreSQL, does it need what Aurora adds? Fast failover, up to 15 low-lag read replicas, storage that grows by itself, cross-Region replication with a promotable secondary (Aurora Global Database), or capacity that follows unpredictable load (Aurora Serverless v2). If the scenario stresses any of these, Aurora wins; otherwise standard RDS is a valid and cheaper answer.

One clue misleads people: "transactions". DynamoDB supports transactions across items and tables too, so transactions alone do not force a relational answer. Joins, flexible queries and referential structure do.

RDS, Aurora or DynamoDB: reading the clues in order
Is every access by a known key, at very large or unpredictable scale, with no servers to manage?
YES →
DynamoDB
↓ no
Must an existing Oracle, SQL Server, Db2 or MariaDB database keep its engine?
YES →
RDS for that engine
↓ no
Relational, with a promotable secondary in another Region?
YES →
Aurora Global Database
↓ no
Intermittent, unpredictable load on a relational database?
YES →
Aurora Serverless v2
↓ no
MySQL or PostgreSQL that needs fast failover, up to 15 replicas or self-growing storage?
YES →
Aurora
↓ no
OTHERWISE →
Standard RDS for MySQL or PostgreSQL: valid and cheaper
Ask from the top; the first yes decides. Joins and flexible queries mean relational, transactions alone do not.

How each exam uses them

We counted the questions in our own practice bank whose scenario involves at least one of the three services. The share differs, and so does the depth.

ExamShareWhat it tests
CLF-C0232 of 325 (10%)Which service matches a description; what AWS manages for you (patching, backups, failover) and what stays yours (data, access, schema).
SAA-C0367 of 395 (17%)Relational versus key-value; Multi-AZ versus read replicas; caching with ElastiCache or DAX; RDS Proxy; Aurora Serverless v2.
DVA-C0382 of 470 (17%)Almost all DynamoDB (72 questions): key design, Query versus Scan, GSIs and LSIs, consistency, Streams.
SOA-C0347 of 395 (12%)Operations: point-in-time recovery, Performance Insights (now part of CloudWatch Database Insights), failover behaviour, capacity modes and auto scaling.
DEA-C0127 of 395 (7%)DynamoDB as a source: partition keys, TTL, Streams, export to S3; Aurora zero-ETL to Redshift and federated queries.
DOP-C0231 of 395 (8%)Resilience and releases: global tables, Aurora Global Database, Multi-AZ DB clusters, schema changes during blue/green.
SAP-C0367 of 509 (13%)Multi-Region design and recovery objectives; single-writer versus multi-writer; cost of managed versus self-managed databases.

Signal words

The scenario says…Usually means
"key-value", "single-digit millisecond at any scale", "schema changes often", "serverless"DynamoDB
"microsecond reads" on an existing DynamoDB table, minimal code changeDynamoDB Accelerator (DAX)
"writes in every Region", "multi-active"DynamoDB global tables (Aurora Global Database has one writer Region)
"MySQL or PostgreSQL compatible", "storage grows automatically", "replicated across three AZs"Aurora
"RPO of seconds across Regions", "promote in minutes", relationalAurora Global Database
"intermittent, unpredictable load" on a relational databaseAurora Serverless v2
"Oracle", "SQL Server", "Db2", "keep the existing engine"RDS for that engine
"many short-lived Lambda connections exhaust the database"RDS Proxy (works with RDS and Aurora)
Going deeper: why Aurora fails over faster than standard RDS

In standard RDS with a Multi-AZ instance, the primary and the standby each have their own storage, kept in step by synchronous replication, and the standby serves no traffic. Aurora separates compute from storage. Every cluster has one shared storage volume, kept as six copies spread across three Availability Zones, and the writer and all replicas read from that same volume. A replica therefore holds no separate copy that must catch up: when the writer fails, a replica is promoted and points at storage that is already current. That is why Aurora replicas lag only slightly behind the writer, why they double as failover targets, and why "add Aurora replicas" answers both read scaling and availability in one step, while standard RDS needs read replicas for reads and Multi-AZ for failover.

Same words, different machinery

"Replica", "standby" and "global" mean different things in each service. Most wrong answers in this area come from mixing them up.

FeatureWhat it is for
RDS Multi-AZ DB instanceAutomatic failover inside a Region. The standby cannot be read.
RDS Multi-AZ DB clusterFaster failover plus two standbys that can serve reads (MySQL and PostgreSQL).
RDS read replicaRead scaling with asynchronous replication, so reads can be slightly stale; can be in another Region.
Aurora replicaRead scaling and failover target at once, behind a reader endpoint.
Point-in-time recoveryUndo a mistake by restoring to a second before it (RDS, Aurora and DynamoDB, up to 35 days). Replicas copy the mistake.

Two rules follow. Replicas and standbys never add write capacity: write limits are fixed with more IOPS, a bigger instance, sharding or a move to DynamoDB. And after a Multi-AZ failover, the endpoint name stays the same but its address changes, so applications must reconnect by name rather than caching the IP.

Beginner tip: create a DynamoDB table with partition key CustomerId and sort key OrderDate, add a dozen items for three customers, then fetch one customer's orders twice in the AWS CLI: once with query and a key condition, once with scan and a filter, both with --return-consumed-capacity TOTAL. The scan reads, and charges for, every item before the filter throws most of them away. That one comparison explains a large part of the DynamoDB questions in DVA-C03 and DEA-C01.

DynamoDB details that decide close calls

  • Capacity: on-demand for unknown or spiky traffic; provisioned with auto scaling for steady, predictable load.
  • Indexes: a global secondary index (GSI) adds a new partition key at any time but is eventually consistent; a local secondary index (LSI) keeps the partition key, must be created with the table and allows strongly consistent reads.
  • Changes and expiry: Streams for reacting to each item change; TTL for automatic deletion of old items without consuming write throughput; export to S3 for analytics without consuming table capacity.
Going deeper: why a table with spare capacity still throttles

DynamoDB stores a table's items in partitions chosen by hashing the partition key, and each partition can serve only a limited amount of reads and writes per second. The table's total capacity is spread across those partitions. If the key has few distinct values (a status field, today's date), most traffic hashes to the same partition and hits that partition's limit while the others sit idle, so the table throttles even though its overall metrics show spare capacity. Adaptive capacity raises throughput for busy partitions and can move hot items to partitions of their own, but no partition goes past its maximum of 3,000 reads or 1,000 writes per second. The fix the exams expect is a high-cardinality key, or adding a random or calculated suffix to spread writes (write sharding), not more capacity or another index.

Practise it

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Counts refer to our own practice bank. Engine support, Aurora features and DynamoDB limits change; check the current AWS documentation and each exam's official guide.