ACA CCE common mistakes
ACA CCE · Understand the traps and how to avoid them.
Treating Alibaba Cloud's elastic compute (ECS) as a drop-in synonym for a generic virtual machine and ignoring that instance families, images, and disks are provisioned as separate decisions.
Candidates carry over mental models from other clouds and assume a single 'create VM' object bundles networking, storage, and image; the exam separates these into distinct product concepts.
What to do instead
Study ECS as a composition of instance type (family/spec), image, system disk, data disk, and security group, and be able to state which component is changed to solve a given scenario.
Choosing an object storage (OSS) bucket ACL or RAM policy that grants public read access when the scenario only requires authenticated users to download files.
Candidates equate 'users need to download' with 'make it public' and forget that signed URLs and RAM roles satisfy the requirement without exposing the bucket.
What to do instead
For any 'users must access objects' scenario, default to private bucket plus either a signed URL or a RAM role/policy, and reserve public-read for genuinely anonymous content.
Confusing the roles of security groups and network ACLs when diagnosing why traffic is blocked between two ECS instances.
Both are described as filters, so candidates pick whichever appears first and do not trace the layered path (security group is instance-level and stateful; network ACL is subnet-level and stateless).
What to do instead
Memorize the layer, scope, and statefulness of each control, and practice tracing a packet from source instance to destination instance naming every control it crosses.
Mixing up the billing models (subscription, pay-as-you-go, and spot/preemptible) when asked which is cheapest for a fault-tolerant batch workload.
Candidates anchor on 'cheapest equals reserved/subscription' because that is true for steady-state workloads, and fail to apply the interruption-tolerance condition.
What to do instead
Build a decision table keyed on workload duration and interruption tolerance: long steady state maps to subscription, unpredictable short work to pay-as-you-go, and interruptible batch to spot.
Assuming a resource that was released or stopped stops incurring charges, including attached disks, elastic IPs, and snapshots.
Candidates reason about the compute instance only and overlook that storage, IP, and snapshot resources persist and bill independently of instance state.
What to do instead
For any cost question, enumerate every attached resource (disks, EIPs, snapshots, bandwidth) and state its billing behavior when the instance is stopped or released.
Answering AI/model-service questions by describing self-managed training infrastructure when the scenario calls for a managed inference or model service.
Candidates over-index on generic machine-learning lifecycle knowledge and do not map the scenario's operational constraints (no ops team, need for an endpoint) to the managed offering.
What to do instead
Classify each AI scenario by whether the customer owns model training, fine-tuning, or only inference, then match to the managed service tier that removes the operational burden described.
Treating the shared responsibility boundary as 'the provider secures everything' or 'the customer secures everything' rather than splitting by layer.
The concept is stated simply in study material, so candidates memorize a slogan instead of being able to place a specific control (patching the guest OS, physical datacenter access, customer data encryption) on the correct side.
What to do instead
Practice placing concrete controls on each side of the boundary until you can classify any named control without hesitation.
Selecting a support plan based on price alone when the scenario specifies response-time or channel requirements.
Candidates compare plans on cost and miss that the differentiators are response-time commitments, available channels, and access to technical account resources.
What to do instead
Compare support plans on the axes the exam tests: response time, channel, and included advisory services, and match the scenario's stated urgency to the tier.
Describing a VPC as if it were a single flat network and ignoring that vSwitches are zone-scoped subnets and route tables govern inter-vSwitch traffic.
Candidates map VPC to a familiar 'virtual network' and skip the zone/subnet structure, so they cannot answer questions about multi-zone placement or connectivity.
What to do instead
Draw a VPC as region-scoped containing zone-scoped vSwitches, then explain how route tables and gateways connect them, and verify placement questions against zone scope.
Answering elasticity and scaling questions with vertical scaling (larger instance) when the scenario describes variable, spiky demand.
Vertical scaling is the intuitive 'make it bigger' answer, and candidates skip evaluating auto scaling groups and load balancing as the horizontal pattern.
What to do instead
For any demand-variability scenario, first check whether horizontal scaling with an auto scaling group plus load balancer satisfies it, and reserve vertical scaling for single-instance limits.