Junior Software Engineer interview prep: practice questions & salary data

An agent loop (also called a ReAct loop or think-act loop) follows a repeated cycle: the agent observes the current state/context, reasons or thinks about what to do, takes an action (e.g., calls a tool), receives an observation back, and repeats until a stopping condition is met. This is the fundamental pattern behind virtually all LLM-based agent frameworks.

A VPC is the foundational networking layer almost every other AWS resource sits in. The pieces that matter in interviews: it's Region-scoped with subnets pinned to one AZ each, public vs private subnets are defined by their route to an Internet Gateway vs a NAT Gateway, and security groups (stateful) vs NACLs (stateless) are the two filtering layers. Being fuzzy on subnets-are-AZ-scoped or on the IGW-vs-NAT distinction is a common tell of someone who hasn't actually built a network.

When a query filters on an indexed column and the selectivity is very high (only ~0.1% of rows match), the query planner will almost always prefer an index scan. An index scan navigates the B-tree to quickly locate the matching rows without reading the entire table. A sequential scan reads every page in the table and is preferred only when a large fraction of rows are expected to match, making random I/O from an index scan more expensive than sequential I/O. Hash join and merge sort scan are join strategies, not applicable to a single-table lookup.

Idempotency means that repeating the same operation multiple times yields the same server state as performing it once. PUT is defined as idempotent by the HTTP specification: sending PUT /users/1 with the same body replaces the resource each time, leaving the server in the identical state after the first call. This is unlike POST, which typically creates a new resource on each request and is therefore NOT idempotent.

Since Python 3.9, built-in collection types like list, dict, tuple, and set can be used directly as generic types (e.g., list[int]), eliminating the need to import List from the typing module. List[int] is the older style (still valid but requires from typing import List). list(int) is a runtime call, not a type hint, and Array[int] does not exist in standard Python.

The Circuit Breaker pattern monitors call failures to a downstream service and, after a threshold is crossed, 'opens' the circuit so that further calls fail immediately rather than waiting for a timeout. This prevents thread pools and other resources in the calling service from being exhausted while the dependency is down. Retries alone worsen the situation by consuming more resources. The Saga and Sidecar patterns address different concerns (distributed transactions and cross-cutting concerns, respectively).

In-context (working) memory is the portion of the LLM's context window that holds the current conversation history, including prior user and assistant messages. Recalling a message from five turns ago is a straightforward in-context retrieval — no external database lookup is required. Episodic memory refers to persisted, retrievable logs of past sessions, semantic memory to general world knowledge, and procedural memory to behavioural rules or skills.

True. TTL deletion is a background process DynamoDB runs itself; AWS's documented expectation is that items are typically removed within 48 hours of expiring, though the exact timing isn't guaranteed and depends on the table's size and activity. Because it's a system-initiated deletion rather than an application write, it doesn't consume the table's provisioned WCUs — this is why TTL is commonly used to cost-effectively prune session data, cache entries, or time-bounded records at scale. Until an item is actually deleted, it's still visible to normal reads, queries, and scans (unless the app filters it out), and it still counts toward storage size and billing during that window.

Because order_id is explicitly the sole primary key, there is no composite key, so 2NF is automatically satisfied — every non-key column depends on the full (single-column) key. The table is not in 3NF, however, because of transitive dependencies: order_id → customer_id → {customer_name, customer_email} and order_id → product_id → product_name. In both chains, an intermediate non-key attribute (customer_id or product_id) determines other non-key attributes, which is the classic 3NF violation. The fix is to decompose the table into separate customers and products tables, keeping only the foreign keys customer_id and product_id in orders.

Refresh tokens are credentials used exclusively with the authorization server's token endpoint to obtain new access tokens (and optionally new refresh tokens) when the current access token expires. They are never sent to the resource server. Only access tokens are presented to the resource server as Bearer tokens in the Authorization header.

pytest runs all three test functions. test_pass succeeds because 1+1 == 2 is True. test_fail raises an AssertionError because 1+1 == 3 is False, so it counts as a failure. test_skip calls pytest.skip(), which raises pytest.skip.Exception internally and marks the test as skipped rather than failed or passed. The final summary line therefore reads 1 passed, 1 failed, 1 skipped.

Horizontal scaling (scaling out) means adding more machines/instances to share the load, as opposed to vertical scaling (scaling up), which means upgrading the existing machine's CPU, RAM, or storage. Using a load balancer to distribute traffic across multiple servers is the canonical example of horizontal scaling. Database sharding and cache invalidation are separate, more specific techniques.

A post-generation output filter that combines regex patterns and Named Entity Recognition (NER) is the most direct mitigation at inference time because it inspects and sanitizes the actual text the model has produced before it reaches the user. Lowering temperature only changes randomness, larger models still memorize and reproduce PII, and rate-limiting reduces volume but does not block PII content.

EC2 Auto Scaling policies are driven by Amazon CloudWatch alarms, and those alarms can be based on EC2 instance metrics (like CPU utilization) or on metrics from integrated services such as an Application Load Balancer (e.g., RequestCountPerTarget). S3 events and IAM password-policy changes are not signals that an Auto Scaling Group listens to, and neither can directly trigger a scaling activity.

Durability is the 'D' in ACID. It ensures that committed transactions are permanently recorded, typically by writing to non-volatile storage (e.g., a write-ahead log or data files), so data survives crashes or power failures. Atomicity ensures all-or-nothing execution, Consistency ensures valid state transitions, and Isolation controls how concurrent transactions see each other's data.