AWS Lambda Interview Questions
Core Concepts
Q: How does Lambda work? What is cold start?
Lambda is serverless compute — you provide code, AWS runs it on demand. No servers to manage.
Execution model:
1.Event triggers Lambda (API Gateway, S3, SQS, CloudWatch Events, etc.)
2.AWS finds/creates execution environment (container)
3.Runs your handler function
4.Returns response or writes to output
5.Environment stays "warm" briefly (reuse for next invocation) or is destroyed
Cold start: First invocation (or after idle period) requires:
•Spin up execution environment (~100ms)
•Download deployment package
•Initialise runtime (Python/Node.js)
•Run initialisation code outside handler
Cold start latency: ~100ms-1s (Java worst, Python/Node.js better, SnapStart for Java).
Mitigation:
•Provisioned Concurrency: pre-warms environments (eliminates cold starts, extra cost)
•Keep deployment package small
•Move heavy initialisation outside handler (connection pooling, SDK clients)
•Use Python/Node.js for latency-sensitive functions
---
Q: Lambda limits and concurrency.
|---|---|
| Max memory | 10,240 MB |
|---|
| Max timeout | 15 minutes |
| Deployment package | 50MB zipped, 250MB unzipped |
| Container image | Up to 10GB |
| Concurrent executions | 1,000 (soft limit, can increase) |
| Ephemeral storage (/tmp) | 512MB default, up to 10GB |
Concurrency types:
•Reserved concurrency: Caps maximum concurrent executions for a function (throttles beyond limit)
•Provisioned concurrency: Pre-initialised environments (warm, no cold start, costs money)
---
Q: Lambda invocation types — synchronous vs asynchronous.
Synchronous (request-response): Caller waits for result.
Triggers: API Gateway, ALB, Lambda (SDK call), Cognito
Asynchronous: Lambda queues event, returns immediately (202). Retries 2x on failure.
Triggers: S3, SNS, CloudWatch Events
Event Source Mapping (stream/queue): Lambda polls source.
Triggers: SQS, Kinesis, DynamoDB Streams, Kafka
Error handling:
•Sync: Error returned to caller
•Async: Retries, then DLQ (Dead Letter Queue)
•Event source mapping: Retries until expiry, BisectOnError
---
Q: How do you connect Lambda to a VPC?
python
# Lambda VPC config - add in function configuration
VpcConfig:
SubnetIds: [subnet-xxx, subnet-yyy] # Private subnets
SecurityGroupIds: [sg-xxx]
# Lambda needs AWSLambdaVPCAccessExecutionRole
# For internet access from VPC: needs NAT Gateway
# For AWS services: use VPC Endpoints (avoid NAT)
Why VPC?: Access RDS, ElastiCache, or other private resources.
VPC caveat: Historically, attaching Lambda to a VPC added a significant ENI-creation cold start penalty. AWS re-architected Lambda's VPC networking in 2019 (Hyperplane ENIs, shared across functions in the same subnet/security-group combination), which removed most of that penalty — VPC-attached functions today are not meaningfully slower to cold-start than non-VPC ones in most cases. The remaining reason to avoid VPC when you don't need it is simplicity and avoiding NAT Gateway costs, not cold-start latency. Use VPC Endpoints for AWS service calls (S3, DynamoDB, etc.) so VPC-attached functions don't need a NAT Gateway just to reach other AWS services.
---
Q: Lambda layers and container images — when to use each?
Layers: Shared code/dependencies packaged as a ZIP.
•Max 5 layers per function
•Separate deployment from updates (update layer once, all functions benefit)
•Good for: shared libraries (pandas, requests), custom runtimes, security tools
Container Images (up to 10GB):
•Full control of environment
•Larger ML models, complex dependencies
•Familiar Docker workflow
•Better for ML inference, large dependency trees
Revision Notes
LAMBDA: serverless, event-driven, max 15min timeout, 10GB memory
COLD START: ~100ms-1s. Mitigate: Provisioned Concurrency, small packages, warm code outside handler.
CONCURRENCY:
Reserved: caps maximum (throttles beyond)
Provisioned: pre-warmed (no cold start, costs money)
Default limit: 1,000 concurrent (can raise)
INVOCATION TYPES:
Sync: API GW, ALB → wait for result
Async: S3, SNS → queued, retries 2x
Event Source Mapping: SQS, Kinesis → Lambda polls
LAYERS: shared code (max 5). Container: large deps/ML models.
VPC: pre-2019 added real cold-start time via ENI creation; largely fixed by Hyperplane ENIs since. Still use VPC Endpoints for AWS services to skip NAT Gateway.
