Architecting an EdTech Platform on AWS: Exam-Day Spikes, Video Cost, and Offline Sync
Education platforms have a traffic profile that breaks default architectures. A practical look at Cognito multi-tenancy, DynamoDB modeling, the assessment burst path, MediaConvert cost levers, and offline sync with AppSync.
Education platforms have a traffic profile that breaks most default architectures. For most of the term, load is flat and modest. Then an exam window opens and every student in a district hits the same endpoint within the same ninety seconds — and unlike an e-commerce spike, you cannot shed load. A student who cannot start their exam is an incident, not a lost conversion.
This article covers the AWS architecture decisions that matter for learning platforms, drawn from building them: identity and multi-tenancy, the assessment path under burst load, video delivery economics, offline sync for low-connectivity regions, and where compliance constrains the design.
The Traffic Shape
Three distinct load patterns coexist:
Pattern | Characteristic | Constraint |
|---|---|---|
Content browsing | Steady, cacheable, read-heavy | Cost per GB delivered |
Live classes | Scheduled, concurrent, latency-sensitive | Sub-3s glass-to-glass |
Assessments | Extreme burst, write-heavy, zero-tolerance | Cannot fail or throttle |
Designing for the average kills you on exam day. Designing everything for peak kills you on cost. The split matters more than any single service choice.
Identity and Multi-Tenancy
Most education platforms are multi-tenant by school or district, with role hierarchies (student, teacher, parent, school admin, district admin) and data isolation requirements that are contractual, not merely good practice.
Amazon Cognito user pools handle this reasonably well if you plan the structure early:
- One pool per environment, not one pool per tenant. Per-tenant pools seem cleaner until you need cross-tenant admin tooling or hit the pools-per-account limit.
- Tenant ID as a custom attribute (
custom:tenant_id), injected into the token via a pre-token-generation Lambda trigger. Every downstream authorization decision reads it from the token rather than doing a lookup.
- Cognito groups for role, mapped to IAM roles when you need fine-grained AWS resource access.
Schools frequently arrive with existing identity infrastructure — Google Workspace for Education, Microsoft Entra ID, or a SAML IdP. Federating these into Cognito is straightforward, but be aware that attribute mapping differs per provider, and you will need per-tenant IdP configuration. Build that as data, not as deployment configuration.
For parent-child account linking (a requirement in most K-12 products), model the relationship in your own data store. Cognito is not a relationship database, and attempts to encode guardianship in custom attributes get ugly quickly.
Data Modeling for Learning Content
The dominant access patterns in a learning platform are:
- Get a course's structure (modules → lessons → items)
- Get one student's progress across a course
- Get all students' progress for one teacher's class
- Get one student's submission history
- Get an assessment's questions in order
DynamoDB single-table design fits these well, because they're all key-based lookups rather than ad-hoc queries. A workable key structure:
PK SK AttributesTENANT#123#COURSE#456 METADATA title, descriptionTENANT#123#COURSE#456 MODULE#01#LESSON#03 content_ref, durationTENANT#123#USER#789 PROGRESS#COURSE#456 completed_items, last_positionTENANT#123#CLASS#321 ENROLLMENT#USER#789 enrolled_at, statusTENANT#123#USER#789 SUBMISSION#ASSESS#654#v2 answers_ref, score, graded_atTenant prefixing on the partition key gives isolation and distributes load across tenants naturally. A GSI inverting PK/SK covers "all students in class X" and "all classes for user Y."
Where DynamoDB stops being the right answer: content search and discovery. Full-text search over course catalogs, transcripts, and materials belongs in OpenSearch Serverless, populated via DynamoDB Streams. Don't try to scan your way there.
Aurora Serverless v2 is worth considering instead if your product does heavy relational reporting — gradebook analytics, cross-course cohort comparison, institutional dashboards. Many platforms end up with both: DynamoDB for the operational path, Aurora or Redshift for analytics.
The Assessment Path
This is the part that requires deliberate engineering.
The problem. 5,000 students start a timed exam at 09:00. Each submits answers continuously (autosave), then submits finally within a narrow window at the end. Any write failure loses student work, which is unrecoverable trust damage.
What works:
Pre-warm nothing, but choose services that don't need warming. API Gateway and Lambda handle the burst, but watch concurrency: 5,000 concurrent students at 1 autosave every 20 seconds is only ~250 requests/sec, which is fine. The danger is the final submission cluster, where several thousand requests can land in a few seconds. Set reserved concurrency on the submission function so it cannot be starved by other functions in the account, and set provisioned concurrency on a schedule if your exam windows are known.
Decouple acceptance from processing. The submission endpoint should do one thing: validate, write to DynamoDB (or SQS), return 200. Grading, analytics, notifications, and gradebook updates all happen asynchronously downstream via EventBridge or DynamoDB Streams. A student's submission must never wait on a grading pipeline.
Use conditional writes for idempotency. Students on flaky connections retry.
ConditionExpression: attribute_not_exists(submission_version) prevents a retry from overwriting a later answer with an earlier one.On-demand capacity for exam tables. Provisioned capacity with autoscaling reacts too slowly for a ninety-second spike — autoscaling operates on CloudWatch metrics with a multi-minute lag. On-demand absorbs the burst immediately. The cost difference is trivial relative to the failure mode.
Multi-question grading with Step Functions. Auto-gradeable items (MCQ, numeric, short-match) run inline. Free-text grading — whether by rubric heuristics or via Amazon Bedrock — belongs in a Step Functions workflow with retry and a human-review branch. Model the human-review escalation from the start; you will need it.
Video: Where the Money Goes
For most education platforms, video is the largest line item after headcount. Three distinct workloads, three different services:
Recorded lectures (VOD). Upload to S3, transcode with AWS Elemental MediaConvert into HLS with multiple renditions, serve through CloudFront. Two cost levers matter more than the rest:
- Ladder design. Generating 6 renditions when your users are 80% mobile wastes transcoding spend and storage. Instrument actual playback resolution before deciding.
- S3 Intelligent-Tiering on the source masters. Lecture recordings are accessed heavily for two weeks and rarely afterward, which is exactly the access pattern Intelligent-Tiering is built for.
Live classes. Amazon IVS is the pragmatic choice for one-to-many teaching (one instructor, many viewers) — managed, low-latency, no infrastructure to run. For genuinely interactive small-group sessions where students need two-way video, Amazon Chime SDK is the right tool. Teams that try to serve both patterns with one service end up over-paying for one of them.
Content protection. Signed CloudFront URLs with short TTLs, tied to the authenticated session. For high-value certification content, DRM through a partner integration. Note that signed URLs prevent casual sharing, not determined piracy — scope the effort to the actual threat.
Transcripts and accessibility. Amazon Transcribe on ingest produces captions (often a legal requirement) and, usefully, searchable text. Indexing transcripts in OpenSearch turns "find where the professor explained eigenvalues" into a solved feature rather than a roadmap item.
Offline Sync
For education app development targeting regions with intermittent connectivity, offline capability is not a nice-to-have — it determines whether the product is usable at all.
AWS AppSync with a GraphQL API and client-side caching handles the common cases. Amplify DataStore gives you local persistence and automatic sync with conflict resolution, and is the fastest path if your data model fits.
Where it gets hard: conflicts. AppSync offers three resolution strategies (optimistic concurrency, automerge, custom Lambda). For learning progress, automerge is usually wrong — if a student completes lesson 5 offline on their phone and lesson 3 offline on a tablet, you want both recorded, not last-write-wins.
The pattern that works: model progress as append-only events, not mutable state. Sync uploads events (
lesson_completed , quiz_attempted , video_position_updated ), and the server derives current state. Conflicts largely disappear because two devices adding different events is not a conflict.For content itself, pre-download course bundles to device storage and validate with checksums. S3 Transfer Acceleration helps in distant regions; CloudFront helps more.
Analytics and Engagement
Learning analytics — time-on-task, drop-off points, question-level difficulty, at-risk student flags — needs a separate path from the operational store.
App events → Kinesis Data Firehose → S3 (Parquet) → Glue Catalog → Athena │ └→ Lambda → DynamoDB (real-time counters)Firehose buffering into partitioned Parquet keeps Athena queries cheap. Real-time needs (a teacher watching live quiz responses) go through a Lambda consumer writing aggregates, not through the analytical path.
If you're implementing standards-based tracking, xAPI (Tin Can) statements map cleanly onto this event model, and using the standard vocabulary from day one saves a migration when an institutional customer asks for LRS interoperability.
Compliance Constraints
These shape architecture, so they belong in the design phase:
- FERPA (US): student education records. Access controls, audit logging, and defined disclosure rules. CloudTrail plus application-level access logs are your evidence.
- COPPA (US, under-13): verifiable parental consent before collecting personal information. This is a product flow, not just a checkbox — build the consent state into the account model.
- GDPR / DPDP Act (EU / India): lawful basis, data minimization, retention limits, and deletion rights. Deletion is the hard one — a student's data spans DynamoDB, S3, OpenSearch, analytics Parquet files, and video recordings. Build a deletion orchestration workflow early; retrofitting it across five stores is genuinely painful.
- Data residency. Institutional customers, particularly in the EU and India, increasingly require in-region storage. Region-per-tenancy is far easier to design for at the start than to add later.
- Accessibility (WCAG 2.1 AA): captions, keyboard navigation, screen-reader semantics. Frequently a procurement requirement for public institutions.
Cost Controls Worth Setting Up Early
- CloudFront in front of everything static. Origin requests are the expensive path.
- S3 lifecycle policies on raw uploads and old recordings.
- Lambda ARM64 (Graviton2) — roughly 20% better price-performance for the same code, and it's a runtime configuration change.
- DynamoDB on-demand for spiky tables, provisioned for predictable ones. The mix beats either alone.
- Cost allocation tags per tenant from day one. When a customer asks what their contract actually costs to serve, you want the answer available rather than reconstructible.
What I'd Do Differently
Three things I'd tell anyone starting an education app development project on AWS:
- Model progress as events from the beginning. It solves offline sync, analytics, and audit logging simultaneously. Retrofitting event sourcing onto mutable progress records is a multi-month project.
- Separate the assessment path physically. Its own Lambda functions with reserved concurrency, its own tables, its own alarms. It has different reliability requirements than the rest of the platform and should not share a failure domain with course browsing.
- Instrument video playback resolution before designing the transcode ladder. Most teams generate renditions nobody watches and pay for it every month.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article