SplitShare — Backend and Observability Engineering
Academic project with independent reliability extensions
SplitShare began as a collaborative CSE370 academic project. The reliability extensions below are my independent work; the evaluation consists of self-reported controlled experiments, not a production or customer benchmark.
Problem Statement
A marketplace depends on both database access and downstream services. The engineering question was how to follow a request across those boundaries, distinguish slow or failed dependencies from healthy behavior, and make degraded database-backed pages report failure clearly.
Architecture
Browser → Django → MariaDB
Django → Flask → MariaDB
Both paths use the marketplace's MariaDB database. OpenTelemetry spans cover Django and Flask: HTTP calls carry trace context between the services, while request and database spans are exported over OTLP/HTTP.
Marketplace Database Flow
SplitShare originated as a collaborative CSE370 database-course marketplace built with Django, SQL and MariaDB/MySQL. The original course application was a team effort, distinct from my independently scoped reliability extensions.
- Browser to Django: Marketplace requests enter the Django application, which coordinates page rendering and database access.
- Django to MariaDB: SQL-backed data operations supply the marketplace pages. A database outage therefore affects what those pages can render.
- Django to Flask to MariaDB: The independently added coupon service introduces an HTTP boundary. Django calls Flask for coupon operations, and Flask accesses the database.
The database flow provides the academic foundation; the additional service boundary makes it possible to study how failures and trace context travel between services.
Independent Reliability Extensions
My scope was the Flask coupon service, cross-service instrumentation, fault experiments and database-outage handling—not sole authorship of the original course application.
- Extended a Django/MariaDB marketplace with a Flask coupon service and OpenTelemetry instrumentation, propagating trace context across HTTP requests and database operations using OTLP/HTTP.
- Designed controlled database-outage and service-boundary fault experiments covering slow responses, downstream errors and service unavailability; compared AI-generated diagnoses with traces, server logs and source code.
- Improved database-outage visibility by recording exceptions on request spans and returning HTTP 503 for degraded page renders, documenting the issue-to-merged-fix verification workflow.
Controlled Observability Evaluation
I used controlled fault injection to examine database outages and the Django-to-Flask service boundary, rather than presenting normal requests alone as evidence of reliability.
- Database outage: Examine request and database spans, server logs and the HTTP response when the database is unavailable.
- Service-boundary faults: Examine intentionally slow responses, downstream errors and coupon-service unavailability.
AI-generated Bluebox diagnoses were compared with traces, server logs and source code to evaluate whether the proposed root cause matched the injected fault and the application behavior.
Methodology: The extension used AI-assisted development with human review. I treated generated diagnoses and proposed changes as hypotheses to check against trace evidence, logs and source code.
The observations are limited to the documented controlled scenarios; they do not establish diagnostic accuracy or production performance gains.
Issue-to-Merged-Fix Verification
- Issue #4 was filed by a bot on . It records the database-outage visibility problem.
- Pull request #5 was merged on . The change records exceptions on request spans and returns HTTP 503 for degraded page renders.
- The documented verification workflow compares the reported problem with traces, server logs, source code and the merged HTTP behavior, rather than treating a generated diagnosis as proof of a fix.
Read the controlled evaluation notes for the self-reported setup, observations and verification workflow.
Academic Credits
- Theory faculty
- Rakin Bin Rabbani
- Lab faculty
- Course teammates
- Thanks to my course teammates for their collective contributions to the original academic marketplace. The course application was collaborative; the reliability extensions above are my independently scoped work.
These credits recognize the academic course project, not participation in or endorsement of the independent observability evaluation.
Independent Feedback Acknowledgment
Thanks to Andrew Norris for the opportunity to share independent Bluebox feedback. This was independent feedback, not employment, a partnership or a vendor endorsement.
Learning Context
A separate CSE370 learning resource is not yet public. The public project below is a practical study companion, not a published version of that course resource.
What I Learned
- Follow a failure across request spans, downstream HTTP calls and database operations before deciding where the fault originated.
- Treat HTTP status codes and recorded exceptions as part of observability: a degraded page should not look like a healthy request.
- Use controlled fault scenarios and human review to evaluate AI-generated diagnoses, keeping the evidence and its limitations explicit.