A fast-growing real-time social platform was expanding beyond its original live-audio experience into personalized recommendations, anonymous web access, creator monetization, and integrations with multiple third-party services. At the same time, the engineering team needed to improve reliability without slowing product development.
That growth created several technical challenges. Recommendation queries used to personalize a live feed were taking more than 600 milliseconds (ms), too slow for a product where users expect immediate results. Deployment rollback behavior was inconsistent. Static abuse rules were becoming less effective as bad-actor behavior shifted. Backlogged task queues were difficult to recover without redeploying services. The team was also evaluating language-identification models without reliable ground-truth data.
The platform's architecture added another layer of complexity. Its Django backend coordinated PostgreSQL, DynamoDB, and other supporting systems. Each technology solved a specific scaling problem, but the growing number of dependencies increased the operational risk of failures propagating through the system.
The client tasked Six Feet Up with strengthening the platform’s technical foundation so the team could keep moving quickly as the product, user base, and operational demands grew.
Six Feet Up addressed performance and reliability issues across the full stack, from application code and data architecture to infrastructure, deployment, and machine-learning systems.
The team analyzed the full request path. The original similarity search scanned a large embedding dataset and exceeded 600 ms. By reducing the amount of data touched and revisiting the matching strategy itself, Six Feet Up brought the query down to approximately 4 ms and ultimately moved toward a lighter confidence-based scoring model that delivered comparable recommendation quality at lower computational cost.
Reliability work focused on shortening the distance between a production problem and a safe response. Six Feet Up implemented blue/green deployment safeguards tied to health-check alarms, automatic rollback, PagerDuty notification, and service-level monitoring. The team also established service-level objectives (SLOs) around availability, API latency, Celery success rates, and queue delays so engineers could respond to service degradation based on measurable thresholds.
Static rules were replaced with data-driven analysis using failure rates, historical baselines, and minimum-volume thresholds, allowing the system to better detect changing abuse patterns while limiting false positives.
Third-party services represented another source of operational exposure. Six Feet Up introduced feature-flag-controlled kill switches with service-specific fallback behavior, allowing engineers to disable a failing integration immediately without waiting for a new deployment.
Anonymous web listening required delivering chat, reactions, presence, and channel state to users who did not have accounts. Six Feet Up designed a parallel real-time delivery approach, stabilized guest sessions, introduced locking to prevent identity collisions, and corrected database-routing issues that caused stale state.
For monetization, Six Feet Up helped build the transaction and monitoring infrastructure behind paid reactions and a virtual currency economy, including settlement logic, payout processing, transaction history, concurrency controls, and production metrics. Those metrics quickly revealed balance-fetch failures as an early source of transaction problems and helped validate that locking thresholds could handle production traffic.
The team supported a language-identification pipeline processing more than 300 TB of audio across 500 GPUs. Queue-based scaling distributed work across available capacity and supported recovery from worker failures, while model-stability testing helped compare candidates when labeled ground truth was unavailable.
The deeper engineering behind the pipeline improvements is covered in Operational Efficiency for AI: Lessons from 500 GPUs.
.webp)
.webp)
The work produced measurable improvements in both platform performance and engineering operations.
Recommendation latency dropped from more than 600 ms to approximately 4 ms, making personalized results practical within a real-time feed while reducing the compute required to produce them.
Production failures became less expensive to manage. Automated rollback protected deployments when error rates crossed defined thresholds, while service-level monitoring gave the engineering team earlier warning of availability, latency, queue, and background-processing problems.
For failures originating with external providers, operational response time improved dramatically. The feature-flag-based circuit breaker reduced mitigation time from roughly 30 min to 30 sec, allowing the team to isolate an unhealthy dependency without deploying new code.
Anonymous web sessions also became substantially more reliable. A session-management defect that had caused approximately 15–20% session loss was reduced to near zero, removing friction from a key acquisition path that allowed prospective users to experience live content without creating an account.
Abuse detection became more adaptive to changing behavior, while the language-identification pipeline could distribute and recover large-scale processing workloads more reliably.
The result was a platform that could support new product capabilities without requiring the engineering team to trade development speed for operational stability.
Explore how Six Feet Up approaches scalable application architecture: sixfeetup.com/app-development.
Geospatial Analysis from 30 Minutes to Seconds
US-based Distribution Company
Six Feet Up, Inc.
11208 Windermere Blvd.
Fishers, IN 46037 - USA
Phone: +1 (317) 861-5948
Email: info@sixfeetup.com
