Major upgrade to recommendation infrastructure
Migrating disparate recommendation algorithms to a single Discovery platform allowed VK to free up around 70,000 CPU cores and cut the product hypothesis testing cycle fivefold. While deploying a new recommendation service previously took weeks of meticulous engineering configuration, launching a typical vertical now fits practically into a single working day.
The platform unifies content ranking, search, and ad targeting tasks. Due to standardization, about 90% of tasks are resolved by modifying configuration files without writing new code, and onboarding of newcomers has accelerated threefold.
GPU ranking under strict Tier A constraints
A key technical breakthrough was implementing real-time GPU inference for heavy neural network models, where response latency is limited to mere milliseconds. The platform is classified under the Tier A reliability level, so the architecture was designed around potential failure scenarios:
- Dynamic batching: the system aggregates incoming single requests into short time windows, balancing GPU utilization with latency at higher percentiles.
- Guaranteed fallback: if the GPU cluster fails or the neural network times out, the service instantly falls back to a lightweight CPU model, keeping the interface operational.
- Resource sharing: GPUs are shared among multiple models concurrently, with idle time utilized for background batch processing.
- End-to-end optimization: feature reuse across different products, smart candidate caching, and profiling critical paths eliminated duplicate computations.
Discovery Portal: rollout safety and speed
To manage dozens of services, the developers launched an internal control portal. It took over routine operations that previously required deep immersion into system topologies:
- pre-validation of configurations before applying them;
- staged canary rollouts with automated evaluation of product metrics and instant rollback upon degradation;
- one-click rollback to any previous revision;
- a unified dashboard for monitoring, profiling, and telemetry collection across all teams.
Thanks to this automation, 15 new verticals were rapidly put into production, including local scenarios like comment section ranking, which were previously cost-ineffective to support with dedicated infrastructure.
Measurable impact across services
Technology unification directly impacted key metrics across VK products:
- VKontakte: time spent in the feed increased by 11%, friend additions from the recommendation block grew by 167%, and reach for original creators surged almost 5-fold. The Clips feed saw a 22% increase in watch time.
- VK Video and Clips: neural ranking implementation delivered +20% watch time on the main video showcase and +5.5% in short videos, alongside a 15% increase in shares.
- Odnoklassniki: +5.8% time spent in the feed and a 1.8% increase in daily active users.
- VK Music: track saves from recommendations grew by a third (+35%), while dislikes dropped by 19%.
- VK Ads: community promotion, mobile app campaigns, and lead generation forms migrated to the shared infrastructure.
The next phase in the Discovery platform's evolution will be fully migrating the ecosystem's internal search services to this unified tech stack.