VK cuts the launch of a new recommendation feed from weeks to a single day

VK's AI team has published a report on Habr about what has changed in Discovery over the past few months. Discovery is the single infrastructure layer behind recommendations, search and advertising across all VK products. The headline result: a new recommendation vertical, meaning a separate feed or list for a specific use case, can now be launched in one day. This routine used to take weeks. About 90% of ranking tasks are solved through configuration, with no new code.

Config instead of weeks of development

Previously, every new feed meant the team had to assemble its own set of services, configure data pipelines, set up monitoring and separately build infrastructure for A/B tests. Now an engineer describes the vertical in a configuration file. It has four blocks:

• a catalog of the objects to be ranked; • candidate sources, where the objects to be shown come from; • a user profile used to select content; • optimization goals, meaning the metrics the model should push up.

From there, the platform deploys the services, data pipelines, monitoring and A/B loop itself. A person writes a description, not infrastructure.

This has produced 15 new verticals. Some are small use cases, such as comment ranking. Such small things used to be not worth launching: weeks of work did not pay off. When a launch takes a day, the economics change, and even narrow tasks get their own recommender.

A portal that keeps production from breaking

If config controls almost everything, a mistake in it hits millions of users. That is why VK built the Discovery Portal around the platform: a single point of control for configs, service topology and monitoring. It works like this:

• changes are validated before they are applied; • rollout is staged, and at each step metrics are compared against the old version; • if metrics drop, the system rolls the change back; • reverting to the previous version takes a single step.

All AI VK developers use the portal, and dozens of changes go through it every day. For engineers, this is the main practical takeaway: without such a safety net, launch speed quickly turns into a string of outages. A fast start and a fast rollback only work as a pair.

GPUs on a millisecond budget

A separate section covers GPUs in the ranking itself. A recommendation has to fit into a few milliseconds, or the feed will stall. The team implemented three mechanisms:

• Dynamic batching. The server briefly accumulates incoming requests and sends them to the GPU as a batch. A GPU processes many objects at once more efficiently than one at a time. • Tail latency control. Average speed says little: users remember the slowest responses. The team tracks the rare long requests specifically and caps them. • A fallback to a simpler model. If the heavy model cannot finish in the allotted time, a lightweight version returns the response. The feed gets a slightly less accurate result, but it does not end up empty.

All three techniques are familiar to engineers who run online inference. Here they have been packaged into a platform and applied to ranking in products with large audiences.

What improved in the metrics

VK reports the following figures after the changes:

• time in the VKontakte feed is up 11%; • friend additions from recommendations are up almost 2.7 times; • viewing time in VK Clips is up 22%; • on the VK Video home page, viewing time is up 20%.

Advertising also runs on Discovery. For community messages, the conversion from impression to target action rose 46%, while the cost of a target action fell 13%. In CPC website advertising (pay per click), target actions are up 56% and their cost is down 24%.

These are the company's own figures. The report does not break out the measurement period, A/B group sizes or the contribution of individual model improvements, so part of the gain may belong not to the platform but to the new models that moved onto it. Still, the claimed range, from +11% to +46%, suggests that the unified layer sped up not only launches but also the experiments themselves.

What's next

The next step is a full migration of search to Discovery. If it succeeds, recommendations, search and advertising will run on the same infrastructure, with a shared portal, monitoring and rollout rules.

The takeaway for other companies is simple. The more ranking tasks are moved into configuration, the cheaper each new vertical becomes and the more boldly teams try them. But a platform pays off only together with validation, staged releases and automatic rollback.