Hemnet is Sweden’s largest property platform, serving millions of users who browse, save, and search through real estate listings every day. Behind the scenes, a GraphQL API handles the bulk of these interactions, powering everything from listing pages and search results to user accounts and saved properties. For years, this API ran on a monolithic GraphQL layer built on top of a Rails stack with PostgreSQL. As the platform grew, so did the pain points, particularly the difficulty of splitting concerns into smaller, independently maintainable chunks. But how do you introduce separation into a backend that was designed to operate as a single entity?
The idea of GraphQL federation had been floating around the engineering organization for some time. Different teams owned different parts of the domain, but they all fed into the same monolithic schema. Development takes time because it's just too much code to make sense of. A federation setup in which each team could own and evolve its own subgraph independently was the obvious architectural goal, but making it work on a live platform with millions of active users seamlessly was far from straightforward.
We had already conducted a small-scale experiment a few years earlier with a very small subset of the users using Apollo. We thought about using it again, just building on the previous work.
At Hemnet, we value OSS a lot, one of our top criteria for choosing a new provider is asking: “Are they Open Source? Do they provide self-hosting options?”, and this is an important part of our culture, because we believe that a tool built by the collective intelligence is way better than to rely on a single person's brilliance, we are not afraid to open issues, send PRs, and help the ecosystem when needed.
The combination of licensing concerns with Apollo, and this desire for an open-source-first approach led the team toward the Hive Platform. The results of the migration were surprising!
The project kicked off in November 2025. A small, focused group was created with a very specific goal: to create the platform for the GraphQL federation. The ambition was clear: take what has been learnt so far, replace the Apollo-based infrastructure with the Hive-based one, integrate it into Hemnet’s existing APIs, and do it all without any user-visible disruption.
So the project was divided into three phases:
It is important to note that we did not immediately split the schema into multiple subgraphs.
The first phase focused purely on replacing the routing layer while keeping the monolithic GraphQL schema structurally unchanged. The entire API was exposed as a single federated subgraph behind the Hive Gateway.
This allowed us to validate performance, stability, and schema governance in isolation without mixing infrastructure migration with domain refactoring. By separating infrastructure replacement from architectural refactoring, we dramatically reduced the blast radius of the migration.
Initially, we wanted to replace the current DNS routes for the gateway directly, but we then realized some of our legacy infrastructure wasn't able to handle some queries, so we needed to phase the rollout using a canary approach. The gateway was implemented behind an edge worker acting as a proxy, and diverting a percentage of the user traffic to the new router or the old API depending on a bucketing strategy.
Installing Hive was simple, so simple we actually second-guessed if we had done it correctly at first, it's just a Docker image with a TypeScript configuration, we load it all and it's there, working. And this setup was already familiar to us since it also mirrored somewhat of the previous tests we did with the Apollo router image, which made the transition even simpler since we already knew some of the configuration options we needed to set. The main differences were related to integrations with our monitoring providers and internal and external authorization, which were very easily solved by implementing custom handlers in Typescript and keeping the request flow as close to the original as possible.
Authentication and Header Propagation
One of the most critical technical requirements during the migration was preserving our existing authentication and authorization behavior.
Our legacy GraphQL API relied on session cookies, internal service tokens, and custom headers used for downstream authorization decisions.
When introducing Hive Gateway into the request path, we ensured complete header transparency. We implemented custom request handlers in the gateway configuration to explicitly forward authentication headers, preserve cookies, and maintain OpenTelemetry trace propagation.
Rather than centralizing authorization in the gateway, we deliberately kept it as a transparent routing layer. Each downstream service continued enforcing authorization rules exactly as before.
This minimized risk and avoided subtle security regressions during the migration.
One of the defining characteristics of this migration was the canary deployment strategy, executed via an edge worker acting as a traffic proxy. Rather than flipping a switch and routing all of Hemnet’s GraphQL traffic through the new gateway at once, we implemented a percentage-based rollout controlled at the edge.
The rollout was complicated by the fact that Hemnet operates two load balancers: one internal (handling server-side rendering) and one external (serving apps and other services). There was no straightforward way to gradually redirect on the external load balancer, so the worker became the control plane for the canary. This approach meant the team could roll back instantly by simply adjusting the percentage at the edge, without touching any infrastructure configuration,which can take time to propagate through the environment.
The rollout followed a careful progression. After initial testing in staging during the first two weeks, the team pushed the gateway to production on the next day it was at 50%. Two days later, it increased to 80%. By the end of the week, it was at 100%. The entire public-facing traffic from our previous API gateway to Hive Gateway took three days. Soon after we also switched our internal API requests to the same gateway and all the billions of requests Hemnet served were now going through our new structure.
Handling Legacy Query Incompatibilities
During rollout, we discovered that a small subset of queries were valid under our legacy Ruby GraphQL implementation but did not comply with Hive’s stricter validation rules.
Rather than blocking the migration, we used the Cloudflare Worker to selectively inspect operation names and route incompatible queries to the legacy Ruby API while all other traffic flowed through Hive.
This gave us time to coordinate with client teams and update non-compliant queries without delaying the overall rollout.
Once those queries were updated, the exclusions were removed and traffic was fully consolidated under Hive Gateway.
The single biggest concern going into this migration was latency. Adding any network layer to a production request path is a risk, especially for a platform where page load times directly affect user engagement and, ultimately, real estate transactions. We expected that moving from a compiled Rust-based Apollo Router to a Node.js-based one with Hive Gateway would come with a measurable performance cost. Oh! We were wrong.
When the gateway reached 50% of production traffic, the results were immediately encouraging. Latency metrics showed GraphQL request latency holding steady at 75ms, and 305ms for p99. These numbers were effectively identical to what we had been seeing with the previous API. To not say we didn't have any increase, our edge worker added a negligible latency on top of it all.
Internal analysis confirmed the picture. Traces showed that the Hive Gateway accounted for the same percentage of time in a request as the previous calls. The average time the gateway itself added to a request hovered well below the 100ms mark, and this included the full round trip of parsing, planning, executing against the monolithic schema, and returning the response.
This was a particularly notable finding because we had explicitly discussed that if the Node.js-based gateway proved too slow, we could fall back to the Apollo Router or explore Hive’s Rust-based query planner. In practice, neither was necessary. We also noted that with the Rust-based query planner feature from Hive, there was potential to reduce the average response time even further below the 60ms mark, but this optimization was not yet critical enough to prioritize. We might still do it in the future.
Beyond latency, the resource footprint was another pleasant surprise. We expected the Node.js gateway to consume significantly more CPU and memory since it's more resource-intensive than a Rust-based one. Instead, we found the Hive Gateway running with less than 30% resource usage than Apollo Router, and it was holding tens of thousands of requests per minute. The resource efficiency reflected strong engineering from the Hive team. The gateway does use more resources than the Rust-based counterpart and these numbers vary depending on the amount of traffic, but it's really not significant enough to say it's worse or to change it.
The observability story was one of the strongest arguments for Hive. The gateway natively integrates with open telemetry, attaching trace correlation to every request. Validation errors and execution errors are tagged on the active span, making it trivial to correlate GraphQL-level failures with infrastructure-level traces. We also built custom plugins for error logging and for worker metrics, ensuring that every layer of the request pipeline was visible.
On the Hive side, the Hive Console provided a centralized view of schema changes, operation performance, and client usage patterns. We found particular value in being able to see which operations were failing the most, which clients were sending the most traffic, and how the schema was evolving over time. Developers now actively use Hive Insights and the schema explorer to understand type evolution and assess whether a proposed change would introduce breaking behavior, visibility that simply did not exist in our previous setup.
Schema Governance in CI/CD
Even before introducing multiple subgraphs, we treated schema management as a first-class concern.
Our monolithic GraphQL schema is built and extracted during CI execution. The pipeline validates the schema locally, checks for breaking and dangerous changes against the Hive registry, and blocks the build if a breaking change is detected.
Only validated schemas are published to the registry.
Schema validation runs directly at the pull request level. Developers can immediately see whether a field removal, nullability change, or type modification would break existing operations and which clients are affected.
This PR-level feedback loop significantly improved developer confidence when evolving the schema and was widely appreciated across the engineering organization.
The migration to Hive Gateway is now complete, with 100% of Hemnet’s public-facing and internal GraphQL traffic flowing through the Gateway. The quantifiable achievements speak for themselves: the website is as fast as it was before, even with an added network layer. There was no perceptible latency regression for end users.
The developer experience improved meaningfully as well. With Hive’s schema registry integrated into CI, developers now get immediate feedback when a schema change would break existing operations. The Hive Console provides visibility into operation performance, client usage patterns, and schema evolution that was previously scattered across multiple tools or simply unavailable.
But the migration was not without its challenges, and being honest about them is what makes this case study worth reading.
The complexity of integrating legacy data and APIs was underestimated. Publishing the GraphQL schema from CircleCI required solving dependency issues, and we had to create lightweight alternatives to avoid booting the entire application just to extract the schema. We also underestimated how strict Hive's validation was with our previous schema, which was using a more relaxed approach. This meant we had to adjust and review several types to make them on par with how strict our new validation rules are now.
The organizational effort required for federation management and team alignment on schema principles was substantial. Convincing people across multiple teams about what federation is and what it is not, and why it matters, consumed more time than any technical challenge. The project needed buy-in from multiple people at multiple organization levels, we had to bring this to the attention of all the developers in the company at once, because we were effectively changing their development platform, and this was significant work.
The cost of knowledge sharing cannot be overstated. Federation introduces new concepts (subgraphs, supergraphs, schema registries, composition) that not every backend developer is familiar with. We had to invest time in documentation, town hall presentations, and architectural boards until we were sure most of the devs had understood the project.
Transform your GraphQL infrastructure like Hemnet. Whether you’re replacing legacy gateways, introducing federation, or improving schema governance, Hive gives you the tools to evolve safely, without disrupting production.
Our migration to Hive Gateway is a story of pragmatic engineering execution. A small team replaced the core routing layer of Sweden’s largest property platform in under two months, with zero user-visible downtime and no latency regression.
This highlights how the partnership between Hemnet and The Guild demonstrated what’s possible when an open-source ecosystem provider and an enterprise consumer collaborate directly: real-time feedback loops, rapid iteration on configuration issues, and a shared investment in getting the details right.
As the company scaled, the need for a robust, flexible, and efficient API architecture became paramount, leading to the adoption of GraphQL with Hive as their central API management solution.
Sound.xyz is revolutionizing the music industry by addressing two critical challenges: the concentration of streaming revenue among top artists and the inadequate compensation per stream.