Skip to content
Blog

· 5 min read

The Socialler years

I joined Socialler early in my career and left a senior. The network, and the people on it, are what grew me.

From March 2023 to September 2026 I worked at Socialler, a consumer social network. I did not arrive as the senior backend engineer. I arrived earlier than that, and Socialler is where I became one. It was the most rewarding stage of my career. My skills moved more in those three and a half years than in any stretch before them, and I am very grateful to the company for the room to do it.

Socialler app icon, a white S on purple.

The work, by the end, was the API and the web, iOS, and Android clients on one GraphQL schema. The path there was a sequence of systems. Each one was a little past what I knew how to hold, which is why they taught me anything.

Identity, the first real boundary

The early lesson was that authentication does not belong inside every handler. Identity became its own service, a hexagonal gateway in front of Logto. Login is an OIDC authorization-code flow with PKCE. Multi-factor authentication, passkeys, and a session per device live there. Native clients keep the refresh token in the Keychain or the Keystore. The browser gets an HttpOnly cookie.

What I learned: a boundary is a place you can point at. Before that, “auth” was a paragraph I copied. After it, a new service did not get to invent its own session. I am grateful they let me pull that out while I was still the person who would have copied the paragraph.

Query, one contract for three clients

query is the client surface. One GraphQL endpoint, POST /graphql. It owns no database and no queue. It composes reads from each owner’s internal API and relays writes to the service that actually stores them. Typed clients are generated from that schema, so web and mobile stopped drifting from the contract.

What I learned: the schema is a product, not a convenience. A field you add on a Tuesday is a promise to three codebases. Generating the clients was the moment I stopped trusting myself to keep them in sync by hand.

Graph and feed, where scale stopped being a word

The social graph is its own service. Follows, counters, and lists live there, not glued onto identity or content, because follow volume does not grow at the same rate as profiles. Writes are idempotent. Reads return ids. Query joins those ids into the card a person sees.

The feed does not share that database. Timelines are materialised in Redis, fan-out on write, filled by a consumer of the graph’s events. Content does not serve the timeline. The feed does.

What I learned, and this is the turn from junior to senior for me: a system that can grow is one whose cost you can name. A query that joins everything feels fine until it does not. Splitting the graph from the feed was the first time I could say what would get expensive, and where.

Contracts, so the wire could not drift

Events between those services are not copied by hand. contracts is one catalog. It emits Python, Go, and TypeScript. A subject carries its version. A breaking change is a new subject and a window where both exist. CI fails if the generated tree drifts.

What I learned: seniority is mostly refusing the shortcut that only I understand. A silent edit to a payload is how half the consumers break on a Friday. The catalog made that edit loud.

Media issues a presigned upload to S3, then a worker builds image variants and transcodes video. When the file is ready it publishes through an outbox. Realtime is the socket on the other side: one connection, many topics, a sequence number so a client that drops can resume. The channel I lived with was media progress, from tick to done or failed.

Search is not a source of truth. It projects posts and profiles into OpenSearch from the events of content, profile, identity, and the graph. It can be rebuilt. Notifications do the same kind of job for a person’s inbox: consume the event, store the row, let query read it.

What I learned: async work needs a story you can tell at 2 a.m. Who accepted the upload, who is processing it, who tells the phone, who can rebuild the index if the projection lies. I got much better at that story at Socialler. I did not have it when I started.

Clubs

Progression is its own service: badges, benefits, and one club per person. It states facts. It does not decide what someone may wear. Appearance listens and grants that on its own.

Deorum club crest, a blue wolf. Felium club crest, a pink cat. Sapientium club crest, a green owl. Titanum club crest, a red reptile.
Deorum, Felium, Sapientium, Titanum. Four clubs, one membership.

What I learned: a product idea still wants a boundary. Clubs were easy to smear across profile. Keeping them separate meant a badge could change without a rewrite of the card that shows it.

The review bar

AWS was provisioned with Terraform. Every service shipped through GitHub Actions on self-hosted runners, with tracing and metrics on the workloads. On iOS and Android I set the architecture, and the review bar around it: conventional commits, code review, and CI.

That last part is the senior work I am proudest of, and the part I could not have done in the first year. A bar only holds if other people trust it. Socialler gave me the time, and the colleagues, to earn that.

I got the title by staying. The skills came from systems that were a step past me, and from people who reviewed the step.

I owe the company a great deal. It was a rewarding place to grow up as an engineer: real users, real clients, and the permission to cut a network until each part could grow. It ended in September 2026. I left more capable than I arrived, and I am very grateful.

views

Comments

0:00