How Sellpy sharded MongoDB for 700 million items in-house with Better Stack
Sellpy is the H&M-backed secondhand marketplace where every one of its 700 million items is unique. Its engineers send over 30 TB of logs a month to Better Stack from more than 150 sources and build their dashboards on top of them with AI.
- Industry
- E-commerce / secondhand marketplace
- Products used
- Logs, Dashboards, Error tracking, MCP
- Switched from
- Papertrail
- Customer since
- April 2022
Where they started
Sellpy works a lot like a large online store, except every item it sells is unique, and the robots in its warehouse were built in-house because nothing on the market fit. Its engineers have always shaped their tools around that.
Sellpy has been a Better Stack customer since 2022. For a while it ran Better Stack and Papertrail side by side to make sure the integration worked smoothly, keeping a week of logs in Papertrail and longer retention in Better Stack. Moving fully to Better Stack was an easy decision after that.
Once all of Sellpy's logs lived in Better Stack, the team saw how much more it could build on top of them. About a year ago, Sofia Tovedal took over responsibility for keeping the whole system running and started building that layer: dashboards and alerts around the signals Sellpy cares about, for many teams with very different needs.
“There's never been an instance where I've thought, okay, we can't do this because the product doesn't fit. We can always build it ourselves.”
How they set it up: everything Sellpy runs, in one place
Sellpy sends over 30 TB of logs a month to Better Stack from more than 150 sources, most of them dev, canary and production copies of each service spread across AWS ECS, AWS CloudWatch, JavaScript applications and Heroku. Scheduled jobs and the eBay integration log there too. Where a system has nothing useful to send, Sofia writes a job that collects the data and logs it. One of those jobs queries MongoDB, Sellpy's main database, for metadata about each collection and shard, data MongoDB does not expose as metrics.
Because Better Stack is built on the OpenTelemetry standard, every source arrives in the same open format, whichever service or language produced it. Sellpy instruments once, with no proprietary agents, and can shape the data around what it wants to know. AI tools already understand the format, which makes building on it fast. Sofia builds most dashboards with AI through the Better Stack MCP, which she uses many times a day.
Her main system dashboard shows availability, the number of dynos running, and throughput and memory per dyno. A cron dashboard tracks every job and its outcome over time. The MongoDB and cron dashboards took about an hour together. MongoDB's counters are cumulative, so the charts have to calculate the deltas between readings. The AI wrote that SQL, and Sofia estimates it would have taken her weeks to work out alone. Sellpy now runs 110 dashboards with 1,385 charts.
Error tracking accepts Sentry's SDKs as is, so Sellpy could test it with almost no setup, and uptime monitoring, escalation policies and webhooks all run from the same account. 46 people across 11 teams use it, all engineers, plus product managers who follow the dashboards too.
Scaling MongoDB in-house, one shard at a time
As Sellpy grew, the team had to scale its MongoDB database horizontally. Sharding a database with 700 million unique items is a project many companies hire consultants for.
The work depended on seeing load at a level MongoDB does not report. Sofia's job writes per-collection and per-shard metadata to Better Stack, and the MongoDB dashboard turns it into load over time for each collection on each shard. That showed the team where the pressure was before they split the data, and whether it moved where they expected afterwards.
The main dashboard reads signals together. Availability alone can mislead: it can show 100% while request throughput sits at zero. A high dyno count can mean a planned backfill, or a database under strain. Reading those side by side is how Sofia has caught issues such as memory leaks and known straight away where the problem was.
“Knowing exactly what load is on what collection, on what shard, has been extremely helpful, and probably was why we could do it alone without using consultants.”
What it costs now: less engineering time per answer
Cost was one reason Sellpy chose Better Stack, and Sofia still counts it in Better Stack's favor. The larger saving is in engineering time. Sellpy built its own AI monitor in about five minutes: an agent that listens for errors from Better Stack, investigates them through the MCP, checks the related code on GitHub and can query the database and other services.
The MCP also runs in Sellpy's development environment. Before an engineer builds a feature, the AI queries production logs to learn the real scale it is working with, which Sofia says often surprises it. Support runs through a shared Slack channel, where Petr from Better Stack helped Sofia speed up slow dashboard queries and the bugs she reports get fixed quickly.
What is next: release dashboards and product data
Release monitoring is already routine. When the team ships a feature, someone asks Better Stack for a dashboard to watch it, and minutes later there is a rollout view that product managers can follow as well. They use the MCP themselves when they want their own views.
Next, Sofia wants to look at product tracking. Real user monitoring would bring user behavior into the same place as the rest of Sellpy's data.
“It's the first time we have all of these signals in one place, and I can tell, at any moment, whether Sellpy is healthy.”
Send your logs to Better Stack and ask the MCP for the views you want.