🌀 From Connector Chaos to Clean Streams: PLAID’s Real-Time Architecture Overhaul
How a team cut Kafka connector costs and scaled cleanly with MongoDB Atlas Stream Processing
The setup
PLAID, a Tokyo-based company behind the real-time analytics platform KARTE, had used MongoDB since 2015. Their product ingests and processes millions of user interaction events across web and mobile in near real-time. The goal was to let customers personalize user experiences with up-to-the-second context.
Originally, they ran MongoDB self-hosted. Then in 2021, they moved to MongoDB Atlas. With Atlas came the ambition to tighten the feedback loop: real-time event ingestion, transformation, and analysis, without the complexity of batch jobs. They wanted everything piped into BigQuery to drive their downstream ML and reporting.
That’s where things started to get messy.
The problem: pipeline sprawl
To stream data from Atlas to BigQuery, they relied on Kafka Connect with custom connectors. The shape of the pipeline looked like this:
MongoDB Atlas → Kafka Source Connector → Kafka → BigQuery Sink Connector
In theory, this worked.
But as the product scaled, reality set in:
Connector count exploded. One per collection, per environment (prod, staging, etc.). Multiply that by services and teams.
Pricing was linear per connector. Each new pipeline meant higher Confluent Cloud costs.
Staging overhead doubled everything. Even feature testing meant provisioning duplicate connectors.
Monitoring was brittle. Kafka Connect didn’t integrate cleanly with Datadog. They had limited visibility into lags and throughput.
Backfills were clunky. There was no clean way to replay historical data into Kafka or BigQuery without hacks.
PLAID hit a tipping point: they were solving data engineering problems that had nothing to do with their core product.
The solution: stream closer to the source
Instead of solving these problems at the Kafka layer, PLAID looked upstream. In particular, they replaced the Kafka Source Connector with MongoDB Atlas Stream Processing (ASP).
