Based on Ravishankar Raju's session at NextQuery 2026, PhoenixAI's Data + AI Infrastructure Summit. Raju is a software engineer on Coinbase's data platform team. Watch the session
About Coinbase
Coinbase is a platform where customers trade and track crypto, on-chain tokens, prediction markets, equities and options. Its apps chart how each asset has moved, and customers use those charts to make a prediction or decide what to do next.
Coinbase's data platform team runs StarRocks as a shared platform for other teams at Coinbase and decides when those teams get more capacity. Coinbase also uses StarRocks for blockchain analytics in fraud detection and compliance, described in a separate case study.
The challenge
Two customer-facing dashboards that can't wait
Coinbase's customer-facing price charts sit behind two dashboards, and both need fresh data, fast queries and high concurrency. Prediction markets are contracts with a live price and odds. A market can settle as soon as 15 minutes after it opens, so customers care most about 1-second candles over the last 48 hours. Portfolio tracking shows the gain on each asset a customer holds over the last hour, the last 24 hours, the last few days or a chosen range.
Live prices come from a different system at Coinbase. The history behind each chart is a separate workload: every customer who opens a dashboard queries it, so it has to stay current and fast at high concurrency.
Why key-value stores and warehouses fell short
Coinbase first served its price charts from key-value stores (DynamoDB and Redis). Custom workers rolled trades into candles for fixed time ranges and wrote the results to the stores. Lookups were fast, and the stores handled high concurrency. The approach did not break on performance. It broke as Coinbase grew: prediction markets, equities, options and on-chain tokens each needed their own workers. Adding an asset class took a lot of work, and every new pipeline was one more thing to maintain and one more thing that could break.
The chart queries are analytical, so a data warehouse such as Snowflake or Databricks could compute them. At the freshness and concurrency Coinbase's dashboards need, the warehouses fell short. For prediction markets, Coinbase had served historical data from a warehouse fed by a daily build, and trades took 5 to 29 hours to arrive. Snowflake and Databricks still handle history, BI and ML at Coinbase, and StarRocks can read the lakehouse in place when needed.

The solution
Why Coinbase chose StarRocks
Coinbase chose StarRocks for its price charts because it serves fast queries at high concurrency on fresh data, and because a new aggregation is a table and a few materialized views rather than another worker pipeline. One engine could serve every asset class. The platform runs on AWS, with one copy of the data on Amazon S3 and a cache on the compute nodes' local NVMe disks.
Trades reach StarRocks through a Kafka Connect sink connector, which writes them from Kafka into Primary Key tables. The tables are partitioned by hour and distributed by asset class, so a query for one asset over a time range reads only the matching partitions and touches very few buckets.

Materialized views and caches serve every chart tier
At Coinbase, a chain of StarRocks materialized views turns trades into chart bars. The 1-second view builds 1-second bars every minute, which keeps charts within a minute of the latest trade. The 1-second bars roll up into 5-minute bars, those into 1-hour bars and then 1-day bars, and each view refreshes automatically as data arrives. Chart services query whichever tier fits the requested time range, and no cron, Airflow or Flink jobs keep the rollups current. For the teams building on the platform, adding an asset class means adding a table and a few views, and the same monitoring covers every asset class.
Customers mostly look at 1-second bars from the last 48 hours, so Coinbase keeps only about 48 hours of them. That fits in the StarRocks compute-node caches on NVMe disks, and the most recent prediction-market queries return within 30 ms. Coinbase measured a p99 of 28 ms on this tier, against 766 ms for cold reads over full history.
As query volume grew, Coinbase's data platform team reached StarRocks concurrency limits. It added a time-series cache with L1 and L2 layers in front of the cluster, which also batches and jitters the calls that go through to StarRocks. The cache takes in almost 70,000 to 80,000 QPS, so the StarRocks cluster only has to be sized for the cache miss rate, not the request rate. That saves Coinbase a lot of cost.
One StarRocks warehouse per team
At first, several Coinbase teams shared one StarRocks warehouse, and with it one compute-node cache and one resource group. A backfill running in that warehouse evicted the hot data another team relied on and sent that team's reads to S3, so queries that should take 50 to 60 ms took 700 ms to a second. One team's query volume could also use up the shared concurrency limit.
Now each larger team, such as retail assets, has its own StarRocks warehouse with its own nodes, resource group and concurrency limits, and scales on its own. A shared warehouse serves small tables and small use cases. The split also lets Coinbase's data platform team attribute compute and cost to each team, so one team's scale-up is not charged to the others.
To watch the warehouses, Temporal jobs query StarRocks through a dedicated monitoring user. They check whether tables are fresh and whether materialized views, tagged by asset class, need alerts, then write the results to Datadog monitors that page the data platform team.
The results
One StarRocks engine now serves the historical charts behind Coinbase's prediction-market and portfolio dashboards:
- Fresh data: a new trade is queryable within a minute. Prediction-market trades used to take 5 to 29 hours to reach the warehouse.
- Fast queries at peak load: chart queries take about 10 ms at the median, with peaks of 6,000 to 7,000 QPS reaching StarRocks after the cache. Prediction-market queries have a median of 89 ms and a p99 of 292 ms.
- One pattern for every asset class: a new asset class is a table and a few views, monitored the same way as the others, and a view design that works for one carries over to the rest.
- Isolated capacity with clear costs: each larger team runs in its own warehouse, scales on its own and sees its own cost.
The figure below compares Coinbase's prediction-market data before and after the move to StarRocks.

Next, Coinbase plans to move its equity and options market data onto StarRocks. Order lookups and other high-QPS customer-facing services that need time-range or point-in-time lookups could follow.