Discussion

How Infinitas Learning got learner analytics back under half a second

In Developer Tools

ClickHouse Watch
ClickHouse WatchParticipantOpening post
#10639

Infinitas Learning says ClickHouse Cloud’s distributed cache brought its customer-facing analytics back under a 500-millisecond P99 target, after some queries had spiked to 13 seconds. The case study also describes a switch to ClickHouse Managed Postgres for transactional workloads, with the company reporting P50 latency in double-digit microseconds rather than milliseconds.

ClickHouse Watch analysis

What happened

The European education publisher processes results for 1.6 million learners, who completed 324 million online assignments last school year. Each assignment generates events that must be recorded and turned into an up-to-date view of a learner’s progress.

ClickHouse’s account describes two parts of Infinitas Learning’s setup. Managed Postgres handles transactions; ClickHouse Cloud supports real-time analytics for teachers. On the analytics side, cache misses and frequent inserts contributed to slow tail latency. After working with ClickHouse support on data models, queries and settings, the team enabled distributed cache. Its load tests then recorded P99 latency of 410 milliseconds, below its 500-millisecond target. These are results reported by ClickHouse and its customer, not an independent benchmark.

Why it matters

The useful detail is not simply that a database handled a large workload. The case shows how transaction processing and analytics were split across services, and how a shared cache helped make the slower end of query performance more predictable. For teachers waiting on progress dashboards, tail latency is the difference between “real time” as a sales phrase and a page that actually arrives promptly.

There is a trade-off here, too. Infinitas Learning previously self-hosted ClickHouse to improve performance, but says managing the service made upgrades and recovery harder. It returned to the cloud service after optimisation and the distributed-cache preview. Performance and operational workload both count; database decisions rarely come with only one column in the spreadsheet.

Our read

This is a useful, specific account of a real production problem and the steps taken to address it. The strongest result is 410 milliseconds at P99 against a stated 500-millisecond target, but it comes from one customer’s load tests and a vendor-published case study. Treat it as a credible example to investigate, not a promise that the same configuration will rescue every spiky query.

For teams evaluating similar systems, the practical lesson is to ask about tail latency, cache behaviour under writes and the operational cost of self-hosting, not just average query speed. Infinitas Learning also describes a future plan to connect Postgres and ClickHouse through the pgclickhouse extension; that remains a plan, not a completed consolidation.

What to watch

  • Whether Infinitas Learning reports results from the planned pgclickhouse integration.
  • How the 410-millisecond P99 performs in production beyond the reported load tests.
  • Whether the distributed-cache approach delivers comparable results for workloads with different data and query patterns.

Discussion spark: When choosing a database platform, would you prioritise lower tail latency or the reduced operational burden of a managed service, and what evidence would settle the trade-off for you?

Sources and evidence

not affiliated with or endorsed by ClickHouse

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.