Questions we are testing
We publish the method before the result. Numbers appear here only once they have been measured on controlled hardware, never estimated.
Current investigations
NVMe behaviour under sustained compaction
How do enterprise and mixed use NVMe devices behave when a state database compacts continuously at high fill levels?
- Identical hosts, identical client version and network
- Fill levels at 50, 70 and 85 percent
- Measure p50, p95 and p99 read latency during compaction windows
- Record block processing delay against a reference node
Cache hit profile by RPC method
Which methods are safely cacheable, at what TTL, and what share of a realistic traffic mix do they represent?
- Replay anonymised traffic shapes against a test pool
- Vary TTL per method class
- Measure hit ratio, staleness and node load reduction
Cloud vs baremetal node throughput
How does block processing throughput compare between network attached storage and local NVMe under the same client and network?
- Matched CPU and memory across environments
- Same client version and configuration
- Measure sync rate, processing delay and IO latency percentiles
ClickHouse insert batching and merge pressure
What batch size keeps part counts stable at a given ingestion rate?
- Fixed ingestion rate, varying batch size
- Track part count, merge queue depth and query latency
- Identify the point where merges stop keeping up
Measured failover recovery time
How long does a realistic region drain and recovery actually take end to end?
- Scheduled drill on a staging fleet
- Measure detection, routing shift and full capacity restoration
- Repeat quarterly and track the trend
Measured, or not published
Infrastructure benchmarks are easy to produce and easy to make meaningless. Ours only leave the lab with the hardware, client version, configuration and method attached, so anyone can reproduce or challenge them.
Until a test is complete, this page shows the question and the method, not a conclusion.