35,812 records per second at saturation, batch 10 x 256B (2.5KB of values per request, group commit off, segment preallocation 64MB), over 15s with 0 errors, scenario produce, peaking at 64 connections (server at 1.17 of 3 cores, generator at 0.33 of 1 cores).
Measured with mix malachi.loadtest, the multi-core BEAM generator. Every number here comes from the recorded run described under Reproduce; none of it is maintained by hand. See Running the Elixir load test for the other ways to drive it.
Throughput
Measure
Value
Records per second
35,812
Operations per second
3,581
Data rate
8.74 MB/s
Records
537,180
Operations
53,718
Peak connections
64
Server CPU (cores)
1.17 of 3
Generator CPU (cores)
0.33 of 1
Duration
15s
Errors
0
Throughput by batch size
Each batch size ran over its own connection ladder, 1 repetition per point, with the points interleaved across batch sizes. Repeating the headline peak moved it -3.0% (35,812 then 34,748 records per second), so a difference between batch sizes smaller than that is noise. Compare batch sizes within this run rather than across runs: on a shared CI runner the published ceiling has moved by more than 30% between runs of unchanged code.