Load simulator

Set a viewer count, pick content, and watch traffic land. The generator is calibrated to the supplied extract — 40s heartbeat cadence, lognormal session durations, background and pause episodes — so this carries the same disorder the real stream has.

viewers

concurrent sessionssteady-state simultaneously active sessions — the natural dial for this workloaddistinct usersfewer users than sessions produces multi-session users, which is what makes session and user concurrency differbot shareshare of sessions on one shared identity, reproducing the observed 95-concurrent outlier

content

no matches

or sample this many catalogue rowspinning a handful manufactures a live-event spike on one asset; sampling is the realistic spread

pace

speed ×ramp-up s

speed 30 = 30 event-seconds per wall second; 0 = as fast as ClickHouse accepts

duration minmax events

both 0 = run until stopped

late fractiondup fraction

measured: 7% out-of-order, 0.465% byte-identical duplicates. Set both to 0 for a perfectly ordered stream when isolating a bug from the disorder.

batch sizeworkers

write pathClickHouse native protocol, straight through chx.Loaderasync insert— for small batches, not 50k ones

ingest

state

idle

rows sent

rows/sec

batches

retries

insert p50

0ms

insert p99

0ms

elapsed

sessions

—

peak conc.

—

still open

—

late

—

duplicates

—

dropped

—

Generator totals are reported when the run completes — the row counters above are live.

sessions active per minute

no traffic in the window yet

This is the heartbeat-lease estimate, not served concurrency. It cannot retract paused or backgrounded time — on the supplied extract the same estimator peaks at 3,162 against an exact 2,305, so read it as a liveness signal only. The exact curve needs the minute layer, which is not deployed yet.