Instrument less, decide faster
Why this matters
Most teams reach for a rewrite when what they actually have is a measurement problem.
The cheapest version of this is a week of instrumentation and a single dashboard nobody has to ask for.
How it works
The client batches events in memory and flushes on a timer, on size, and on process exit.
Ordering is preserved per session; across sessions it is explicitly not guaranteed.
What it costs
Roughly four kilobytes on the wire per thousand events, and one long-lived connection.
The part people underestimate is cardinality: every unique property value is a column somewhere.
Where to go next
Start with three events you would actually act on, then resist adding a fourth for a month.